JPH09127977A - Voice recognition method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To sharply reduce the calculation quantity by reading the reference probability, and obtaining the forward probability of the present frame number when the distance between the voice feature vector of the present frame number and the voice feature vector of the standard frame number becomes a threshold value or below. SOLUTION: When the present frame number (t) is the terminal frame number T or below, a collation section obtains the distance dts between voice feature vectors xt , xqs of the present frame number (t) and the standard frame number qs. When the distance dts is a threshold value DTS or below, the output probability Bji (xt ) of the present frame number (t) is approximately equal to the reference probability Bji, and the reference probability Bji is not rewritten. The output probability Bji is not calculated by an equation, each reference probability Bji is read out, and the logarithmic forward probability Cit is obtained. The forward probability Cit is obtained with the output probability Bji (xqs ) of the standard frame number qs.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method using a Hidden Markov Model as a standard pattern for recognition and verification.

[0002]

2. Description of the Related Art Hidden Markov Model (Hidden
Markov Model. Hereinafter, HMM) can appropriately express patterns including fluctuations such as voice patterns, which vary with vocalization speed, individual differences in vocalization, and articulatory coupling.
Therefore, it is widely used in the field of voice recognition.
HMMs used in speech recognition have several states, such as S _0.
˜S _3, and a probability a _ij of transition from the state S _i to the state S _j and an output probability b _ij (x) of the speech feature vector x output at the time of the transition, and generally, the output probability b _ij ( x) is expressed as an uncorrelated mixed normal distribution composed of a plurality of normal distributions.

In the voice recognition method using the HMM, a voice feature vector x is calculated for each frame of a voice section from a voice signal.
_t , and then output probability b of speech feature vector x _{t
Find ij} (x _t ). As the output probability b _ij (x _t ), typically b _ij (x _t ) = Σ {λ _ijm b _ijm (x _t )} is calculated. Here, λ _ijm is the weight of the m-th normal distribution in the uncorrelated mixed normal distribution, and b _ijm (x _t ) is the speech feature vector x obtained from the m-th normal distribution in the uncorrelated mixed normal distribution.
Indicates the output probability of _t (output probability without weighting).

Then, the likelihood between the time series of the voice feature vector x _t extracted from the start frame to the end frame of the voice section and the HMM is calculated as the output probability b _ij (x _t ) of each voice feature vector x _t. Use to find. The likelihood is calculated for each HMM prepared as a standard pattern, and the maximum likelihood H is obtained.
The category assigned to the MM is used as the recognition result.

[0005]

However, the output probability b _ij (x _t ) = Σ {λ _ijm b of the speech feature vector x _t.
ijm (x _t )} requires enormous calculation, and thus it is difficult to quickly calculate the likelihood between the time series of the speech feature vector x _t and the HMM.

Therefore, it has been desired to more simply obtain the output probability b _ij (x _t ) of the voice feature vector x _t while suppressing the error.

[0007]

In order to solve the above-mentioned problems, the speech recognition method according to the inventions of claims 1 to 8 is a time series of speech feature vectors extracted from the start frame to the end frame of a voice section. The likelihood ln {P (x ₁ , x, x between x ₁ , x ₂ , ..., X _T and the Hidden Markov model
_2, ..., seek x _T)}, a category that is given to the hidden Markov model to obtain the maximum likelihood, the speech recognition method according to the recognition result for the speech signal in the speech segment,

[0008]

(Equation 4)

However, i: i = 1, 2, ..., I j: j = 1, 2, ..., J Φ _i : Probability that the initial state is S _i in the Hidden Markov model a _ji : Hidden Markov model At the state S _j to the state S _i in the above, x _t : the speech feature vector extracted in the t-th frame in the speech section (1 ≦ t ≦ T, and the first frame corresponds to the speech section). The start frame and the T-th frame represent the end frame of the speech section) b _ji (x _t ): speech feature vector x output when transitioning from state S _j to state S _i in the Hidden Markov model
Output probability of _t c _it : Time series of speech feature vector x ₁ , x ₂ , ..., Which starts transition from the initial state in Hidden Markov model
Forward probability of outputting x _t and reaching state S _i * i: Likelihood ln {P (x (x)) is obtained by using each equation represented by state number i given to state S _i that is the final state in the Hidden Markov model. ₁ , x ₂ , ...
., X _T )} is obtained, the following processing is performed.

<Invention of Claims 1 and 2> That is, in the speech recognition method of the invention of Claim 1, a storage unit for storing the reference frame number qs and the reference probability b _ji is provided, and the reference probability b _ji. , The forward probability c _it in each case of t = 1, 2, ..., T is sequentially obtained.

And (1). When the t = 1, is initialized to 1 reference frame number qs, all j, for i, the output probability b _ji the (x _t) determined from the hidden Markov model the output probability b _ji the (x _t) writing the initial value of the reference probability b _ji, after completion of writing of the reference probability b _ji reads each reference probability b _ji seek forward probability c _it processes (1A)
Then, after the process (1A) is completed, the process (1B) of adding 1 to the current frame number t is performed.

(2). When 2 ≦ t ≦ T, the voice feature vector x _t of the current frame number _t and the reference frame number qs
Of the voice feature vector x _qs of the
When S is compared with S and the comparison result is dts> DTS, the reference frame number qs is rewritten to the current frame number t, and the output probabilities b _ji (x _t ) for all j and i.
Rewriting the reference probability b _ji determined from hidden Markov model to the output probability b _ji (x _t), determine the forward probability c _it reads each reference probability b _ji after rewriting completion of the reference probability b _ji, the comparison If the result is dts ≦ DTS, reads each reference probability b _ji without rewriting the reference probability b _ji seek forward probability c _it processing (1C), after the processing (1C), the current frame A process (1D) of adding 1 to the number t is performed.

As described above, in the first aspect of the invention, the initial value of the reference probability b _ji is the output probability b _ji (x ₁ ) obtained from the Hidden Markov model in the starting frame, and the reference frame number q
The initial value of s is the frame number 1 of the start frame.

Then, the voice feature vector x _t of the current frame number t and the voice feature vector x qs of the reference frame number _qs
The distance dts between and is compared to a threshold DTS. dts>
In the case of DTS, the reference frame number qs and the reference probability b _ji are rewritten, and the rewritten reference probability b _ji is performed.
Is read to obtain the forward probability c _it . dts ≦ DTS
In the case of, rewriting of the reference frame number qs and the reference probability b
The rewriting of _ji is not performed, and the reference probability b _ji that has not been rewritten is read to obtain the forward probability c _it .

Therefore, the reference probability b stored in the storage unit
_ji is the output probability b _ji (x _t ) obtained from the Hidden Markov model in the frame with the reference frame number qs.

If dts> DTS, the distance dt
s is a case where since exceeding the threshold DTS audio feature vector x _t of the current frame number t does not approximate to the speech feature vector x _qs reference frame number qs before rewriting, thus the output probability b _ji (x of the current frame number t _t ) cannot be approximated by the output probability b _ji (x _qs ) of the reference frame number qs before rewriting, that is, the reference probability b _ji . Therefore, the reference probability b _ji is rewritten to the output probability b _ji (x _t ) of the current frame number t, and the rewritten reference probability b _ji is read to obtain the forward probability c _it . Further, since the reference probability b _ji is rewritten to the output probability b _ji (x _t ) of the current frame number t, the standard frame number qs is rewritten to the current frame number t.

If dts≤DTS, the distance dts is less than or equal to the threshold value DTS, and therefore the voice feature vector x _t of the current frame number t is not rewritten to the reference frame number qs.
Is approximately equal to the speech feature vector x _{qs of,} and the output probability b _ji (x _t ) of the current frame number t is
It is approximately equal to the output probability b _ji (x _qs ) of the reference frame number qs, that is, the reference probability b _ji . Therefore, the reference probability b _ji
Without rewriting, the reference probability b _ji is read to obtain the forward probability c _it . Since the reference probability b _ji is not rewritten, the reference frame number qs is not rewritten.

In this way, in the case of dts> DTS, after the reference probability b _ji is rewritten, therefore, after the output probability b _ji (x _t ) of the current frame number t is calculated from the Hidden Markov model, The reference probability b _ji is read to obtain the forward probability c _it . Furthermore, the distance dts ≦ threshold value DTS
In the case of, the reference probability b _ji is not rewritten, and thus the reference probability b _ji is read out without performing the calculation for obtaining the output probability b _ji (x _t ) of the current frame number t from the Hidden Markov model. Since c _it is obtained, the forward probability c _it
It is possible to reduce the calculation amount while suppressing the error of.

The error of the forward probability c _{it in} this case is d
The forward probability c obtained without performing the calculation of the output probability b _ji (x _t ) from the Hidden Markov model when ts ≦ DTS
_It is the difference between _it and the forward probability c _it obtained without such simplification of the operation.

As the threshold value DTS is increased, the amount of calculation reduction increases, but the error of the forward probability c _it increases. Therefore, the forward probability c is within the error range practically desired.
_It is necessary to set the value of the threshold value DTS so that it can be obtained.

According to the speech recognition method of the second aspect of the invention, the following processing is performed in the speech recognition method of the first aspect of the invention.

(1). When t = 1, the reference frame number qs is initialized to 1, the skip number skips is initialized to 0, and the output probabilities b _ji (x
_t ) is calculated from the Hidden Markov model and the output probability b _ji (x
_t ) is written as the initial value of the reference probability b _ji , and the reference probability b
After the writing of _ji is completed, each reference probability b _ji is read out to obtain the forward probability c _it (1A), and after the process (1A) is finished, a process (1B) of adding 1 to the current frame number t is performed.

(2). 2 ≦ t when the ≦ T, together with comparing the number of skips skips a threshold NSKIPS, threshold DTS distance dts between the speech feature vector x _qs speech feature vector x _t and the reference frame number qs of the current frame number t And the comparison result is skips> NSK.
When IPS or dts> DTS, the skip number skips is initialized to 0, the reference frame number qs is rewritten to the current frame number t, and the output probabilities b _ji (x _t ) are set for all j and i. hidden rewritten reference probability b _ji determined from Markov model to the output probability b _ji (x _t), determine the forward probability c _it reads each reference probability b _ji after rewriting completion of the reference probability b _ji, the comparison result Is skips ≦ NSKIPS and dts ≦ DTS, 1 is added to the skip number skips, and
Without rewriting the reference probability b _ji reads each reference probability b _ji seek forward probability c _it processing (1C), after the processing (1C), the process of adding 1 to the current frame number t (1D ) And do.

As described above, in the second aspect of the present invention, the initial value of the reference probability b _ji is set to the output probability b _ji (x ₁ ) obtained from the Hidden Markov model in the starting frame, and the reference frame number q
The initial value of s is the frame number 1 of the start frame, and the initial value of the skip number skips is 0.

Then, the skip number skips is set to the threshold value NSK.
With comparison with IPS, comparing the distance dts between the speech feature vector x _qs speech feature vector x _t and the reference frame number qs of the current frame number t and the threshold DTS. s
If kips> NSKIPS or dts> DTS, the skip number skips is initialized, the reference frame number qs is rewritten, and the reference probability b _ji is rewritten, and the rewritten reference probability b _ji is read to obtain the forward probability c _it . . When skips ≦ NSKIPS and dts ≦ DTS, the skip number skips is counted up, the reference frame number qs is not rewritten, and the reference probability b _ji is not rewritten, but the reference probability b that is not rewritten.
_ji is read to obtain the forward probability c _it .

Therefore, the reference probability b stored in the storage unit
_ji is the output probability b _ji (x _t ) obtained from the Hidden Markov model in the frame with the reference frame number qs.

If dts> DTS, the distance dt
s is a case where since exceeding the threshold DTS audio feature vector x _t of the current frame number t does not approximate to the speech feature vector x _qs reference frame number qs before rewriting, thus the output probability b _ji (x of the current frame number t _t ) cannot be approximated by the output probability b _ji (x _qs ) of the reference frame number qs before rewriting, that is, the reference probability b _ji . Therefore, the reference probability b _ji is rewritten to the output probability b _ji (x _t ) of the current frame number t, and the rewritten reference probability b _ji is read to obtain the forward probability c _it . Further, since the reference probability b _ji is rewritten to the output probability b _ji (x _t ) of the current frame number t, the standard frame number qs is rewritten to the current frame number t. Number of skips skips
Represents the number of times the reference probability b _ji was not rewritten within the range of skips ≦ NSKIPS, so the skip number skips is initialized.

In the case of skips> NSKIPS, the number of times skips in which the reference probability b _ji has not been rewritten exceeds the threshold value NSKIPS, so that the time gap between the current frame number t and the reference frame number qs becomes large, thus increasing the error. Is likely to. Therefore, the reference probability b _ji is rewritten in order to reduce the error. Therefore, the reference probability b
_{Since ji} is rewritten to the output probability b _ji (x _t ) of the current frame number t, the reference frame number qs is rewritten to the current frame number t. The skip number skips is skips ≦
Since it represents the number of times the reference probability b _ji was not rewritten within the range of NSKIPS, the number of skips s
Initialize the kips.

Skips ≦ NSKIPS and dts ≦ D
For TS, a case where the audio feature vector x _qs reference frame number qs is not performed rewriting a speech feature vector x _t of the current frame number t is equal to approximately since at dts ≦ DTS, therefore the current frame The output probability b _ji (x _t ) of the number t is approximately equal to the output probability b _ji (x _qs ) of the reference frame number qs that is not rewritten, that is, the reference probability b _ji . Moreover, skips ≦ NSKIPS, and therefore the number of times skips at which the reference probability b _ji has not been rewritten is less than or equal to the threshold value NSKIPS, so that the time gap between the current frame number t and the reference frame number qs is small, and thus the error can increase. The sex is low. So without rewriting the reference probability b _ji, it reads the reference probability b _ji by obtaining the forward probability c _it. Therefore, since the reference probability b _ji is not rewritten, the reference frame number qs is not rewritten. The skip number skips is skip
Since it represents the number of times the reference probability b _ji was not rewritten within the range of s ≦ NSKIPS, 1 is added to the skip number skips to count up the skip number skips.

As described above, in the case of skips> NSKIPS or dts> DTS, after the reference probability b _ji is rewritten, the output probability b _{ji of the} current frame number t is accordingly.
After performing a calculation to obtain (x _t ) from the Hidden Markov model, the reference probability b _ji is read to obtain the forward probability c _it . Furthermore, skips ≦ NSKIPS and dts ≦ DT
In the case of S, the reference probability b _ji is read out without rewriting the reference probability b _ji , and thus the output probability b _ji (x _t ) of the current frame number t is not calculated from the Hidden Markov model. Since the probability c _it is obtained, the forward probability c
_The amount of calculation can be reduced while suppressing the error of _it .

The error of the forward probability c _{it in} this case is s
In the case of kips ≦ NSKIPS and dts ≦ DTS, the output probability b _ji (x _t ) is obtained without performing the calculation for obtaining from the Hidden Markov model, and the forward probability c _it is obtained without simplifying such calculation. This is the difference from the forward probability c _it .

As the threshold value DTS is increased, the amount of reduction in calculation increases, but the error of the forward probability c _it increases. Therefore, the forward probability c is within the error range practically desired.
_It is necessary to set the value of the threshold value DTS so that it can be obtained.

<Invention of Claims 3 to 6> Furthermore, in the speech recognition method of the invention of Claim 3, the state S _j which is the transition source in the Hidden Markov model has one of a stationary part and a transient part. s is added to the reference frame number q
s, a transition part reference frame number qt, and a reference probability b _ji are provided in a storage unit, and the reference probability b _ji is used to t =
The forward probability c _it for each of 1, 2, ..., T is sequentially obtained.

And (1). When t = 1, the constant part reference frame number qs and the transient part reference frame number qt are initialized to 1, and all j and i are
Output Write probabilities b _ji (x _t) the probability that output determined from Hidden Markov Models b _ji the (x _t) as the initial value of the reference probability b _ji, after completion of writing of the reference probability b _ji reads each reference probability b _ji A process (2A) for obtaining the forward probability c _it and a process (2B) for adding 1 to the current frame number t are performed after the process (2A) is completed.

(2). When the 2 ≦ t ≦ T, the distance dts between the speech feature vector _qs speech feature vector x _t and the constant part reference frame number qs of the current frame number t is compared with a threshold DTS, the comparison result is dts> DTS between when made, processing of rewriting the constant part reference frame number qs to the current frame number t and (2C), an audio feature vector x _qt speech feature vector x _t and the transient portion reference frame number qt the current frame number t The distance dtt of the threshold value DTT is compared with the threshold value DTT, and when the comparison result is dtt> DTT, the transition part reference frame number qt is rewritten to the current frame number t (2D) and the processes (2C) and (2D). After the end, the output probability b _ji (x
_t ) is compared with the process (2E) for determining the type s _assigned to the transition source S _j of the state transition and the process (2C) when the type determination result of the process (2E) is a stationary part. If the result is dts> DTS, the type determination result is obtained j
For, regarding all i, the output probability b _ji (x _t ) is obtained from the Hidden Markov model, and the reference probability b _ji is rewritten to the output probability b _ji (x _t ), and the type determination result of the process (2E) is the stationary part. If the comparison determination result of the process (2C) is dts ≦ DTS, the reference probability b _ji is not rewritten for j for which the type determination result is obtained, and the process (2
When the type determination result of E) is the transition part, the process (2
If the comparison result of D) is dtt> DTT, the output probability b _ji (x _t ) is obtained from the Hidden Markov model for all i with respect to j for which the type determination result is obtained, and the reference probability b
_{If ji} is rewritten to the output probability b _ji (x _t ), and the type determination result of the process (2E) is the transition part, and the comparison determination result of the process (2D) is dtt ≦ DTT, the type determination For j that has obtained the result, a process (2F) in which the reference probability b _ji is not rewritten, and j = 1, 2, ...
The process (2F) is performed for each j, and when the process (2F) is completed for all j, the process (2G) of reading out each reference probability b _ji to obtain the forward probability c _it , and after the process (2G), Processing for adding 1 to the current frame number t (2H)
And.

As described above, in the invention of claim 3, j = 1,
For each j of 2, ..., J, the type s _assigned to the transition source S _j of the state transition that gives the output probability b _ji (x _t ) is determined.

When the type s is the stationary part, and the distance dts related to the stationary part exceeds the threshold value DTS, the output probability b _ji (x _t ) of the current frame number t is obtained for j that obtained the type.
From the Hidden Markov model, and the reference probability b _ji is rewritten to the output probability b _ji (x _t ). After that, the reference probability b _ji is read to obtain the forward probability c _it . If the distance dts related to the stationary part is equal to or less than the threshold value DTS, the reference probability b _ji is not rewritten for j for which the type s is obtained, and accordingly, the output probability b _ji (x _t ) of the current frame number t The reference probability b _ji is read out and the forward probability c _it is calculated without calculating H from the Hidden Markov model. Therefore, with respect to j for which the determination result that the type s is the stationary part is obtained, it is possible to reduce the amount of calculation while suppressing the error of the forward probability c _it .

The error of the forward probability c _{it in} this case means that j is obtained as a result of the determination that the type s is a stationary part.
dts ≦ a forward probability c _it got an output probability b _ji (x _t) without operation for obtaining the hidden Markov model in the case of DTS, such forward probability c _it obtained without simplification of operation Is the difference between.

The reference probability b _ji is rewritten according to the comparison result of the distance dts related to the stationary part and the threshold value DTS for the following reason. That is, if dts> DTS, the constant part reference frame number qs is rewritten. However, since dts> DTS, the voice feature vector x _t of the current frame number _t is the constant part reference frame number qs before the rewriting. not approximate the speech feature vector x _qs, hence the speech feature vector x _t of the current frame number t greater change from speech feature vector x _qs constant region reference frame number qs, output probability b _ji the current frame number t (x _t ) cannot be approximated by the reference probability b _ji . Also the distance dts
If ≦ threshold value DTS, the constant part reference frame number qs is not rewritten. However, since dts ≦ DTS, the voice feature vector x _t of the current frame number _t is the constant part reference frame number qs that is not rewritten. Since the speech feature vector x _qs is approximately equal to the speech feature vector x _qs, and therefore the speech feature vector x _t of the current frame number t changes little from the speech feature vector x _qs of the stationary part reference frame number qs,
The output probability b _ji (x _t ) of the current frame number t can be approximated by the reference probability b _ji .

Similarly, if the type s is the transient part, the transient
If the distance dtt related to the part exceeds the threshold value DTT,
The output probability of the current frame number t for j that has obtained another s
b_ji(x_t) From the Hidden-Markov model and reference
Rate b_jiIs the output probability b_ji(x _t), And after that,
Reference probability b_jiAnd the forward probability c_itAsk for. Ma
If the distance dtt related to the transition part is less than or equal to the threshold value DTT,
For example, for j that obtained the type s, the reference probability b_jiBook of
Output probability of the current frame number t without changing
b_ji(x_t) From the Hidden Markov model,
Rate b_jiAnd the forward probability c_itAsk for. This
Therefore, regarding j that obtained the determination result that the type s is a transient part,
And the forward probability c_itError while suppressing the amount of calculation
Can be made.

The error of the forward probability c _{it in} this case means that j is obtained as a result of the judgment that the type s is a transient part.
dtt ≦ a forward probability c _it got an output probability b _ji (x _t) without operation for obtaining the hidden Markov model in the case of DTT, such forward probability c _it obtained without simplification of operation Is the difference between.

The reason why the reference probability b _ji is rewritten according to the comparison result of the distance dtt and the threshold value DTT relating to the transition part is as follows. That is, if dtt> DTT, the transient part reference frame number qt is rewritten. However, since dtt> DTT, the voice feature vector x _t of the current frame number _t is the transient part reference frame number qt before the rewriting. not approximate the speech feature vector x _qt, hence the speech feature vector x _t of the current frame number t greater change from speech feature vector x _qt transient portion reference frame number qt, output probability b _ji the current frame number t (x _t ) cannot be approximated by the reference probability b _ji . Also, the distance dtt
If ≦ threshold value DTT, the transient reference frame number qt is not rewritten. However, since dtt ≦ DTT, the voice feature vector x _t of the current frame number _t is the transient reference frame number qt of which rewriting is not performed. Since the voice feature vector x _qt is approximately equal to the voice feature vector x _qt, and therefore the voice feature vector x _t of the current frame number t has a small change from the voice feature vector x _qt of the transition part reference frame number qt,
The output probability b _ji (x _t ) of the current frame number t can be approximated by the reference probability b _ji .

Further, in the invention of claim 3, the reason why the threshold value DTS when the type s is a stationary part and the threshold value DTT when the type s is a transient part are individually set is as follows.
It depends on the following reasons.

That is, since the change of the voice feature vector x _t detected time-sequentially is large in the transition part of the voice signal, the threshold value DTT used when the type s is the transition part.
It is desired to reduce the error of the forward probability c _it by reducing.

On the other hand, in the stationary part of the audio signal, since the change of the audio feature vector x _t detected in time sequence is small, the threshold value DT used when the type s is the stationary part.
Even if S is increased, the error of the forward probability c _it can be reduced. Increasing the threshold value DTS contributes to a reduction in the amount of calculation.

Therefore, it is more effective to use a small value for the threshold value DTT used when the type s is a transient part and to use a large value for the threshold value DTS used when the type s is a stationary part. It is possible to reduce the calculation amount while reducing the error of the forward probability c _it .

According to the voice recognition method of the invention of claim 4, in the voice recognition method of the invention of claim 3, the process (2E) is performed after the processes (2C) and (2D) are completed.

As described above, in the invention of claim 4, dts,
Based on the comparison result of DTS, the reference frame number of the stationary part qs
Is performed (2C) and the transition part reference frame number qc is rewritten according to the comparison result of dtc and DTC (2D), and then the output probability b _ji (x _t ) of the current frame number t is calculated. A process (2E) of determining the type s _assigned to the transition source S _j of the given state transition is performed. Therefore, the rewriting process (2C) of the stationary part reference frame number qs and the rewriting process (2D) of the transition part reference frame number qc are completed before the determination process (2E) of the type s,
Since it is not performed for each j of j = 1, 2, ..., J, the processing amount can be reduced. After performing the determination process (2E) for the type s, the rewriting process (2C),
If (2D) is performed, these rewriting processes (2C) and (2D) are performed for each individual j, which increases the processing amount.

Although the processing amount increases, in the invention of claim 3, the processing (2C) and (2D) may be performed after the processing (2E) is completed.

According to the speech recognition method of the fifth aspect of the present invention, in the speech recognition method of the third aspect, the state S _j that is the transition source in the Hidden Markov model is either a steady part or a transient part. A type s is given, and a storage unit is provided for storing a stationary part reference frame number qs, a transition part reference frame number qt, and a reference probability b _ji, and using the reference probability b _ji , t = 1, 2, ... ..., the forward probability c in each case of T
_It is calculated sequentially.

And (1). When t = 1, the constant part skip number skips and the transient part skip number skipt are initialized to 0, and the constant part reference frame number qs and the transient part reference frame number qt are initialized to 1 and all j are initialized. for i, the output probability b _ji (x _t) writes an initial value of the reference probability b _ji the output probability b _ji (x _t) determined from the hidden Markov models, reference probability b each reference probability after completion of writing of _ji b A process (2A) of reading _ji to obtain the forward probability c _it and a process (2B) of adding 1 to the current frame number t are performed after the process (2A) is completed.

(2). When the 2 ≦ t ≦ T, as well as comparing the constant region skip number skips a threshold NSKIPS, the distance between the speech feature vector x _qs speech feature vector x _t and the constant part reference frame number qs of the current frame number t dts is compared with the threshold value DTS, and the comparison result is sk.
When ips> NSKIPS or dts> DTS, the constant part skip number skips is initialized to 0, and the constant part reference frame number qs is rewritten to the current frame number t, and the comparison result is skips ≦ NSKIPS and dts ≦ DTS. If, then the number of skips in the stationary part sk
1 and adds the process (2C) in ips, while comparing the transient portion skip number skipt a threshold NSKIPT, the speech feature vector x _qt speech feature vector x _t and the transient portion reference frame number qt the current frame number t Distance d
tt is compared with the threshold value DTT, and the comparison result is skippt.
> NSKIPT or dtt> DTT,
The transition part skip number skipt is initialized to 0, and the transition part reference frame number qt is rewritten to the current frame number t, and the comparison result is skip≤NSKIPT and dt.
When t ≦ DTT, the number of skips in transition part skippt
Processing of adding 1 to (2D), processing (2C), (2
After the end of D), a process of determining the type s _assigned to the transition source S _j of the state transition that gives the output probability b _ji (x _t ) for each j of j = 1, 2, ..., J. (2E) and processing (2E)
When the type determination result of is a stationary part, processing (2C)
Comparison result of skips> NSKIPS or dts
> DTS, for j for which the type determination result is obtained, the output probability b _ji (x _t ) is obtained from the Hidden Markov model for all i, and the reference probability b _ji is set as the output probability b _ji (x
_t )) and the type determination result of the process (4E) is the stationary part, the comparison result of the process (2C) is skips ≦
If NSKIPS and dts ≦ DTS, the reference probability b _ji is not rewritten for j for which the type determination result is obtained, and if the type determination result of the process (2E) is a transient part, the process (2D) Comparison result of skip> NSKI
If PT or dtt> DTT, the output probability b _ji for all i with respect to j for which the type determination result is obtained.
(x _t ) is obtained from the Hidden Markov model, the reference probability b _ji is rewritten to the output probability b _ji (x _t ), and when the type determination result of the process (2E) is the transient part, the process (2D) If the comparison result is skipt ≦ NSKIPT and dtt ≦ DTT, a process (2F) in which the reference probability b _ji is not rewritten for j for which the type determination result is obtained, and j = 1,
2, ..., Performing the processing (2F) for each j of J,
When the process (2F) is completed for all j, a process of reading the reference probabilities b _ji to obtain the forward probability c _it (2
G) and after the processing (2G) is finished, 1 is set to the current frame number t.
Is performed (2H).

As described above, in the invention of claim 5, j = 1,
For each j of 2, ..., J, the type s _assigned to the transition source S _j of the state transition that gives the output probability b _ji (x _t ) is determined.

When the determination result of the type s is the stationary part,
The skip number skips related to the stationary part is the threshold value NSKIP.
If S is exceeded or the distance dts related to the stationary part exceeds the threshold DTS, the output probability b _ji (x _t ) of the current frame number t is obtained from the Hidden-Markov model for j for which the type is obtained, and the reference probability b _ji is rewritten to the output probability b _ji (x _t ), and after that, the reference probability b _ji is read to obtain the forward probability c _it . In addition, when the determination result of the type s is a stationary part, the skip number ski related to the stationary part
If ps is equal to or less than the threshold value NSKIPS and the distance dts related to the stationary part is equal to or less than the threshold value DTS, the reference probability b _ji is not rewritten for j for which the type s is obtained.
Therefore, the output probability b _ji (x _t ) of the current frame number t is not obtained from the Hidden Markov model, but the reference probability b _ji is read to obtain the forward probability c _it . Therefore, with respect to j for which the determination result that the type s is the stationary part is obtained, it is possible to reduce the amount of calculation while suppressing the error of the forward probability c _it .

The error of the forward probability c _{it in} this case means that j is obtained as a result of the judgment that the type s is a stationary part.
In the case of skips ≦ NSKIPS and dts ≦ DTS, the output probability b _ji (x _t ) was obtained without performing the calculation for obtaining from the Hidden Markov model, and the forward probability c _it was obtained without performing such calculation simplification. This is the difference between the forward probability c _it .

The reference probability b _ji is rewritten according to the comparison result of the distance dts related to the stationary part, the threshold value DTS and the comparison result of the skip number skips and the threshold value NSKIPS for the following reason.

When dts> DTS, the voice feature vector x _t of the current frame number t does not approximate to the voice feature vector x _qs of the reference frame number qs, and therefore the voice feature vector x _t of the current frame number _t is the reference. the change from speech feature vector x _qs frame number qs is large, can not be approximated by reference probability b _ji the output probability b _ji the current frame number _t (x t). Therefore, the reference probability b _ji is rewritten.

In the case of skips> NSKIPS, the number of times skips when the distance dts becomes equal to or less than the threshold value DTS exceeds the threshold value NSKIPS, so that the time gap between the current frame number t and the reference frame number qs becomes large, and thus the error increases. Since the possibility is high, the reference probability b _ji is rewritten in order to reduce the error.

Skips ≦ NSKIPS and dts ≦ D
In the case of TS, since dts ≦ DTS, the voice feature vector x _t of the current frame number _t is the reference frame number qs.
Since the approximate speech feature vector x _qs, hence speech feature vector x _t of the current frame number t is less change from speech feature vector x _qs reference frame number qs, output probability b _ji (x of the current frame number t _t ) can be approximated by the reference probability b _ji . Moreover, skips ≦ NSKIPS, and thus the number of times the distance dts becomes equal to or less than the threshold value DTS sk
Since ips is less than or equal to the threshold value NSKIPS, the time gap between the current frame number t and the reference frame number qs is small, and therefore the error is unlikely to increase. Therefore, in order to reduce the calculation amount, the reference probability b _ji is not rewritten.

Similarly, when the determination result of the type s is the transient part, the skip number skipt related to the transient part is the threshold NS.
Distance dtt that exceeds KIPT or is related to transition
Is greater than the threshold DTT, the output probability b _ji (x _t ) of the current frame number t is obtained from the Hidden-Markov model for j that has obtained the type, and the reference probability b _ji is the output probability b _ji (x _t ), And after that, the reference probability b _ji is read to obtain the forward probability c _it . In addition, when the determination result of the type s is the transient part, the number of skips s related to the transient part
If kipt is less than or equal to the threshold value NSKIPT and the distance dtt related to the transition part is less than or equal to the threshold value DTT, the type s
For j obtained, the reference probability b _ji is read out without rewriting the reference probability b _ji , and thus the output probability b _ji (x _t ) of the current frame number t is not obtained from the Hidden Markov model. ask for _it . For this reason, with respect to j for which the determination result that the type s is the transient part is obtained, the forward probability c
_The amount of calculation can be reduced while suppressing the error of _it .

The error of the forward probability c _{it in} this case means that j is obtained as a result of judging that the type s is a transient part.
In the case of skipt ≤ NSKIPT and dtt ≤ DTT, the output probability b _ji (x _t ) is obtained without performing the operation for obtaining from the Hidden-Markov model, and the forward probability c _it is obtained without performing such operation simplification. This is the difference between the forward probability c _it .

The reference probability b _ji is rewritten according to the comparison result of the distance dtt related to the transitional part, the threshold value DTT and the comparison result of the skip number skipt and the threshold value NSKIPT for the following reason.

When dtt> DTT, the voice feature vector x _t of the current frame number t does not approximate to the voice feature vector x _qt of the reference frame number qt, so the voice feature vector x _t of the current frame number _t is the reference. the change from speech feature vector x _qt frame number qt is large, can not be approximated by reference probability b _ji the output probability b _ji the current frame number _t (x t). Therefore, the reference probability b _ji is rewritten.

In the case of skipt> NSKIPT, the number of times the distance dtt becomes the threshold value DTT or less skipt exceeds the threshold value NSKIPT, so that the time gap between the current frame number t and the reference frame number qt becomes large, and thus the error increases. Since the possibility is high, the reference probability b _ji is rewritten in order to reduce the error.

Skippt NSKIPT and dtt≤D
In the case of TT, since dtt ≦ DTT, the voice feature vector x _t of the current frame number _t is the reference frame number qt.
Since the approximate speech feature vector x _qt, hence speech feature vector x _t of the current frame number t is less change from speech feature vector x _qt reference frame number qt, output probability b _ji (x of the current frame number t _t ) can be approximated by the reference probability b _ji . In addition, skipt ≦ NSKIPT, and therefore the number of times the distance dtt becomes equal to or less than the threshold value DTT
Since ipt is less than or equal to the threshold value NSKIPT, the time gap between the current frame number t and the reference frame number qt is small, which reduces the possibility of increasing the error. Therefore, in order to reduce the calculation amount, the reference probability b _ji is not rewritten.

Further, in the invention of claim 5, the thresholds DTS, NSKIPS and type s when the type s is a stationary part
The reason why the threshold values DTT and NSKIPT are individually set in the case where is a transition part depends on the following reason.

That is, since the change of the voice feature vector x _t detected in time sequence is large in the transition part of the voice signal, the threshold value DT used when the type s is the transition part.
It is desirable to reduce the error of the forward probability c _it by reducing T and NSKIPT.

On the other hand, in the stationary part of the audio signal, since the change of the audio feature vector x _t detected in time sequence is small, the threshold value DT used when the type s is the stationary part.
Even if S and NSKIPS are increased, the error of the forward probability c _it can be reduced. Threshold DTS, NSKIPS
Increasing the value contributes to the reduction of the calculation amount.

Therefore, the thresholds DTT and NSKIPT having a small value are used when the type s is the transient part, and the threshold DTS used when the type s is the stationary part,
By using a large value for NSKIPS, it is possible to more effectively reduce the error of the forward probability c _it and reduce the calculation amount.

Further, in the voice recognition method of the invention of claim 6, in the voice recognition method of the invention of claim 5, the process (2E) is performed after the processes (2C) and (2D) are completed.

As described above, in the invention of claim 6, the skip
A process (2C) of initializing or counting up the constant part skip number skips and rewriting the constant part reference frame number qs according to the comparison result of s, NSKIPS and the comparison result of dts, DTS, and skipt, NSKI.
Depending on the comparison result of PT and the comparison result of dtt and DTT, the process (2D) of initializing or counting up the transition part skip number skipt and rewriting the transition part reference frame number qt is performed. Frame number t
The process (2E) of determining the type s _assigned to the transition source S _j of the state transition giving the output probability b _ji (x _t ) of Therefore, the processes (2C) and (2D) related to the number of skips and the reference frame number are compared with the determination process (2
The processing amount can be reduced because the processing is finished before performing step E) and is not performed for each j of j = 1, 2, ..., J. When the rewriting process (2C) and (2D) relating to the skip number and the reference frame number are performed after the determination process (2E) of the type s, the skip number and the reference frame number are calculated for each j. Since the processes (2C) and (2D) related to are performed, the processing amount increases.

Although the processing amount increases, in the invention of claim 5, after the processing (2E) is completed, the processing (2C) and (2) are performed.
You may make it perform D).

<Invention of Claims 7 to 8> Further, in the voice recognition method of the invention of Claim 7, the forward probability reference frame number qc, the output probability reference frame number qs, and the reference probability b _ji are stored. A storage unit is provided, and the forward probability c _it in each case of t = 1, 2, ..., T is sequentially obtained using the reference probability b _ji .

And (1). When t = 1, the forward probability reference frame number qc and the output probability reference frame number q
Writing is initialized s to 1 in all j, for i, as the initial value of the reference probability b _ji the output probability b _ji (x _t) the probability that output determined from Hidden Markov Models b _ji (x _t) , Each reference probability b after writing the reference probability b _ji
Processing to read _ji and obtain forward probability c _it (3A)
Then, after the process (3A) is completed, the process (3B) of adding 1 to the current frame number t is performed.

(2). 2 ≦ t when the ≦ T, the distance between the speech feature vector x _qc of forward probabilities reference frame number qc and speech feature vectors x _t of the current frame number t dtc
Processing (3C) for comparing the value with the threshold value DTC, and processing (3C)
When the result of comparison of dtc ≦ DTC is satisfied, the forward probability c _it is assumed to be equal to the forward probability c _{i (t−1)} of the immediately preceding frame, and the calculation for obtaining the forward probability c _it is ended (3
D) and the process (3C) result in dtc> DTC, the forward probability reference frame number qc is rewritten to the current frame number t (3E), and the current frame number after the process (3E) is finished. The distance dts between the voice feature vector x _{t of t} and the voice feature vector x _qs of the output probability reference frame number qs is compared with the threshold DTS, and when the comparison result is dts> DTS, the output probability reference frame number qs is rewritten to the current frame number t, and the output probabilities b _ji (x _t ) are obtained from the Hidden Markov model for all j and i, and the reference probabilities b _ji are output probabilities b _ji (x _t ).
, And each reference probability b _ji is read after the reference probability b _ji is rewritten to obtain the forward probability c _{it. When the} comparison result is dts ≦ DTS, each reference probability b _ji is not rewritten. A process (3F) of reading the probability b _ji to obtain the forward probability c _it and a process (3G) of adding 1 to the current frame number t after the process (3D) or (3F) are completed.

As described above, in the invention of claim 7, the initial value of the reference probability b _ji is the output probability b _ji (x ₁ ) obtained from the Hidden-Markov model in the start frame. Then, the initial value of the forward probability reference frame number qc and the initial value of the output probability reference frame number qs are set as the frame number 1 of the start frame.

[0077] and compares the distance dtc between the speech feature vector x _qc of forward probabilities reference frame number qc and speech feature vectors x _t of the current frame number t and the threshold DTC. If dtc ≦ DTC, the forward probability c _it of the current frame number t is assumed to be equal to the forward probability c _{i (t−1)} of the immediately preceding frame, and the calculation of the forward probability c _it ends.

If dtc> DTC, the forward probability reference frame number qc is rewritten to the current frame number t,
Thereafter, comparing the distance dts between the speech feature vector x _qs speech feature vector x _t and the reference frame number qs of the current frame number t and the threshold DTS. If dts> DTS, the reference frame number qs is rewritten and the reference probability b _ji is rewritten, and the rewritten reference probability b _ji is read to obtain the forward probability c _it . If dts ≤ DTS,
The rewriting of the reference frame number qs and the rewriting of the reference probability b _ji are not performed, but the reference probability b _ji that has not been rewritten is read to obtain the forward probability c _it .

[0079] dtc For ≦ DTC, when the distance dtc is equal to or less than the threshold DTC and the speech feature vector x _qc of forward probabilities reference frame number qc and speech feature vectors x _t of the current frame number t is equal to approximately Therefore, since the forward probability c _it of the current frame number t changes from the voice feature vector x _qc of the forward probability reference frame number qc, the forward probability c _it of the current frame number t becomes.
Can be approximated by the forward probability c _{i (t-1)} of the immediately preceding frame. So forward probability c _it the current frame number t is as equal to the forward probability c _{i (t-1)} of the previous frame, and terminates the operation for obtaining the forward probability c _it.

[0080] The dtc> For DTC, distance dtc the case since exceeds a threshold DTC that the speech feature vector x _t and forward probabilities reference frame number qc speech feature vector x _qc of the current frame number t does not approximate, thus Since the speech feature vector x _t of the current frame number t has a large change from the speech feature vector x _qc of the forward probability reference frame number qc, the forward probability c _it of the current frame number t is the forward probability C _{i (t -1)} cannot be approximated. Therefore, since the reference probability b _ji is read and the forward probability c _it of the current frame number t is calculated, the forward probability reference frame number qc is rewritten to the current frame number t.

The reference probability b _ji stored in the storage unit is
The output probability is the output probability b _ji (x _t ) obtained from the Hidden Markov Model in the frame of the reference frame number qs.

When dts> DTS, the distance dt
s is a case where since exceeding the threshold DTS audio feature vector x _t of the current frame number t and the speech feature vector x _qs for rewriting the previous output probabilities reference frame number qs is not approximate, hence speech feature vectors of the current frame number t x _t is the speech feature vector x _qs of the output probability reference frame number _qs
Therefore, the output probability b _ji (x _t ) of the current frame number t cannot be approximated by the output probability b _ji (x _qs ) of the output probability reference frame number qs before rewriting, that is, the reference probability b _ji . Therefore, the output probability b _ji (x _t ) of the current frame number t
From the Hidden Markov model, the reference probability b _ji is rewritten to the output probability b _ji (x _t ), and then the reference probability b _ji is read to obtain the forward probability c _it . The reference probability b _ji is
Since the output probability b _ji (x _t ) of the current frame number t is rewritten, the output probability reference frame number qs is changed to the current frame number t.
Rewrite

[0083] dts For ≦ DTS is the distance dts is equal to or less than the threshold DTS and speech feature vectors x _qs output probabilities reference frame number qs is not performed rewriting a speech feature vector x _t of the current frame number t approximate Therefore, the voice feature vector x _t of the current frame number t has a small change from the voice feature vector x _qs of the output probability reference frame number qs, and thus the output probability b _ji (x _t ) can be approximated by the output probability b _ji (x _qs ) of the output probability reference frame number qs, that is, the reference probability b _ji . So without rewriting the reference probability b _ji, it reads the reference probability b _ji by obtaining the forward probability c _it. Since the reference probability b _ji is not rewritten, the output probability reference frame number qs is not rewritten.

Thus, in the case of dtc≤DTC, the forward probability c _it of the current frame number t is assumed to be equal to the forward probability c _{i (t-1)} of the immediately preceding frame, and the calculation of the forward probability c _it is completed. The calculation for _{obtaining the} output probability b _ji (x _t ) is not performed. And if dtc> DTC, dts> DTS
If so, after the reference probability b _ji is rewritten, and accordingly, the output probability b _ji (x _t ) of the current frame number t is calculated from the Hidden Markov model, and then the reference probability b _ji
Is read to obtain the forward probability c _it . Also dtc> D
In the case of TC, if dts ≦ DTS, the reference probability b _ji is not rewritten, and thus the output probability b _ji (x _t ) of the current frame number t is not calculated from the Hidden Markov model, and the reference probability since reads b _ji seek forward probability c _it, while suppressing the error of the forward probability c _it, it is possible to reduce the amount of calculation.

The error of the forward probability c _{it in} this case is d
When tc ≦ DTC or dts ≦ DTS, the output probability b _ji (x _t ) is obtained without performing the calculation for obtaining from the Hidden Markov model, and the forward probability c _it is obtained without simplifying such calculation. This is the difference from the forward probability c _it .

As the threshold values DTC and DTS are increased, the amount of reduction in calculation increases, but the error in the forward probability c _it increases. Therefore, the threshold DTC, so that the forward probability c _it can be obtained within the error range practically desired,
It is necessary to determine the value of DTS.

In the voice recognition method of the eighth aspect of the invention, the following processing is performed in the voice recognition method of the seventh aspect of the invention.

(1). When t = 1, the forward probability reference frame number qc and the output probability reference frame number qs are set to 1, respectively, and the forward probability skip number skipc,
Is initialized to output probability skip number skips to 0 respectively, all j, for i, the output probability b _ji (x _t) from hidden Markov model determined reference probability b _ji the output probability b _ji (x _t) Write as an initial value, and the reference probability b _ji
After the completion of the writing, the reference probability b _ji is read out to obtain the forward probability c _it (3A), and after the process (3A) is finished, the current frame number t is incremented by 1 (3B).

(2). When the 2 ≦ t ≦ T, the forward probability skip number skipc with is compared with a threshold value NSKIPC, the distance between the speech feature vector x _qc of forward probabilities reference frame number qc and speech feature vectors x _t of the current frame number t Process of comparing dtc with threshold DTC (3C)
And the comparison result of the processing (3C) is skipc ≦ NSKIP
When C and dtc ≦ DTC, the forward probability c _it is assumed to be equal to the forward probability c _{i (t-1)} of the immediately preceding frame, and the operation for obtaining the forward probability c _it is ended, and the forward probability skip number skippc, output Probability skip number ski
The process (3D) of adding 1 to ps and the process (3D)
The comparison result of C) is skipc> NSKIPC or d
When tc> DTC, the forward probability skip number sk
A process of initializing ipc to 0 and rewriting the forward probability reference frame number qc to the current frame number t (3E)
And, after the processing (3E) ends, the output probability skip number ski
with comparing ps a threshold NSKIPS, the distance dts between the speech feature vector x _qs output probabilities reference frame number qs with speech feature vectors x _t of the current frame number t is compared with a threshold DTS, the comparison result is skips > NS
When KIPS or dts> DTS, the output probability skip number skips is initialized to 0, the output probability reference frame number qs is rewritten to the current frame number t, and the output probability b _ji (for all j and i). x _t ) is calculated from the Hidden Markov model, the reference probability b _ji is rewritten to the output probability b _ji (x _t ), each reference probability b _ji is read out after the reference probability b _ji is rewritten, and the forward probability c _it is calculated. The comparison result is skips ≦ NSKIPS and dt.
Output probability skip number skip if s ≦ DTS
In addition to adding 1 to s, each reference probability b _ji is read without rewriting the reference probability b _ji (x _t ) and the forward probability c _it.
Process (3F) and process (3D) or (3
After the end of F), a process (3G) of adding 1 to the current frame number t is performed.

As described above, in the invention of claim 8, the initial value of the reference probability b _ji is the output probability b _ji (x ₁ ) obtained from the Hidden Markov model in the start frame. Then, the initial value of the forward probability reference frame number qc and the initial value of the output probability reference frame number qs are set as the frame number 1 of the start frame. Also, the forward probability skip number skip
The initial value of c and the initial value of the output probability skip number skips are set to 0, respectively.

Then, the forward probability skip number skippc is compared with the threshold value NSKIPC, and the current frame number t
The distance dtc between the voice feature vector x _{t of the above} and the voice feature vector x _qc of the forward probability reference frame number qc is compared with the threshold value DTC. skipc ≤ NSKIPC and dt
If c ≦ DTC, the forward probability c of the current frame number t
_It is assumed that _it is equal to the forward probability c _{i (t−1)} of the immediately preceding frame, and the calculation of the forward probability c _it is completed, and the forward probability skip number skippc and the output probability skip number skips are incremented. .

In addition, skippc> NSKIPC or d
When tc> DTC, the forward probability reference frame number q
c is rewritten to the current frame number t, the forward probability skip number skipc is initialized, and then the output probability skip number skips is compared with the threshold value NSKIPS, and the voice feature vector x _t of the current frame number _t and the output probability reference frame number are compared. The distance dts between _qs and the voice feature vector x _qs is compared with a threshold DTS. skips> NSK
In the case of IPS or dts> DTS, the output probability skip number skips is initialized, the output probability reference frame number qs is rewritten, and the reference probability b _ji is rewritten, and the rewritten reference probability b _ji is read and the forward probability c _it. Ask for. In the case of skips ≦ NSKIPS and dts ≦ DTS, the output probability skip number skips is counted up, and the reference probability not rewritten without rewriting the output probability reference frame number qs and the reference probability b _ji . Read b _ji to obtain the forward probability c _it .

Skipc≤NSKIPC and dtc≤D
In the case of TC, since dtc ≦ DTC, the voice feature vector x _t of the current frame number t is close to the voice feature vector x _qc of the forward probability reference frame number qc, and thus the voice feature vector x _t of the current frame number _t is Since the change in the forward probability reference frame number qc from the voice feature vector x _qc is small, the forward probability c _it of the current frame number t can be approximated by the forward probability c _{i (t-1)} of the immediately preceding frame. Moreover, s
Since kipc ≦ NSKIPC, the number of times ski in which the forward probability c _{i (t-1)} of the immediately preceding frame is not rewritten is ski.
Since pc is less than or equal to the threshold value NSKIPC, the time gap between the current frame number t and the forward probability reference frame number qc becomes small. Therefore, the error is unlikely to increase, so that the forward probability c of the current frame number t should be reduced in order to reduce the calculation amount.
_It is assumed that _it is equal to the forward probability c _{i (t-1)} of the immediately preceding frame, and the calculation for obtaining the forward probability c _it ends. Therefore, since neither the reference probability b _ji is read out to obtain the forward probability c _it nor the output probability b _ji is rewritten, neither the forward probability reference frame number qc nor the output probability reference frame number qs is rewritten. The forward probability skip number Skipc is complete the calculation of the forward probability c _it is approximated by the forward probability c _{i (t-1)} of the immediately preceding frame forward probability c _it the current frame number t in a range of a skipc ≦ NSKIPC The number of forward probability skips qc is calculated by adding 1 to the number of forward probability skips qc.
Count up. Further, the output probability skip number sk
Since ips represents the number of times the reference probability b _ji was not rewritten within the range of skips ≦ NSKIPS, 1 is added to the output probability skip number skips to count up the output probability skip number skips.

[0094] The dtc> For DTC, distance dtc the case since exceeds a threshold DTC that the speech feature vector x _t and forward probabilities reference frame number qc speech feature vector x _qc of the current frame number t does not approximate, thus Since the voice feature vector x _t of the current frame number t has a large change from the voice feature vector x _qc of the forward probability reference frame number qc, the forward probability c _it of the current frame number t is the forward probability c _{i (t of the} immediately preceding frame. _-1) cannot be approximated. Therefore, since the reference probability b _ji is read and the forward probability c _it of the current frame number t is calculated, the forward probability reference frame number qc is rewritten to the current frame number t. In addition, the forward probability skip number skipc is
It represents the number of times that the forward probability c _it of the current frame number t is approximated by the forward probability c _{i (t-1)} of the immediately preceding frame within the range of pc ≦ NSKIPC to complete the calculation of the forward probability c _it . , The forward probability skip number skipc is initialized to 0.

If skippc> NSKIPC, the number of times skippc is the threshold NSKIPC when the forward probability c _it of the current frame number t is approximated by the forward probability c _{i (t-1)} of the immediately preceding frame and the calculation of the forward probability c _it is completed. Since the current frame number t exceeds the forward probability reference frame number qc, the error is likely to increase. Therefore, in order to reduce the error, the reference probability b _ji is read and the forward probability c _it. Is performed. Therefore, the forward probability reference frame number qc is rewritten to the current frame number t. Further, the forward probability skip number skipc is sk
Current frame number t within the range of ipc ≦ NSKIPC
Forward probability c _it of the previous frame forward probability c _{i (t-1)}
Since it represents the number of times that the calculation of the forward probability c _it is completed by approximating with, the forward probability skip number skippc is initialized.

The reference probability b _ji stored in the storage unit is
The output probability is the output probability b _ji (x _t ) obtained from the Hidden Markov Model in the frame of the reference frame number qs.

If dts> DTS, the distance dt
s is the case where the audio feature vector x _qs output probabilities reference frame number qs before rewriting the audio feature vector x _t of the current frame number t since exceeding the threshold DTS is not approximate, the audio feature vector x of the current frame number t _{Since t} has a large change from the voice feature vector x _qs of the output probability reference frame number qs, the output probability b _ji (x
_t ) is the output probability b _ji (x
_qs ) That is, it cannot be approximated by the reference probability b _ji . Therefore, the reference probability b _ji is rewritten to the output probability b _ji (x _t ) of the current frame number t, and the rewritten reference probability b _ji is read to obtain the forward probability c _it . Further, since the reference probability b _ji is rewritten to the output probability b _ji (x _t ) of the current frame number t, the output probability reference frame number qs is rewritten to the current frame number t. Then, the output probability skip number skips is skip
Since it represents the number of times the reference probability b _ji was not rewritten within the range of s ≦ NSKIPS, the output probability skip number skips is initialized.

In the case of skips> NSKIPS, the number of times skips, in which the reference probability b _ji is not rewritten, exceeds the threshold value NSKIPS, so that the time lag between the current frame number t and the output probability reference frame number qs becomes large, and thus the error occurs. Is likely to increase, the reference probability b _ji is rewritten in order to reduce the error. Therefore, since the reference probability b _ji is rewritten to the output probability b _ji (x _t ) of the current frame number t, the output probability reference frame number qs is rewritten to the current frame number t. And the output probability skip number sk
Since ips represents the number of times the reference probability b _ji is not rewritten within the range of skips ≦ NSKIPS, the output probability skip number skips is initialized.

Skips ≦ NSKIPS and dts ≦ D
For TS, a case where the audio feature vector x _t of the current frame number t and the speech feature vector x _qs output probabilities reference frame number qs equal to approximately since at dts ≦ DTS, therefore the current frame number t Output probability b _ji (x
_t ) is the output probability b _ji (x of the output probability reference frame number qs
_qs ), that is, approximately equal to the reference probability b _ji . Moreover, skips ≦ NSKIPS, and therefore the reference probability b _ji
The number of times skips was not rewritten is the threshold value NSK
Since it is IPS or less, the time gap between the current frame number t and the output probability reference frame number qs is small, and therefore the possibility of increasing the error is low. So without rewriting the reference probability b _ji, it reads the reference probability b _ji by obtaining the forward probability c _it. Therefore, since the reference probability b _ji has not been rewritten, the output probability reference frame number qs is not rewritten. The output probability skip number skips is
Reference probability b _ji within the range of skips ≦ NSKIPS
Since it represents the number of times that
The number of skips skips is incremented by adding 1 to the output probability skip number skips.

As described above, in the case of skipc≤NSKIPC and dtc≤DTC, the forward probability c _it of the current frame number t is equal to the forward probability c _{i (t-1)} of the immediately preceding frame, and the forward probability c _it is calculated. And the calculation for _obtaining the output probability b _ji (x _t ) is not performed. See skipc
If> NSKIPC or dtc> DTC, sk
If ips> NSKIPS or dts> DTS, the reference probability is changed after the reference probability b _ji is rewritten, and thus the output probability b _ji (x _t ) of the current frame number t is calculated from the Hidden-Markov model. Read b _ji to obtain the forward probability c _it . Furthermore, skipc> N
If SKIPC or dtc> DTC, skip
If s ≦ NSKIPS and dts ≦ DTS, the reference probability b _ji is not rewritten, and accordingly the current frame number t
The reference probability b _ji is read out to obtain the forward probability c _it without performing the operation of obtaining the output probability b _ji (x _t ) of the forward probability c _it from the Hidden Markov model. Therefore, the error of the forward probability c _it is suppressed and the amount of calculation is reduced. Can be made.

The error of the forward probability c _{it in} this case is s
If kipc ≦ NSKIPC and dtc ≦ DTC, or skips ≦ NSKIPS and dts ≦ DTS
In the case of, the difference between the forward probability c _it obtained without performing the calculation of the output probability b _ji (x _t ) from the Hidden Markov model and the forward probability c _it obtained without such simplification of the calculation. Is.

As the thresholds DTC and DTS are increased, the amount of reduction in calculation increases, but the error in the forward probability c _it increases. Therefore, the threshold DTC, so that the forward probability c _it can be obtained within the error range practically desired,
It is necessary to determine the value of DTS.

[0103]

BEST MODE FOR CARRYING OUT THE INVENTION

<First Embodiment of Invention of Claim 1> FIG. 1 is a functional block diagram showing a configuration example of a speech recognition apparatus suitable for use in implementing the first embodiment of the invention of Claim 1.

The voice recognition apparatus 10 shown in FIG.
2, sound processing unit 14, voice section detection unit 16, collation unit 18
And a reference information storage unit 20.

The dictionary unit 12 stores a plurality of Hidden Markov models prepared for each category as standard patterns for recognition and matching. The reference information storage unit 20 stores the reference frame number qs and the reference probability b _ji .

The sound processing section 14 extracts a voice feature vector from the input voice signal for each frame of a fixed time width.
The voice section detection unit 16 detects a voice section from the input voice signal.

The collating unit 18 implements the first embodiment of the invention of claim 1, and is a time series x ₁ and x _{2 of} the voice feature vectors extracted from the start frame to the end frame of the voice section. , ..., The likelihood ln {P (x ₁ , x ₂ , ..., x _T )} between x _T and the Hidden Markov model is
The category given to the Hidden Markov model that has been obtained using the following equations (1) to (3) and has the maximum likelihood is used as the recognition result for the voice signal in the voice section.

[0108]

(Equation 5)

However, i: i = 1, 2, ..., I j: j = 1, 2, ..., J Φ _i : Probability that the initial state is S _i in the Hidden Markov model a _ji : Hidden Markov model At the state S _j to the state S _i in the above, x _t : the speech feature vector extracted in the t-th frame in the speech section (1 ≦ t ≦ T, and the first frame corresponds to the speech section). The start frame and the T-th frame represent the end frame of the speech section) b _ji (x _t ): speech feature vector x output when transitioning from state S _j to state S _i in the Hidden Markov model
Output probability of _t c _it : Time series of speech feature vector x ₁ , x ₂ , ..., Which starts transition from the initial state in Hidden Markov model
Forward probability of outputting x _t to the state S _i * i: State number i given to the state S _i that is the final state in the Hidden Markov model When storing the likelihood, it is stored in the reference information storage unit 20. Using the given reference probability b _ji , the forward probability c _it in each case of t = 1, 2, ..., T is sequentially obtained as follows.

(1). When the t = 1, is initialized to 1 reference frame number qs, all j, for i, the output probability b _ji the (x _t) determined from the hidden Markov model the output probability b _ji the (x _t) The reference probability b _ji is written as an initial value, and each reference probability b _ji is written after the reference probability b _ji is written.
And a process (1A) for obtaining the forward probability c _it ,
After the end of the process (1A), a process (1B) of adding 1 to the current frame number t is performed.

(2). When 2 ≦ t ≦ T, the voice feature vector x _t of the current frame number _t and the reference frame number qs
Of the voice feature vector x _qs of the
When S is compared with S and the comparison result is dts> DTS, the reference frame number qs is rewritten to the current frame number t, and the output probabilities b _ji (x _t ) for all j and i.
From the Hidden Markov model, the reference probability b _ji is rewritten to the output probability b _ji (x _t ), each reference probability b _ji is read out after the reference probability b _ji is rewritten, and the forward probability c _it is calculated. If There where the dts ≦ DTS, without rewriting of the reference probability b _ji reads each reference probability b _ji seek forward probability c _it processing (1C), after the processing (1C), the current frame number The process of adding 1 to t (1D) is performed.

FIG. 2 is a diagram for explaining the Hidden Markov Model. Hidden Markov Model (hereinafter, HMM) stored in the dictionary unit 12 is
Voice recognition Represents one unit of voice signal. One unit of speech recognition can be a word unit, a phoneme unit, or another unit, but here, it is a word unit. A plurality of HMMs are prepared for each category z, and the HMMs and the categories z are associated with each other and stored in the dictionary unit 12.

The HMM has a set 1 of states consisting of a total of I states S _{1 to} S _I and a set 2 of speech feature vectors x.
, A set 3 of state transition probabilities a _ji, a set 4 of output probabilities b _ji (x), a set 5 of initial state probabilities Φ _i , and a set 6 of final states F. However,

[0114]

(Equation 6)

I: i = 1, 2, ..., I j: j = 1, 2, ..., J a _ji : Probability of transition from state S _j to state S _i b _ji (x): state S _j probability speech feature vector x is output in transition to state S _i from .PHI _i: initial state in the example of the probability for example FIG. 2 is a S _i, a ₁₂ state from the state S ₁ is S ₂
Transitions to probability and b ₁₂ (x) The probability, speech feature vector x is output when a transition from the state S ₁ to state S ₂ is a ₂₂ probability and b ₂₂ transitions from state S ₂ to state S ₂ is
(x) represents the probability that the audio feature vector x is output when a transition from the state S ₂ to state S _2.

The sets 1 to 6 for defining the HMM are individually obtained for each category z by a statistical method. That is, various voice signals are collected as voice signals corresponding to the category z, for example, voice signals are collected by age or sex, or voice signals having different voicing methods are collected.
Sets 1 to 6 expressing the statistical properties of these audio signals are obtained.

The output probability b _ji (x) is expressed by using a non-correlated mixed normal distribution consisting of a plurality of normal distributions that are uncorrelated with each other, and each of these normal distributions is a function of the voice feature vector x. There is. The decorrelated mixed normal distribution has the advantage of being easy to handle mathematically and having high expressiveness.

Next, the operation of the voice recognition apparatus 10 will be described, and the flow of processing of the voice recognition method of this embodiment will be specifically described.

The sound processing section 14 uses the input voice signal to output a voice feature vector x _t = (x _t 1, x _t 2, ...) For each frame.
,, x _t p) are extracted. Where p is the voice feature vector x
order and x _t 1 to x _t p to _t represents a vector component of the speech feature vector x _t. t is a number given to the frame from which the voice feature vector x _t is extracted. HM described later
At the stage of matching with M, the frames are rewritten in ascending order with the frame number t of the starting frame of the voice section as 1, but at the time of the acoustic processing, the frame number t may be added so that each frame can be identified.

As the vector component of the voice feature vector x _t , for example, the one obtained from each filter output when the input voice signal is input to the band filter group consisting of a plurality of band pass filters having different center frequencies, the input voice Power spectrum component obtained by Fourier analysis of the signal,
Alternatively, the LPC cepstrum coefficient obtained by the linear prediction analysis, that is, the LPC analysis of the input speech signal can be used. Here, an example of extracting the voice feature vector x _t using a bandpass filter group will be described.

The acoustic processing section 14 converts the input audio signal from an analog signal to a digital signal, and outputs the converted input audio signal through a band filter group to a frequency band (channel) corresponding to each band pass filter. Separated into signal components,
The total number p of signal components x1 ...
Get xp. Next, the acoustic processing unit 14 rectifies the signal component x1 and obtains an average value of the rectified signal component x1 (absolute value of the signal component x1) in frame units. This average value is obtained by dividing the rectified signal component x1 by the time width of one frame. The average value of the signal component x1 obtained in the t-th frame is extracted as component x _t 1 of the audio feature vector x _t. Similarly, from the remaining signal components x2 ～Xp, it extracts the component x _t 2~x _t p of a speech feature vector x _t.

Next, the voice section detecting section 16 includes the sound processing section 1
Based on the voice feature vector x _t from 4, the start frame and the end frame of the voice section are detected, and the section information indicating which frame is the start frame and the end frame of the voice section is generated. The voice section is a section in which a voice signal for one unit of voice recognition is included here.

The collation unit 18 inputs the section information and the speech feature vector x _t from the speech section detection unit 16, and the time series x of the speech feature vector x _t extracted from the start frame to the end frame of the speech section. Generate ₁ , x ₂ , ..., X _T. At this time, the frame number t of the start frame is set to 1, and the frame numbers t from the start frame to the end frame of the voice section are rewritten in ascending order.

The collation unit 18 then calculates the vector time series x ₁ ,
x ₂ , ..., X _T and the HMM stored in the dictionary unit 12
And the likelihood ln {P (x ₁ , x ₂ , ..., X _T )} between
The category z given to each HMM of the dictionary unit 12 is obtained individually, and the category z assigned to the HMM having the maximum likelihood is output as a recognition result.

Here, P (x ₁ , x represented by the equation (1)
₂ , ..., x _T ) is the vector time series x in the HMM.
It is the probability that ₁ , x ₂ , ..., x _T will appear.

[0126]

(Equation 7)

The c _iT in the equation (1) starts transition from the initial state in the HMM and the vector time series x ₁ , x ₂ , ...
..., the forward probability of outputting x _T to reach the state S _i , * i is i (the number _i assigned to the state S _i belonging to the final state F) that satisfies S _i εF, and thus i = The maximum forward probability c _iT among the forward probabilities c _iT of * i is _defined as the appearance probability P (x ₁ , x ₂ , ..., X _T ).

The forward probability c _iT is approximately obtained by the Viterbi algorithm using the recurrence formulas shown in equations (2) to (3). c _i0 = Ф _i (2)

[0129]

(Equation 8)

In the HMM, there is one or more state transitions that output the voice feature vector x _t . Therefore, there is one or a plurality of transition paths from the initial state to output the vector series x _{1 to} x _t to reach the state S _i , and in most cases, there are a plurality of transition paths. Therefore, as shown in Expression (3), the maximum c _{j (t-1)} a _ji b _ji (x of the c _{j (t-1)} a _ji b _ji (x _t ) calculated for each transition path is obtained. Let _t ) be the forward probability c _it . This calculation method is called the Viterbi method.

The output probability b _ji (x _t ) in the equation (3) is defined as in the following equation (4).

[0132]

(Equation 9)

However, m = 1, 2, ..., M g _jim (x _t ): A voice feature vector calculated from the m-th normal distribution in the uncorrelated mixed normal distribution consisting of M normal distributions. Weighting probability of x _t The weighting probability g _jim (x _t ) in the expression (4) is expressed by the following expression (5)-
It is expressed using (7).

G _jim (x _t ) = λ _jim b _jim (x _t ) ... (5) b _jim (x _t ) = (2π) ^{−p / 2} _{│ρ jim} │ ^-1/2 exp {−D _jimt ^{2/2} ...... (6)} D jimt 2 = (x t -μ jim) 'ρ jim -1 (x t -μ jim) ...... (7) λ jim: weight b _jim of the m-th normal distribution (x _t ): _Unweighted probability of the speech feature vector x _t calculated from the m-th normal distribution ρ _jim : _Covariance / covariance matrix of the m-th normal distribution μ _jim : _{Mean of} the m-th normal distribution Vector D _jimt : Mahalabis's general distance (x _t −μ _jim ) ′: transposed matrix of (x _t −μ _jim ), which represents the distance between the voice feature vector x _t and the m-th normal distribution. Various types of b _ji (x _t ) can be used, and in addition to the formula (4), for example, a formula defined as the following formula (8) may be used. Equation (8) is the total number M
Number of regular weights probability in the distribution uncorrelated Gaussian Mixture consisting calculated from individual normal distribution g _ijm largest weighted probability g _jim (x _t) of the (x _t), the output probability b _ji (x _t)
It means to detect as.

[0135]

(Equation 10)

Further logarithmic transition probability A _ji = ln
(A _ji ), logarithmic output probability B _ji (x _t ) = ln {b
_ji (x _t )} and the logarithmic forward probability C _it = ln
When expressed as (c _it ), the likelihoods are modified by modifying equations (1) to (3).
Equations (9) to (11) relating to the calculation of ln {P (x ₁ , x ₂ , ..., X _t )} are obtained.

[0137]

[Equation 11]

Since equations (9) to (11) are recurrence equations of t, the logarithmic forward probability C _it when t = 1, 2, ..., T is expressed by the following equations (12) to (12). It can be calculated sequentially as in 16).

[0139]

(Equation 12)

The HMM matching unit 18 uses i = 1, 2, ... I
Logarithmic forward probability C _iT of t = T for all i in
, The maximum C _iT among the logarithmic forward probabilities C _iT with i = * i is _calculated as the likelihood ln {P (x ₁ , x ₂ , ..., X
_T )}. For all the HMMs stored in the dictionary unit 12, the likelihood ln {P (x ₁ ,
x ₂ , ..., x _T )} is obtained and the maximum likelihood is obtained.
The category z given to the time series x ₁ , x
₂ , ..., x _T is output as the recognition result for the input voice signal.

Next, in the first embodiment of the invention of claim 1, time series x ₁ , x ₂ , ... Of HMM and speech feature vector.
,, the flow of the process of obtaining the likelihood between x _T and 1
The flow of processing focused on individual HMMs will be described. FIG. 3 and FIG. 4 are diagrams showing the flow of processing focusing on this one HMM. In this example, the output probability b _ji (x _t ), the forward probability c _it, and the reference probability b _ji are logarithmized output probability B _ji (x _t ), logarithmic forward probability C _it, and logarithmic reference probability, respectively. B _ji and i = j = 1, 2, ..., I.

When the section information and the speech feature vector x _t are input from the speech section detection unit 16, the collation unit 18 sets i =
For all i of 1, 2, ..., I, the initial value C _i0 of the forward probability logarithmized is set according to the equation (10) (S).
1).

Next, the collation unit 18 initializes the current frame number t to t = 1 in order to process the start frame of the voice section (S2).

Next, the collation unit 18 determines that j = 1, 2, ..., J.
And i = 1, 2, ..., I, logarithmized output probabilities B _ji (x ₁ ) are obtained according to equations (4) to (7) (S3), and the output probabilities B _{ji are obtained.} (x ₁ ) is written as the initial value of the logarithmic reference probability B _ji (S4).

In the reference information storage section 32, j = 1, 2, ...
, J and i = 1, 2, ..., I are respectively provided with storage areas save B _ji for storing the reference probabilities B _ji . Therefore, the reference information storage unit 32 stores B ₁₁ , B ₁₂ , ...
_{..., B 1I, B 21,} B 22, ......, B 2I, ......, B J1,
B _J2 , ..., B _JI respectively have J × I storage areas for individually storing. Therefore, in the figure, the reference probability B _ji
The process of storing the initial value of is expressed as save B _ji = B _ji (x ₁ ).

Next, the collation unit 18 determines the reference frame number qs.
Is initialized to the current frame number 1 (S5), and then i =
Logarithmic forward probability C _i1 is obtained for all i of 1, 2, ..., I according to equation (11) (S6).

Next, the collation unit 18 adds 1 to the current frame number t in order to process the next frame of the voice section (S7), and thereafter, the current frame number t and the frame number T of the end frame are compared. Are compared to determine whether the processing has been completed for all the frames in the voice section (S8).

(1-1A: When t≤T in S8) If the current frame number t is equal to or less than the end frame number T in S8, the processing has not been completed for all the frames in the voice section, so the comparison is performed. part 18 a distance dts between the speech feature vector x _qs speech feature vector x _t and the reference frame number qs of the current frame number t, determined according to the following equation (17) (S9).

[0149]

(Equation 13)

However, x _t k: vector component of the voice feature vector x _t of the current frame number t x _qs k: voice feature vector x _qs of the reference frame number _qs
Next, the matching unit 18 compares the distance dts with the threshold value DTS and determines whether or not these vectors x _t and x _qs are approximately equal (S10).

[0151] When the S10 by the distance dts exceeds the threshold DTS, the output probability of the speech feature vector x _t and the reference frame number qs speech feature vector x _qs and does not approximate Thus current frame number t of the current frame number t since B _ji (x _t) can not be approximated by the reference probability B _ji, so that the rewriting of the reference probability B _ji. Therefore, the reference frame number qs is rewritten to the current frame number t (S11). After that, j =
, J, and i = 1, 2, ..., I, for all j and i, logarithmic output probabilities B _ji (x _t ) are obtained according to equations (4) to (7), The reference probability B _ji is rewritten to the output probability B _ji (x _t ) (S12). After the reference probability B _ji has been rewritten, each reference probability B _ji is read, and i =
Forward probability C _it for all i of 1, 2, ..., I
Is calculated according to the equation (11) (S13). After that, the process returns to S7 so as to perform the process for the next frame of the voice section. The process of rewriting the reference probability B _{ji in} S12 is represented as save B _ji = B _ji (x _t ) in the figure.

Reference probability B read in S13 in this case
_ji is the output probability B _ji (x _t ) of the current frame number t obtained in S12. Therefore, in S13 of this case, the forward probability C is calculated using the output probability B _ji (x _t ) of the current frame number t.
_It will ask for it.

If the distance dts is less than or equal to the threshold value DTS in S10, the voice feature vector x _t of the current frame number t is approximately equal to the voice feature vector x _qs of the reference frame number qs. Output probability B _ji
Because (x _t) is equal approximately to a reference probability B _ji, rewriting of the reference probability B _ji becomes not performed. Therefore, the reference probabilities B _ji are read out without calculating the output probabilities B _ji (x _t ) using the equations (4) to (7), and all i of i = 1, 2, ... The forward probability C _it which is logarithmized is _expressed by the equation (11).
(S13). After that, the process returns to S7 so as to perform the process for the next frame of the voice section.

Reference probability B read in S13 in this case
_ji is the output probability B _ji (x _qs ) obtained in the frame of the reference frame number qs. Therefore, in S13 in this case, the forward probability C _it is used by using the output probability B _ji (x _qs ) of the reference frame number qs. Will be asked.

(1-1B: When t> T in S8) When the current frame number t is larger than the frame number T of the end frame in S8, i = 1, 2, ... since finished sought forward probability C _iT, wherein i = * i maximum forward probability C _iT of the forward probability C _iT comprising the following (9), the speech feature vector time series _{x 1, x 2, ......,} x
Likelihood ln {P (x ₁ , x ₂ , ..., X between _T and HMM
_T )}, and after that, the process of calculating the likelihood for the HMM is ended (end).

The collation unit 18 obtains the likelihood (forward probability C _iT ) by performing the processes of S1 to S13 shown in FIGS. 3 and 4 for all HMMs stored in the dictionary unit 12. , And the category of the HMM having the maximum likelihood, the time series x ₁ , x ₂ , ... Of the speech feature vector.
, X _T are output to a device (not shown) in the next stage as a recognition result for the extracted input voice signal.

As described above, the likelihood ln {P (x ₁ , x ₂ , ...
, X _T )} = C _iT in the process of obtaining the distance dts
Is less than or equal to the threshold value DTS, the forward probability C _it is calculated without performing the calculation of the output probabilities B _ji (x _t ) from the equations (4) to (7), so that the amount of calculation can be significantly reduced. Moreover, the simplification of such calculation is that the distance dts is the threshold value DTS.
Since it is performed in the following case, the error of the forward probability C _it can be reduced even if the calculation is simplified.

According to the simulation result of the inventor of this application, the calculation amount for _obtaining the output probability B _ji (x _t ) is
Example in which there is no significant difference in recognition accuracy of voice recognition between the case where the threshold value DTS is set to be about 1/5 of the case where the calculation is not simplified and the case where the calculation is not simplified There were many.

<Second Embodiment of the Invention of Claim 1> As a voice recognition apparatus suitable for carrying out the second embodiment of the invention of claim 1, the collating unit 18 is configured as follows. It is possible to use the voice recognition device 10 having the same configuration as that described above.

That is, the collation unit 18 uses the reference probability b _ji stored in the reference information storage unit 20 when calculating the likelihood, and the forward probability in each case of t = 1, 2, ..., T. c
_It is calculated sequentially as follows.

(1). When t = 1, the reference frame number qs is initialized to 1, the skip number skips is initialized to 0, and the output probabilities b _ji (x
_t ) is calculated from the Hidden Markov model and the output probability b _ji (x
_t ) is written as the initial value of the reference probability b _ji , and the reference probability b
After the writing of _ji is finished, a process (1A) of reading out each reference probability b _ji to obtain a forward probability c _it and a process (1B) of adding 1 to the current frame number t after the process (1A) are completed. .

(2). 2 ≦ t when the ≦ T, together with comparing the number of skips skips a threshold NSKIPS, threshold DTS distance dts between the speech feature vector x _qs speech feature vector x _t and the reference frame number qs of the current frame number t And the comparison result is skips> NSK.
When IPS or dts> DTS, the skip number skips is initialized to 0, the reference frame number qs is rewritten to the current frame number t, and the output probabilities b _ji (x _t ) are set for all j and i. the reference probability b _ji determined from hidden Markov models rewritten to the output probability b _ji (x _t), determine the forward probability c _it reads each reference probability b _ji after rewriting completion of the reference probability b _ji, the comparison result There when the skips ≦ NSKIPS and dts ≦ DTS, as well as adding 1 to the skip number skips, reference probability b _ji each reference probability b _ji without rewriting the
And a process (1C) for obtaining the forward probability c _it ,
After the end of the process (1C), a process (1D) of adding 1 to the current frame number t is performed.

Next, in the second embodiment of the invention as claimed in claim 1, time series x ₁ , x ₂ , ... Of HMM and voice feature vector.
,, the flow of the process of obtaining the likelihood between x _T and 1
The flow of processing focused on individual HMMs will be described. FIG. 5 and FIG. 6 are diagrams showing the flow of processing focusing on this one HMM. In this example, the output probability b _ji (x _t ), the forward probability c _it, and the reference probability b _ji are logarithmized output probability B _ji (x _t ), logarithmic forward probability C _it, and logarithmic reference probability, respectively. B _ji and i = j = 1, 2, ..., I.

When the section information and the speech feature vector x _t are input from the speech section detecting section 16, the collation section 18 sets i =
For all i of 1, 2, ..., I, the initial value C _i0 of the forward probability logarithmized is set according to the equation (10) (S).
1).

Next, the collation unit 18 initializes the current frame number t to t = 1 in order to process the start frame of the voice section (S2).

Next, the collation unit 18 uses j = 1, 2, ..., J.
And i = 1, 2, ..., I, logarithmized output probabilities B _ji (x ₁ ) are obtained according to equations (4) to (7) (S3), and the output probabilities B _{ji are obtained.} (x ₁ ) is written as an initial value of the logarithmic output probability B _ji (S4).

In the reference information storage section 32, j = 1, 2, ...
, J and i = 1, 2, ..., I, respectively, save B _ji is provided as a storage area for storing the reference probability B _ji for each j and i. Therefore, the reference information storage unit 32 stores B ₁₁ , B ₁₂ , ...
_{..., B 1I, B 21,} B 22, ......, B 2I, ......, B J1,
B _J2 , ..., B _JI respectively have J × I storage areas for individually storing. Therefore, in the figure, the reference probability B _ji
The process of storing the initial value of is expressed as save B _ji = B _ji (x ₁ ).

Next, the collating section 18 determines the reference frame number qs.
To the current frame number 1 and skip number sk
ips is initialized to 0 (S5). After that, i = 1,
The logarithmic forward probability C _i1 is _calculated for all i of 2, ..., I according to the equation (11) (S6).

Next, the collation unit 18 adds 1 to the current frame number t in order to process the next frame in the voice section (S7), and thereafter, the current frame number t and the frame number T of the end frame. Are compared to determine whether the processing has been completed for all the frames in the voice section (S8).

(1-2A: When t ≦ T in S8) If the current frame number t is less than or equal to the end frame number T in S8, the processing has not been completed for all the frames in the voice section, so skip. Number skips and threshold NSKIPS
A comparison judgment with is performed (S9).

In S9, the skip number skips is the threshold NSK.
When the IPS is exceeded, the current frame number t and the reference frame number qs have a large time gap, and thus the error is likely to increase. Therefore, in order to reduce the error, the reference probability B _ji is reduced.
Will be rewritten. Therefore skip number ski
The ps is initialized to 0 and the reference frame number qs is rewritten to the current frame number t (S10). After that, j =
, J, and i = 1, 2, ..., I, for all j and i, logarithmic output probabilities B _ji (x _t ) are obtained according to equations (4) to (7), The reference probability B _ji is rewritten to the output probability B _ji (x _t ) (S11). This reference probability B
After the completion of rewriting _ji , each reference probability B _ji is read, and i =
Forward probability C _it for all i of 1, 2, ..., I
Is calculated according to the equation (11) (S12). After that, the process returns to S7 so as to perform the process for the next frame of the voice section. The process of rewriting the reference probability B _{ji in} S11 is represented as save B _ji = B _ji (x _t ) in the figure.

Reference probability B read in S12 in this case
_ji is the output probability B _ji (x _t ) of the current frame number t obtained in S11. Therefore, in S12 of this case, the forward probability C is calculated using the output probability B _ji (x _t ) of the current frame number t.
_It will ask for it.

In step S9, the skip number skips is the threshold value NSK.
If the IPS or less, the matching unit 18 obtains the distance dts between the speech feature vector x _qs speech feature vector x _t and the reference frame number qs of the current frame number t (S1
3) The obtained distance dts is compared with the threshold value DTS to determine whether or not these vectors x _t and x _qs are approximately equal (S14).

[0174] When the S14 by the distance dts exceeds the threshold DTS, the output probability of the speech feature vector x _t and the reference frame number qs speech feature vector x _qs and does not approximate Thus current frame number t of the current frame number t since B _ji (x _t) can not be approximated by the reference probability B _ji, so that the rewriting of the reference probability B _ji. Therefore, the processes of S10 to S12 are performed, and thereafter, the process returns to the process of S7 to perform the process for the next frame of the voice section.

If the distance dts is less than or equal to the threshold value DTS in S14, the voice feature vector x _t of the current frame number _t.
Is approximately equal to the speech feature vector x _qs of the reference frame number qs, and thus the output probability B _ji (x _t ) of the current frame number t
Is approximately equal to the reference probability B _ji , the reference probability B
_ji will not be rewritten. Therefore, 1 is added to the skip number skips to count up the skip number skips (S15), and then the output probability B _ji (x _t ) is not calculated using the equations (4) to (7), Read out the reference probability B _ji, and for all i of i = 1, 2, ..., I,
The logarithmic forward probability C _it is calculated according to the equation (11) (S12). After that, the process returns to S7 so as to perform the process for the next frame of the voice section.

Reference probability B read in S12 in this case
_ji is the output probability B _ji (x _t ) obtained in the frame of the reference frame number qs. Therefore, in S12 in this case, the forward probability C _it is used by using the output probability B _ji (x _qs ) of the reference frame number qs. Will be asked.

(1-2B: When t> T in S8) When the current frame number t is larger than the frame number T of the end frame in S8, i = 1, 2, ... since finished sought forward probability C _iT, the maximum forward probability C _iT of i = * i consisting forward probability C _iT according to equation (9), the time series x _1, x ₂ of the speech feature vector, ...,
Likelihood ln {P (x ₁ , x ₂ , ..., Between x _T and HMM
x _T )}, and after that, the process of calculating the likelihood for the HMM is terminated (end).

The collating unit 18 performs the processes of S1 to S15 shown in FIGS. 5 and 6 for all the HMMs stored in the dictionary unit 12 to calculate the likelihood (forward probability C).
_iT ), and the maximum likelihood is detected from the calculated likelihoods.
The largest category of likelihood the resulting HMM, time series x _1, x ₂ of the audio feature vector, ..., as the recognition result for the input speech signal obtained by extracting the x _T, without the next stage of the device (shown ).

As described above, the likelihood ln {P (x ₁ , x ₂ , ...
, X _T )} = C _iT , the number of skips skips becomes equal to or less than the threshold value NSKIPS and the distance dt
When s is equal to or less than the threshold value DTS, the forward probability C _it is calculated without performing the calculation of the output probability B _ji (x _t ) from the equations (4) to (7), so that the calculation amount can be significantly reduced. . Moreover, such a simplification of the calculation is based on the skip number skips.
Is less than or equal to the threshold NSKIPS and the distance dts is the threshold D
Since it is performed when it is equal to or smaller than TS, the error of the forward probability C _it can be reduced even if the calculation is simplified.

The invention of claim 1 can be applied to all speech recognition apparatuses that perform matching processing in frame units.

<First Embodiment of Invention of Claim 3> FIG. 7 is a functional block diagram showing an example of the configuration of a voice recognition apparatus suitable for carrying out the first embodiment of the invention of Claim 3.

The voice recognition device 22 shown in FIG.
4, sound processing unit 26, voice section detection unit 28, collation unit 30
And a reference information storage unit 32.

The dictionary unit 24 stores a plurality of Hidden Markov models prepared for each category as standard patterns for recognition and matching. In the Hidden Markov model, the state S _j that is the transition source of the state transition that gives the output probability b _ji (x) of the voice feature vector x is given a type s of either a stationary part or a transient part. The reference information storage unit 32 stores the reference frame number qs for the stationary part and the reference frame number q for the transient part.
Store t and the reference probability b _ji .

The sound processing unit 26 extracts a voice feature vector from the input voice signal for each frame of a fixed time width.
The voice section detection unit 28 detects a voice section from the input voice signal.

The collating unit 30 implements the first embodiment of the invention of claim 3, and is a time series x ₁ and x _{2 of} the voice feature vectors extracted from the start frame to the end frame of the voice section. , ..., The likelihood ln {P (x ₁ , x ₂ , ..., x _T )} between x _T and the Hidden Markov model is
The category given to the Hidden Markov model that has been obtained using the following equations (1) to (3) and has the maximum likelihood is used as the recognition result for the voice signal in the voice section.

[0186]

[Equation 14]

However, i: i = 1, 2, ..., I j: j = 1, 2, ..., J Φ _i : Probability that the initial state is S _i in the Hidden Markov model a _ji : Hidden Markov model At the state S _j to the state S _i in the above, x _t : the speech feature vector extracted in the t-th frame in the speech section (1 ≦ t ≦ T, and the first frame corresponds to the speech section). The start frame and the T-th frame represent the end frame of the speech section) b _ji (x _t ): speech feature vector x output when transitioning from state S _j to state S _i in the Hidden Markov model
Output probability of _t c _it : Time series of speech feature vector x ₁ , x ₂ , ..., Which starts transition from the initial state in Hidden Markov model
Forward probability of outputting x _t to the state S _i * i: State number i assigned to the state S _i that is the final state in the Hidden Markov model When storing the likelihood, it is stored in the reference information storage unit 32. The forward probability c _it in each case of t = 1, 2, ..., T is sequentially obtained using the reference probability b _ji that has been described as follows.

(1). When t = 1, the stationary part reference frame number qs and the transient part reference frame number qt are initialized to 1, and the output probabilities b _ji (x _t ) are obtained from the Hidden Markov model for all j and i. writing the output probability b _ji the (x _t) as the initial value of the reference probability b _ji, performs processing for calculating the forward probability c _it reads each reference probability b _ji after completion of writing of the reference probability b _ji the (2A). After the processing (2A) is completed, processing (2B) for adding 1 to the current frame number t is performed.

(2). When the 2 ≦ t ≦ T, the distance dts between the speech feature vector _qs speech feature vector x _t and the constant part reference frame number qs of the current frame number t is compared with a threshold DTS, the comparison result is dts> if the DTS, the process of rewriting the constant part reference frame number qs to the current frame number t and (2C), the speech feature vector x _qt speech feature vector x _t and the transient portion reference frame number qt the current frame number t If the distance dtt is compared with the threshold value DTT and the comparison result is dtt> DTT,
A process (2D) for rewriting the transition part reference frame number qt to the current frame number t is performed, and after completion of these processes (2C) and (2D), j = 1, 2, ... A process (2E) of determining the type s _assigned to the transition source S _j of the state transition that gives the output probability b _ji (x _t ) is performed.

When the type determination result of the process (2E) is the stationary part, the comparison result of the process (2C) is dts>
In the case of DTS, for j for which the type determination result is obtained, the output probability b _ji (x _t ) is obtained from the Hidden Markov model for all i, and the reference probability b _ji is the output probability b _ji (x _t ).
When the type determination result of the process (2E) is a stationary part, the comparison result of the process (2C) is dts ≦ DTS.
In this case, the reference probability b _ji is not rewritten for j for which the type determination result is obtained, and when the type determination result of the process (2E) is the transition part, the comparison result of the process (2D) is dtt. > DTT, for j for which the type determination result is obtained, for all i, output probabilities b _ji (x _t ) are obtained from Hidden Markov models and reference probabilities b _ji are output probabilities b _ji (x _t ). When the type determination result of the process (2E) is a transition part, the comparison result of the process (2D) is changed to dtt.
If ≦ DTT, a process (2F) of not rewriting the reference probability b _ji is performed for j that has obtained the type determination result.

Then, the processing (2F) is performed for each j of j = 1, 2, ..., J, and the processing (2F) is performed for all j.
After the above, the reference probability b _ji is read out and the forward probability c _it is calculated (2G). After the processing (2G) is completed, the processing (2H) of adding 1 to the current frame number t is performed.

FIG. 8 is a diagram for explaining the Hidden Markov Model. The Hidden Markov Model (hereinafter, HMM) stored in the dictionary unit 24 is
Voice recognition Represents one unit of voice signal. One unit of speech recognition can be a word unit, a phoneme unit, or another unit, but here, it is a word unit. A plurality of HMMs are prepared for each category z, and the HMM and the category z are associated with each other and stored in the dictionary unit 24.

The HMM has a set 1 of states consisting of a total of I states S _{1 to} S _I and a set 2 of speech feature vectors x.
, A set 3 of state transition probabilities a _ji, a set 4 of output probabilities b _ji (x), a set 5 of initial state probabilities Φ _i , and a set 6 of final states F. In the HMM, the transition source S _{j of the} state transition that gives the output probability b _ji (x) is given a type s of either a steady part or a transient part. However,

[0194]

(Equation 15)

I: i = 1, 2, ..., I j: j = 1, 2, ..., J a _ji : Probability of transition from state S _j to state S _i b _ji (x): state S _j probability speech feature vector x is output in transition to state S _i from .PHI _i: initial state in the example of the probability for example FIG. 2 is a S _i, a ₁₂ state from the state S ₁ is S ₂
Transitions to probability and b ₁₂ (x) The probability, speech feature vector x is output when a transition from the state S ₁ to state S ₂ is a ₂₂ probability and b ₂₂ transitions from state S ₂ to state S ₂ is
(x) represents the probability that the audio feature vector x is output when a transition from the state S ₂ to state S _2. Furthermore, the output probability b ₁₁
For the transition source S _{1 of the} state transition S ₁ → S ₁ that gives (x), the transient part as the type s, and for the state transition S ₁ → S ₂ that gives the output probability b ₁₂ (x), , The stationary part is added as the type s.

Sets 1 to 6 for defining the HMM are individually obtained for each category z by a statistical method. That is, various voice signals are collected as voice signals corresponding to the category z, for example, voice signals are collected by age or sex, or voice signals having different voicing methods are collected.
Sets 1 to 6 expressing the statistical properties of these audio signals are obtained. At this time, it is also checked whether the state transition that gives the output probability b _ji (x) is the steady part or the transient part of the audio signal,
A type s of either a steady part or a transient part is given to the transition source S _{j of the} state transition.

The output probability b _ji (x) is expressed by using a non-correlated mixed normal distribution consisting of a plurality of normal distributions that are uncorrelated with each other, and each of these normal distributions is a function of the speech feature vector x. There is. The decorrelated mixed normal distribution has the advantage of being easy to handle mathematically and having high expressiveness.

Next, the operation of the voice recognition device 22 will be described, and the flow of processing of the voice recognition method of this embodiment will be specifically described.

The sound processing unit 26, from the input speech signal, the speech feature vector x _t = (x _t 1, x _t 2, ...) For each frame.
,, x _t p) are extracted. Where p is the voice feature vector x
order and x _t 1 to x _t p to _t represents a vector component of the speech feature vector x _t. t is a number given to the frame from which the voice feature vector x _t is extracted. HM described later
At the stage of matching with M, the frames are rewritten in ascending order with the frame number t of the starting frame of the voice section as 1, but at the time of the acoustic processing, the frame number t may be added so that each frame can be identified.

The vector component of the voice feature vector x _t is, for example, one obtained from each filter output when an input voice signal is input to a band filter group consisting of a plurality of band pass filters having different center frequencies, or input voice signals. Power spectrum component obtained by Fourier analysis of the signal,
Alternatively, the LPC cepstrum coefficient obtained by the linear prediction analysis, that is, the LPC analysis of the input speech signal can be used. Here, an example of extracting the voice feature vector x _t using a bandpass filter group will be described.

The acoustic processing unit 26 converts the input audio signal from an analog signal to a digital signal, and outputs the converted input audio signal through a band filter group to a frequency band (channel) corresponding to each band pass filter. Separated into signal components,
The total number p of signal components x1 ...
Get xp. Next, the acoustic processing unit 26 rectifies the signal component x1 and obtains an average value of the rectified signal component x1 (absolute value of the signal component x1) in frame units. This average value is obtained by dividing the rectified signal component x1 by the time width of one frame. The average value of the signal component x1 obtained in the t-th frame is extracted as component x _t 1 of the audio feature vector x _t. Similarly, from the remaining signal components x2 ～Xp, it extracts the component x _t 2~x _t p of a speech feature vector x _t.

Next, the voice section detecting section 28 uses the sound processing section 2
Based on the voice feature vector x _t from 6, the start frame and the end frame of the voice section are detected, and the section information indicating which frame is the start frame and the end frame of the voice section is generated. The voice section is a section in which a voice signal for one unit of voice recognition is included here.

The collating unit 30 inputs the section information and the voice feature vector x _t from the voice section detecting unit 28, and the time series x of the voice feature vector x _t extracted from the start frame to the end frame of the voice section. Generate ₁ , x ₂ , ..., X _T. At this time, the frame number t of the start frame is set to 1, and the frame numbers t from the start frame to the end frame of the voice section are rewritten in ascending order.

Then, the matching unit 30 calculates the vector time series x ₁ ,
x ₂ , ..., X _T and the HMM stored in the dictionary unit 24
And the likelihood ln {P (x ₁ , x ₂ , ..., X _T )} between
The category z given to each HMM of the dictionary unit 24 is individually obtained, and the category z assigned to the HMM having the maximum likelihood is output as a recognition result.

Here, P (x ₁ , x shown in equation (1)
₂ , ..., x _T ) is the vector time series x in the HMM.
It is the probability that ₁ , x ₂ , ..., x _T will appear.

[0206]

(Equation 16)

C _iT in the equation (1) is a vector time series x ₁ , x ₂ , ...
..., the forward probability of outputting x _T to reach the state S _i , * i is i (the number _i assigned to the state S _i belonging to the final state F) that satisfies S _i εF, and thus i = The maximum forward probability c _iT among the forward probabilities c _iT of * i is _defined as the appearance probability P (x ₁ , x ₂ , ..., X _T ).

The forward probability c _iT is approximately obtained by the Viterbi algorithm using the recurrence formulas shown in the equations (2) to (3). c _i0 = Ф _i (2)

[0209]

[Equation 17]

In the HMM, there are one or more state transitions that output the voice feature vector x _t . Therefore, there is one or a plurality of transition paths from the initial state to output the vector series x _{1 to} x _t to reach the state S _i , and in most cases, there are a plurality of transition paths. Therefore, as shown in Expression (3), the maximum c _{j (t-1)} a _ji b _ji (x of the c _{j (t-1)} a _ji b _ji (x _t ) calculated for each transition path is obtained. Let _t ) be the forward probability c _it . This calculation method is called the Viterbi method.

The output probability b _ji (x _t ) in the equation (3) is defined as the following equation (4).

[0212]

(Equation 18)

However, m = 1, 2, ..., M g _jim (x _t ): A speech feature vector calculated from the m-th normal distribution in the uncorrelated mixed normal distribution consisting of M normal distributions. Weighting probability of x _t The weighting probability g _jim (x _t ) in the expression (4) is expressed by the following expression (5)-
It is expressed using (7).

G _jim (x _t ) = λ _jim b _jim (x _t ) ... (5) b _jim (x _t ) = (2π) ^{−p / 2} _{│ρ jim} │ ^-1/2 exp {−D _jimt ^{2/2} ...... (6)} D jimt 2 = (x t -μ jim) 'ρ jim -1 (x t -μ jim) ...... (7) λ jim: weight b _jim of the m-th normal distribution (x _t ): _Unweighted probability of the speech feature vector x _t calculated from the m-th normal distribution ρ _jim : _Covariance / covariance matrix of the m-th normal distribution μ _jim : _{Mean of} the m-th normal distribution Vector D _jimt : Mahalabis's general distance (x _t −μ _jim ) ′: transposed matrix of (x _t −μ _jim ), which represents the distance between the voice feature vector x _t and the m-th normal distribution. Various types of b _ji (x _t ) can be used, and in addition to the formula (4), for example, a formula defined as the following formula (8) may be used. Equation (8) is the total number M
Number of regular weights probability in the distribution uncorrelated Gaussian Mixture consisting calculated from individual normal distribution g _jim largest weighted probability g _jim (x _t) of the (x _t), the output probability b _ji (x _t)
It means to detect as.

[0215]

[Equation 19]

Further logarithmic transition probability A _ji = ln
(A _ji ), logarithmic output probability B _ji (x _t ) = ln {b
_ji (x _t )} and the logarithmic forward probability C _it = ln
When expressed as (c _it ), the likelihoods are modified by modifying equations (1) to (3).
Equations (9) to (11) relating to the calculation of ln {P (x ₁ , x ₂ , ..., X _t )} are obtained.

[0219]

(Equation 20)

Since equations (9) to (11) are recurrence equations of t, the logarithmic forward probability C _it when t = 1, 2, ... It can be calculated sequentially as in 16).

[0219]

(Equation 21)

The HMM matching unit 30 uses i = 1, 2, ...,
Logarithmic forward probability C _{iT of} t = T for all i in I
, The maximum C _iT among the logarithmic forward probabilities C _iT with i = * i is _calculated as the likelihood ln {P (x ₁ , x ₂ , ..., X
_T )}. For all the HMMs stored in the dictionary unit 24, the likelihood ln {P (x ₁ ,
x ₂ , ..., x _T )} is obtained and the maximum likelihood is obtained.
The category z given to the time series x ₁ , x
₂ , ..., x _T is output as the recognition result for the input voice signal.

Next, in the first embodiment of the invention of claim 3, time series x ₁ , x ₂ , ... Of HMM and speech feature vector.
,, the flow of the process of obtaining the likelihood between x _T and 1
The flow of processing focused on individual HMMs will be described. 9 to 11 are diagrams showing the flow of processing focused on this one HMM. In this example, the output probability b _ji (x _t ), the forward probability c _it, and the reference probability b _ji are logarithmized output probability B _ji (x _t ), logarithmic forward probability C _it, and logarithmic reference probability, respectively. B _ji and i = j = 1, 2, ..., I.

When the section information and the speech feature vector x _t are inputted from the speech section detecting section 28, the collating section 30 receives i =
For all i of 1, 2, ..., I, the initial value C _i0 of the forward probability logarithmized is set according to the equation (10) (S).
1).

Next, the collating unit 30 initializes the current frame number t to t = 1 in order to process the start frame of the voice section (S2).

Next, the matching unit 30 determines that j = 1, 2, ..., J.
And i = 1, 2, ..., I, logarithmized output probabilities B _ji (x ₁ ) are obtained according to equations (4) to (7) (S3), and the output probabilities B _{ji are obtained.} (x ₁ ) is written as the initial value of the logarithmic reference probability B _ji (S4).

In the reference information storage section 32, j = 1, 2, ...
, J and i = 1, 2, ..., I are respectively provided with storage areas save B _ji for storing the reference probabilities B _ji . Therefore, the reference information storage unit 32 stores B ₁₁ , B ₁₂ , ...
_{..., B 1I, B 21,} B 22, ......, B 2I, ......, B J1,
B _J2 , ..., B _JI respectively have J × I storage areas for individually storing. Therefore, in the figure, the reference probability B _ji
The process of storing the initial value of is expressed as save B _ji = B _ji (x ₁ ).

Next, the collating unit 30 initializes the stationary part reference frame number qs and the transient part reference frame number qt to the current frame number 1 (S5), and after that, i = 1,
The logarithmic forward probability C _i1 is _calculated for all i of 2, ..., I according to the equation (11) (S6).

Next, the collating unit 30 adds 1 to the current frame number t in order to process the next frame of the voice section (S7), and thereafter, the current frame number t and the frame number T of the end frame are compared. Are compared to determine whether the processing has been completed for all the frames in the voice section (S8).

(2-1A: When t ≦ T in S8) If the current frame number t is equal to or less than the end frame number T in S8, the processing has not been completed for all the frames in the voice section, so the comparison is performed. part 30 a distance dts between the speech feature vector x _qs speech feature vector x _t and the constant part reference frame number qs of the current frame number t, determined according to the following equation (17) (S9).

[0229]

(Equation 22)

However, x _t k: vector component of the voice feature vector x _t of the current frame number t x _qs k: vector component of the voice feature vector x _qs of the constant part reference frame number qs Dts and threshold D related to
It is determined whether these vectors x _t and x _qs are approximately equal by comparing with TS (S10).

If the distance dts exceeds the threshold value DTS in S10, the voice feature vector x _t of the current frame number t does not approximate to the voice feature vector x _qs of the stationary part reference frame number qs, and thus the voice of the current frame number t is Feature vector x
_t is the voice feature vector x of the stationary part reference frame number qs
_Since the change from _qs is large, the reference frame number q
s is rewritten to the current frame number t and the comparison result mode related to the stationary part Information TR representing dts> DTS as s
Write the UE (S11).

If the distance dts is less than or equal to the threshold value DTS in S10, the voice feature vector x _t of the current frame number _t.
Is the voice feature vector x _qs of the stationary part reference frame number _qs
Therefore, since the voice feature vector x _t of the current frame number t has a small change from the voice feature vector x _{qs of} the stationary part reference frame number qs, the stationary part reference frame number qs is not rewritten and the stationary part reference frame number qs is not rewritten. Comparison result mode related to department Information indicating dts ≦ DTS as s
Write FALSE (S12).

[0233] Once finished the process of S11 or S12, then matching unit 30 a distance dtt between the speech feature vector x _qt speech feature vector x _t and the transient portion reference frame number qt the current frame number t, the following It is obtained according to the equation (18) (S13).

[0234]

(Equation 23)

However, x _t k: vector component of the voice feature vector x _t of the current frame number t x _qt k: vector component of the voice feature vector x _qt of the transition part reference frame number qt Next, the matching unit 30 determines the transition part. Distance dtt and threshold D related to
It is compared with TT to determine whether these vectors x _t and x _qt are approximately equal (S14).

If the distance dtt exceeds the threshold value DTT in S14, the voice feature vector x _t of the current frame number t is not approximated to the voice feature vector x _qt of the transition part reference frame number qt, and thus the voice of the current frame number t is Feature vector x
_t is the speech feature vector x of the transition part reference frame number qt
_Since the change from _qt is large, the reference frame number of the transition part q
t is rewritten to the current frame number t and the comparison result mode related to the transient part Information TR representing dtt> DTT as t
Write the UE (S15).

If the distance dtt becomes less than or equal to the threshold value DTT in S14, the voice feature vector x _t of the current frame number _t.
Is the speech feature vector x _qt of the reference frame number _qt
Therefore, since the voice feature vector x _t of the current frame number t has a small change from the voice feature vector x _qt of the reference frame number qt of the transient part, the transition reference frame number qt of the transient part is not rewritten and the transient reference frame number qt is not changed. Comparison result mode related to department Information representing dtt ≦ DTT as t
FALSE is written (S16).

[0238] Once finished the step S15 or S16, then matching unit 30, the transition number j of the original S _j (number j is a number that is given to the transition source S _j state transition in hidden Markov models) Initial and Set the value to 1 (S17),
Then, it is determined whether or not the number _j of the transition source S _j exceeds the maximum number J (J = I here) (S18).

If j ≦ J in S18, the collating unit 30
Determines whether the type s _assigned to the transition source S _j is a steady part or a transient part (S19).

If the type determination result in S19 is the stationary part, the collating unit 30 next compares the comparison result mode related to the stationary part. By referring to s, it is determined whether or not the distance dts related to the stationary part exceeds the threshold value DTS (S20).

The comparison result mode in S20 s is dts> DTS
If the information TRUE indicating that the current frame number is t, the voice feature vector x _t of the current frame number _t is the stationary part reference frame number q
the change from speech feature vector x _qs speech feature vector x _t is the constant part reference frame number qs of s speech feature vector x _qs not approximate the thus current frame number t is large, the output probability B of the current frame number t _ji (x _t ) is the reference probability B
It cannot be approximated by _ji . Therefore, the matching unit 30 uses j = 1, 2,
.., J and i = 1, 2, ..., For all j and i of I, logarithmicized output probabilities B _ji (x _t ) are _expressed by equations (4) to (7).
According to the reference probability B _ji , and the output probability B _ji (x _t )
(S21). Next, the collation unit 30 adds 1 to the number j of the transition source S _{j in} order to process the next number j (S22), and thereafter performs the process of S18.
The process of rewriting the reference probability B _{ji in} S21 is represented by save B _ji = B _ji (x _t ) in the figure.

The comparison result mode in S20 s is dts ≦ DTS
If the information FALSE is that the voice feature vector x _t of the current frame number t is approximately equal to the voice feature vector x _qs of the stationary part reference frame number qs, the voice feature vector x of the current frame number t is _{Since t} has a small change from the speech feature vector x _qs of the stationary part reference frame number qs, the output probability B _ji (x _t ) of the current frame number t can be approximated by the reference probability B _ji . Therefore, the matching unit 30 does not perform the process of S21, and therefore the output probability B _ji (x _t ) is _calculated by the equation (4).
~ (7) is not performed and the process of rewriting the reference probability B _ji is not performed, 1 is added to the number j of the transition source S _j to perform the process for the next number j (S22), and thereafter, The process of S18 is performed.

If the type determination result of S19 is the transient part, the collating part 30 next determines the comparison result mode relating to the transient part. With reference to t, it is determined whether or not the distance dtt related to the transition portion has exceeded the threshold value DTT (S23).

In S23, the comparison result mode t is dtt> DTT
If the information TRUE indicating that the current frame number is t, the speech feature vector x _t of the current frame number _t is the transition part reference frame number q.
Since t is speech feature vector x _t speech feature vector x _qt without approximation therefore current frame number t of large changes from speech feature vector x _qt transient portion reference frame number qt, output probability B of the current frame number t _ji (x _t ) is the reference probability B
It cannot be approximated by _ji . Therefore, the matching unit 30 uses j = 1, 2,
.., J and i = 1, 2, ..., For all j and i of I, logarithmic output probabilities B _ji (x _t ) are _expressed by equations (4) to (7).
According to the reference probability B _ji , and the output probability B _ji (x _t )
(S21). Next, the collation unit 30 adds 1 to the number j of the transition source S _{j in} order to process the next number j (S22), and thereafter performs the process of S18.

Comparison result mode in S23 t is dtt ≦ DTT
If the information FALSE is that the voice feature vector x _t of the current frame number t is approximately equal to the voice feature vector x _qt of the transition part reference frame number qt, the voice feature vector of the current frame number t is Since x _t has a small change from the speech feature vector x _qt of the transition part reference frame number qt, the output probability B _ji (x _t ) of the current frame number t can be approximated by the reference probability B _ji . Therefore, the matching unit 30 uses S2
Therefore, the process for the next number j is performed without performing the process of 1 and thus without performing the process of _{obtaining the} output probability B _ji (x _t ) according to the equations (4) to (7) and the process of rewriting the reference probability B _ji. Therefore, 1 is added to the number j of the transition source S _j (S2
2) Then, the process of S18 is performed.

When the processing of S19 to S23 is completed for all j of j = 1, 2, ..., J, the determination result of j> J (here J = I) is obtained in the processing of S18.
If j> J in S18, then the matching unit 30 reads out the reference probabilities B _ji and calculates the forward probability C _it for all i of i = 1, 2, ..., I according to the equation (11). Ask (S24). After that, the process returns to S7 so as to perform the process for the next frame of the voice section.

(2-1B: When t> T in S8) If the current frame number t is larger than the frame number T of the terminating frame in S8, i = 1, 2, ... since finished sought forward probability C _iT, wherein i = * i maximum forward probability C _iT of the forward probability C _iT comprising the following (9), the speech feature vector time series _{x 1, x 2, ......,} x
Likelihood ln {P (x ₁ , x ₂ , ..., X between _T and HMM
_T )}, and after that, the process of calculating the likelihood for the HMM is ended (end).

The collating unit 30 performs the processes of S1 to S23 shown in FIGS. 9 to 11 for all the HMMs stored in the dictionary unit 24 to calculate the likelihood (forward probability C).
_iT ), and the category of the HMM for which the maximum likelihood is obtained is set to the time series x ₁ , x ₂ , ..., Of the speech feature vector.
As a recognition result for the input speech signal in which x _T is extracted,
Output to the next-stage device (not shown).

As described above, the likelihood ln {P (x ₁ , x ₂ , ...
, X _T )} = C _iT in the process of transition _Sj
If the distance dts is less than or equal to the threshold value DTS when is a stationary portion, or if the distance dtt is less than or equal to the threshold value DTT when the transition source S _j is a transient portion, the output probability B _ji (x _t Since the forward probability C _it is calculated without performing the calculation of () from equations (4) to (7), the amount of calculation can be significantly reduced. Moreover, simplification of such an operation is based on the transition source S _j
Is a stationary part, the distance dts is less than or equal to the threshold value DTS, or the transition source S _j is a transient part, the distance dt is
If t is less than or equal to the threshold value DTT, it is performed.
Even if the calculation is simplified, the error of the forward probability C _it can be reduced.

Further, since the change of the voice feature vector x _t that is extracted in time sequence in the transient portion of the voice signal is large,
When the type s of the transition source S _j is a transient part, it is desired to reduce the error of the forward probability C _it by setting the threshold value DTT related to the transient part to be small.

On the other hand, since the change of the voice feature vector x _t which is extracted in time series in the stationary part of the audio signal is small, when the type s of the transition source S _j is the stationary part,
Forward probability C even if the threshold value DTS related to the stationary part is increased
it is possible to reduce the error of _it.

Therefore, a large value is used as the threshold value DTS related to the steady part, and the threshold value DT related to the transient part is used.
By using T having a small value, _it is possible to reduce the amount of calculation while minimizing the error of the forward probability C _it .

According to the simulation result of the inventor of this application, in this example shown in FIGS. 9 to 11, the amount of calculation for obtaining the forward probability C _it is about 1 / third that when the calculation is not simplified. Even if the threshold value DTS related to the stationary part and the threshold value DTT related to the transient part are determined so as to be 5, the recognition of the voice recognition is performed in this example shown in FIG. 9 to FIG. 11 and the case where the calculation is not simplified. There were many cases where not only the accuracy did not differ significantly, but rather the recognition accuracy improved.

<Second Embodiment of the Invention of Claim 3> As a voice recognition apparatus suitable for carrying out the second embodiment of the invention of claim 3, the collating unit 30 is configured as follows. It is possible to use the voice recognition device 10 having the same configuration as that described above.

That is, the matching unit 30 determines the reference probability b _ji stored in the reference information storage unit 32 when obtaining the likelihood.
, The forward probability c _it in each case of t = 1, 2, ..., T is sequentially obtained as follows.

(1). When t = 1, the constant part skip number skips and the transient part skip number skipt are initialized to 0, and the constant part reference frame number qs and the transient part reference frame number qt are initialized to 1 and all j are initialized. , I, the output probability b _ji (x _t ) is obtained from the Hidden Markov model, and the output probability b _ji (x _t ) is referred to as the reference probability b _ji.
Writing the initial value, it reads out the reference probability b _ji after completion of writing of the reference probability b _ji by obtaining the forward probability c _it process performs (2A).

After the end of the process (2A), a process (2B) of adding 1 to the current frame number t is performed.

(2). When the 2 ≦ t ≦ T, as well as comparing the constant region skip number skips a threshold NSKIPS, the distance between the speech feature vector x _qs speech feature vector x _t and the constant part reference frame number qs of the current frame number t dts is compared with the threshold value DTS, and the comparison result is sk.
When ips> NSKIPS or dts> DTS, the constant part skip number skips is initialized to 0, and the constant part reference frame number qs is rewritten to the current frame number t, and the comparison result is skips ≦ NSKIPS and dts ≦ DTS. If, then the number of skips in the stationary part sk
1 and adds the process (2C) in ips, while comparing the transient portion skip number skipt a threshold NSKIPT, the speech feature vector x _qt speech feature vector x _t and the transient portion reference frame number qt the current frame number t Distance d
tt is compared with the threshold value DTT, and the comparison result is skippt.
> NSKIPT or dtt> DTT,
The transition part skip number skipt is initialized to 0, and the transition part reference frame number qt is rewritten to the current frame number t, and the comparison result is skip≤NSKIPT and dt.
When t ≦ DTT, the number of skips in transition part skippt
The process of adding 1 to (2D) is performed.

After the processes (2C) and (2D) are completed,
For each j of j = 1, 2, ..., J, a process (2E) of determining the type s given to the transition source S _j of the state transition giving the output probability b _ji (x _t ) is performed.

When the type determination result of the process (2E) is the stationary part, the comparison result of the process (2C) is skip.
If s> NSKIPS or dts> DTS, for j for which the type determination result is obtained, for all i,
When the output probability b _ji (x _t ) is obtained from the Hidden Markov model, the reference probability b _ji is rewritten to the output probability b _ji (x _t ), and when the type determination result of the process (2E) is the stationary part, the process The comparison result of (2C) is skips ≦ NSKIPS and d
If ts ≦ DTS, the reference probability b _ji is not rewritten for j for which the type determination result is obtained, and the process (2E) is performed.
When the type determination result of is a transition part, processing (2D)
Comparison result of skip> NSKIPT or dtt
If> DTT, for j for which the type determination result is obtained, for all i, output probabilities b _ji (x _t ) are obtained from Hidden Markov models and reference probabilities b _ji are output probabilities b _ji (x
_t ), and when the type determination result of the process (2E) is the transition part, the comparison result of the process (2D) is skip≤
If NSKIPT and dtt ≦ DTT, the process (2F) of not rewriting the reference probability b _ji is performed for j that has obtained the type determination result.

Then, the process (2F) is performed for each j of j = 1, 2, ..., J, and the process (2F) is performed for all j.
After the step F) is completed, each reference probability b _ji is read out and the forward probability c _it is calculated (2G).

After the processing (2G) is completed, processing (2H) is performed to add 1 to the current frame number t.

Next, in the second embodiment of the invention of claim 3, the time series x ₁ , x ₂ , ... Of HMM and speech feature vector.
,, the flow of the process of obtaining the likelihood between x _T and 1
The flow of processing focused on individual HMMs will be described. 12 to 14 are diagrams showing the flow of processing focusing on this one HMM. In this example, the output probability b _ji (x _t ), the forward probability c _it, and the reference probability b _ji are logarithmized output probability B _ji (x _t ), logarithmic forward probability C _it, and logarithmic reference probability, respectively. B _ji and i = j = 1, 2, ..., I.

When the section information and the speech feature vector x _t are input from the speech section detection unit 28, the matching unit 30 receives i =
For all i of 1, 2, ..., I, the initial value C _i0 of the forward probability logarithmized is set according to the equation (10) (S).
1).

Next, the collating unit 30 initializes the current frame number t to t = 1 in order to process the start frame of the voice section (S2).

Next, the collating unit 30 determines that j = 1, 2, ..., J.
And i = 1, 2, ..., I, logarithmized output probabilities B _ji (x ₁ ) are obtained according to equations (4) to (7) (S3), and the output probabilities B _{ji are obtained.} (x ₁ ) is written as the initial value of the logarithmic reference probability B _ji (S4).

In the reference information storage section 32, j = 1, 2, ...
, J and i = 1, 2, ..., I are respectively provided with storage areas save B _ji for storing the reference probabilities B _ji . Therefore, the reference information storage unit 32 stores B ₁₁ , B ₁₂ , ...
_{..., B 1I, B 21,} B 22, ......, B 2I, ......, B J1,
B _J2 , ..., B _JI respectively have J × I storage areas for individually storing. Therefore, in the figure, the reference probability B _ji
The process of storing the initial value of is expressed as save B _ji = B _ji (x ₁ ).

Next, the matching unit 30 determines the number of regular part skips sk.
ips and the number of skips in the transition part skipt are initialized to 0, and the reference frame number qs of the stationary part and the reference frame number qt of the transition part are initialized to the current frame number 1 (S5), and then i = 1. , ..., For all i in I, the forward probability C _i1 which is logarithmized is expressed by the formula (1
It is obtained according to 1) (S6).

Next, the collating unit 30 adds 1 to the current frame number t in order to process the next frame of the voice section (S7), and thereafter, the current frame number t and the frame number T of the end frame are added. Are compared to determine whether the processing has been completed for all the frames in the voice section (S8).

(2-2A: When t ≦ T in S8) If the current frame number t is less than or equal to the end frame number T in S8, the processing is not completed for all the frames in the voice section, so that the steady state is performed. Copy skip number skips and threshold NSK
A comparison judgment with IPS is performed (S9).

If the skip count skips of the stationary part exceeds the threshold value NSKIPS in S9, the distance dt related to the stationary part
The number of times s is less than or equal to the threshold DTS skips is the threshold NS
Since the time gap between the current frame number t and the constant part reference frame number qs is increased beyond KIPS, the error is likely to increase. Therefore, the number of regular part skips s
Initialize kips to 0, rewrite the constant part reference frame number qs to the current frame number t, and compare the result As s, skips> NSKIP
Information TRUE indicating that S or dts> DTS was true
Is written (S10).

[0272] When the constant region skip number skips is equal to or less than the threshold NSKIPS in S9, then matching unit 30 of the audio feature vector x _qs speech feature vector x _t and the constant part reference frame number qs of the current frame number t Distance dt
s is obtained according to the equation (17) (S11), and then the distance dts related to the stationary part is compared with the threshold value DTS to determine whether or not these vectors x _t and x _qs are approximately equal (S12). .

If the distance dts exceeds the threshold value DTS in S12, the voice feature vector x _t of the current frame number t does not approximate to the voice feature vector x _qs of the stationary part reference frame number qs, and therefore the voice of the current frame number t Feature vector x
_t is the voice feature vector x of the stationary part reference frame number qs
The change from _qs is large. Therefore, the number of skips in the steady part ski
ps is initialized to 0 and the reference frame number q of the stationary part
s is rewritten to the current frame number t, and the comparison result mode related to the stationary part Information s indicating that skips> NSKIPS or dts> DTS is written as s (S10).

If the distance dts is less than or equal to the threshold DTS in S12, the voice feature vector x _t of the current frame number _t
Is the voice feature vector x _qs of the stationary part reference frame number _qs
Speech feature vector x _t of approximately equal therefore current frame number t is small variation from the speech feature vector x _qs constant region reference frame number qs to. Then, 1 is added to the number of skips in the steady part skips to skip the number of skips in the steady part
s is counted up and the comparison result mode related to the stationary part As s, skips ≦ NSKIPT and dts
Write information FALSE indicating that ≤ DTS (S
13).

After the processing of S10 or S13 is completed, the collating unit 30 next makes a comparison judgment between the transient part skip number skipt and the threshold value NSKIPT (S14).

If the number of skips in the transitional part skip exceeds the threshold value NSKIPT in S14, the distance d related to the transitional part
The number of times tt becomes equal to or less than the threshold DTT, skipt is the threshold N
Since the time frame between the current frame number t and the transitional section reference frame number qt becomes larger than the SKIPT,
The error is likely to increase. Therefore, the transition part skip number skipt is initialized to 0, the transition part reference frame number qt is rewritten to the current frame number t, and the comparison result mode relating to the transition part is set. As t, skippt> NSKI
Information TR showing that PT or dtt> DTT
Write the UE (S15).

If the number of skips in the transitional part skip is less than or equal to the threshold value NSKIPT in S14, the collation unit 30 next determines the speech feature vector x _t of the current frame number t and the speech feature vector x qt of the reference frame number of the transitional part _qt . Distance d
tt is obtained according to the equation (18) (S16), and then the distance dtt related to the transient part is compared with the threshold value DTT to determine whether or not these vectors x _t and x _qt are approximately equal (S17). .

If the distance dtt exceeds the threshold value DTT in S17, the voice feature vector x _t of the current frame number t is not approximated to the voice feature vector x _qt of the transition part reference frame number qt, and therefore the voice of the current frame number t. Feature vector x
_t is the speech feature vector x of the transition part reference frame number qt
The change from _qt is large. Therefore, the number of skips in the transition part ski
pt is initialized to 0 and the transition part reference frame number q
t is rewritten to the current frame number t, and the comparison result mode related to the transition part Information t indicating that skip> NSKIPT or dtt> DTT is written as t is written (S15).

If the distance dtt is less than or equal to the threshold value DTT in S17, the voice feature vector x _t of the current frame number _t.
Is the speech feature vector x _qt of the reference frame number _qt
Speech feature vector x _t of approximately equal therefore current frame number t is small variation from the speech feature vector x _qt transient portion reference frame numbers qt to. Therefore, 1 is added to the number of skips in the transition part skip, and the number of skips in the transition part skip
t is counted up and the comparison result mode related to the transient part As t, skipt ≦ NSKIPT and dtt
Write information FALSE indicating that ≦ DTT (S
18).

[0280] Once finished the step S15 or S18, then matching unit 30, the transition number j of the original S _j (number j is a number that is given to the transition source S _j state transition in hidden Markov models) Initial and Set the value to 1 (S19),
Then, it is determined whether or not the number _j of the transition source S _j exceeds the maximum number J (J = I here) (S20).

If j ≦ J in S20, the collating unit 30
Determines whether the type s _assigned to the transition source S _j is a steady part or a transient part (S21).

If the type determination result in S21 is the stationary part, the collating part 30 next compares the comparison result mode related to the stationary part. By referring to s, it is determined what the skip number skips related to the stationary part, the comparison result of the threshold NSKIPS and the distance dts, and the comparison result of the threshold DTS are (S2).
2).

Comparison result mode in S22 s is skips> N
If the information is TRUE indicating that SKIPS or dts> DTS, the collation unit 30 determines that j = 1, 2, ...
For all j and i of J and i = 1, 2, ..., I,
The logarithmic output probability B _ji (x _t ) is obtained according to equations (4) to (7), and the reference probability B _ji is rewritten to the output probability B _ji (x _t ) (S23). Next, the collation unit 30 adds 1 to the number j of the transition source S _{j in} order to process the next number j (S24), and thereafter performs the process of S20. S2
The process of rewriting the reference probability B _{ji in} 3 is sa in the figure.
It is expressed as ve B _ji = B _ji (x _t ).

If skips> NSKIPS, the number of times s that the distance dts related to the stationary part becomes equal to or less than the threshold value DTS.
Since the kips exceeds the threshold value NSKIPS, there is a large time gap between the current frame number t and the constant part reference frame number qs, and thus the error is likely to increase. Therefore, in order to reduce the error, the reference probability B _ji is rewritten.

If dts> DTS, the voice feature vector x _t of the current frame number _t is the stationary part reference frame number q.
the change from speech feature vector x _qs speech feature vector x _t is the constant part reference frame number qs of s speech feature vector x _qs not approximate the thus current frame number t is large, the output probability B of the current frame number t _ji (x _t ) is the reference probability B
It cannot be approximated by _ji . Therefore, the reference probability B _ji is rewritten.

Comparison result mode in S22 s is skips ≤ N
Information indicating that SKIPS and dts ≦ DTS were satisfied
If it is FALSE, the collation unit 30 performs the process of S23.
Without, and therefore output probability B_ji(x_t) According to equations (4)-(7)
The processing to obtain is also the reference probability B _jiAlso rewrite
Instead of performing the process for the next number j, the transition source S
_j 1 is added to the number j of (S24), and then S20
Perform processing.

Skips ≦ NSKIPS and dts ≦ D
If TS, skips ≦ NSKIPS, so the number of times the distance dts related to the stationary part is equal to or less than the threshold value DTS sk
Since ips does not exceed the threshold value NSKIPS, and therefore the time gap between the current frame number t and the constant part reference frame number qs is small, the error is unlikely to increase. Moreover, since dts ≦ DTS, the voice feature vector x _t of the current frame number t becomes approximately equal to the voice feature vector x _qs of the stationary part reference frame number qs, and therefore the current frame number t
Voice feature vector x _t of the stationary part reference frame number qs
The change from speech feature vector x _qs is small, the output probability B _ji (x _t) of the current frame number t may be approximated by the reference probability B _ji. Therefore, the reference probability B _ji is read out without being rewritten to obtain the forward probability C _it .

If the type determination result of S21 is the transitional part, the collation unit 30 next determines the comparison result mode related to the transitional part. With reference to t, it is determined what the skip number skipt related to the transition part, the comparison result of the threshold value NSKIPT, the distance dtt, and the comparison result of the threshold value DTT are (S2).
5).

Comparison result mode in S25 t is skippt> N
If the information is TRUE indicating that SKIPT or dtt> DTT, the collation unit 30 determines that j = 1, 2, ...
For all j and i of J and i = 1, 2, ..., I,
The logarithmic output probability B _ji (x _t ) is obtained according to equations (4) to (7), and the reference probability B _ji is rewritten to the output probability B _ji (x _t ) (S23). Next, the collation unit 30 adds 1 to the number j of the transition source S _{j in} order to process the next number j (S24), and then performs the process of S20.

If skipt> NSKIPT, the number of times s that the distance dtt related to the transient portion becomes equal to or less than the threshold value DTT.
Since kipt exceeds the threshold value NSKIPT, there is a large time gap between the current frame number t and the transition part reference frame number qt, and thus the error is likely to increase. Therefore, the reference probability B _ji is rewritten in order to reduce the error.

If dtt> DTT, the voice feature vector x _t of the current frame number _t is the transition part reference frame number q.
Since t is speech feature vector x _t speech feature vector x _qt without approximation therefore current frame number t of large changes from speech feature vector x _qt transient portion reference frame number qt, output probability B of the current frame number t _ji (x _t ) is the reference probability B
It cannot be approximated by _ji . Therefore, the reference probability B _ji is rewritten in order to reduce the error.

In S25, the comparison result mode t is skip ≦ N
Information indicating that SKIPT and dtt ≦ DTT
If it is FALSE, the matching unit 30 does not perform the process of S23, and therefore does not perform the process of _{obtaining the} output probability B _ji (x _t ) according to the equations (4) to (7) or the process of rewriting the reference probability B _ji. In order to perform processing for the next number j, the transition source S
adds 1 to _j number j (S24), thereafter, it performs the processing of S20.

Skipt ≤ NSKIPT and dtt ≤ D
If TT, then skip ≦ NSKIPT, so the number of times sk that the distance dtt related to the transition part is equal to or less than the threshold value DTT
Since ipt does not exceed the threshold value NSKIPT, and therefore the time gap between the current frame number t and the transition part reference frame number qt becomes small, the possibility of increasing the error is low. Moreover, since dtt ≦ DTT, the voice feature vector x _t of the current frame number t becomes approximately equal to the voice feature vector x _qt of the reference frame number qt of the transition part, and thus the voice feature vector x _t of the current frame number _t is the transition part. the change from speech feature vector x _qt reference frame number qt is small,
The output probability B _ji (x _t ) of the current frame number t can be approximated by the reference probability B _ji . Therefore, the reference probability B _ji is not rewritten.

When the processing of S20 to S25 is completed for all j of j = 1, 2, ..., J, the determination result of j> J (here J = I) is obtained in the processing of S20.
If j> J in S20, then the matching unit 30 reads out the reference probabilities B _ji and calculates the forward probability C _it for all i of i = 1, 2, ..., I according to the equation (11). Ask (S26). After that, the process returns to S7 so as to perform the process for the next frame of the voice section.

(2-2B: When t> T in S8) When the current frame number t is larger than the frame number T of the end frame in S8, i = 1, 2, ... since finished sought forward probability C _iT, wherein i = * i maximum forward probability C _iT of the forward probability C _iT comprising the following (9), the speech feature vector time series _{x 1, x 2, ......,} x
Likelihood ln {P (x ₁ , x ₂ , ..., X between _T and HMM
_T )}, and after that, the process of calculating the likelihood for the HMM is ended (end).

The collating unit 30 detects all the HMMs stored in the dictionary unit 24 for each HMM as shown in FIGS.
, The likelihood (forward probability C _iT ) is obtained by performing the processes of S1 to S26, and the category of the HMM for which the maximum likelihood is obtained is the time series x ₁ , x ₂ , ... Of the speech feature vector.
, X _T are output to a device (not shown) in the next stage as a recognition result for the extracted input voice signal.

As described above, the likelihood ln {P (x ₁ , x ₂ , ...
, X _T )} = C _iT in the process of transition _Sj
Is a stationary part, the skip number skips is the threshold NS
If KIPS or less and the distance dts is less than or equal to the threshold DTS, the reference probability B _ji is read out without performing the calculation of the output probability B _ji (x _t ) from the equations (4) to (7), and the forward probability C _it. Ask for. If the number of skips skip is less than or equal to the threshold value NSKIPT and the distance dtt is less than or equal to the threshold value DTT when the transition source S _j is the transition portion, the output probability B
_Since the forward probability C _it is calculated without performing the calculation of _ji (x _t ) from the equations (4) to (7), the calculation amount can be significantly reduced. Moreover, simplification of such an operation is based on the transition source S _j
Is a stationary part, the skip number skips is the threshold NS
The number of skips s when the distance dts is less than or equal to KIPS and the distance dts is less than or equal to the threshold value DTS, or the transition source S _j is a transition part.
kipt is less than or equal to the threshold value NSKIPT and the distance dtt
Is less than or equal to the threshold DTT, the error in the forward probability C _it can be reduced even if the calculation is simplified.

Further, since the change of the voice feature vector x _t extracted in time sequence in the transient portion of the voice signal is large,
When the type s of the transition source S _j is the transient part, it is desired to reduce the error of the forward probability C _it by setting the thresholds NSKIPT and DTT related to the transient part to be small.

On the other hand, since the change in the voice feature vector x _t extracted in time-sequential order in the stationary part of the audio signal is small, when the type s of the transition source S _j is the stationary part,
Even if the thresholds NSKIPS and DTS related to the stationary part are increased, the error of the forward probability C _it can be reduced.

Therefore, the threshold value NSKIPS relating to the stationary part,
By using a large value for DTS and a small value for the threshold values NSKIPT and DTT related to the transient part, _it is possible to reduce the amount of calculation while minimizing the error of the forward probability C _it .

The invention of claim 3 can be applied to all speech recognition apparatuses that perform matching processing in frame units.

The stationary part given to the transition source S _j ,
The type s of the transitional part can be determined as described below, for example.

The first example focuses on one parameter b _jim (x _t ) that determines the output probability b _ji (x _t ).
As shown in the equation (6), b _jim (x _t ) = (2π) ^{−p / 2} |
A ^{_{-1/2 exp {D jimt 2/2}} }, the magnitude of the dispersion and test covariance matrix in the equation (6) | | ρ _jim when a exceeds the threshold value THL which defines any suitably | [rho _jim , The type s of the transition source S _j that gives the output probability b _ji (x _t ) is determined to be the transient part, and the size | ρ _jim | of the variance-covariance matrix is the threshold T.
When it becomes HL or less, the type s of the transition source S _j that gives the output probability b _ji (x _t ) is determined to be a stationary part. Therefore, in this case, the size of the variance / covariance matrix | ρ _jim |
And the comparison judgment between this | ρ _jim | and the threshold value THL is
This means the determination of the type s.

In the second example, when the state transition giving the output probability b _ji (x _t ) corresponds to the state transition of the vowel, the information indicating that it is a stationary part for the transition source S _{j of the} state transition. To
In addition, when the state transition giving the output probability b _ji (x _t ) corresponds to the state transition of a consonant, information indicating that it is a transition part is given to the transition source S _{j of the} state transition in advance. It is a thing.

The third example is the case where the state transition giving the output probability b _ji (x _t ) corresponds to the state transition of a vowel and the state transition of a consonant other than p, t, k and r. If the state transition that gives the information indicating that the transition source S _{j is a} stationary part and that the output probability b _ji (x _t ) corresponds to the state transition of the consonants p, t, k, and r, Information indicating that it is a transitional part is added in advance to the transition source S _j of the transition.

<First Embodiment of the Invention of Claim 7> FIG.
FIG. 8 is a functional block diagram showing a configuration example of a voice recognition device suitable for use in carrying out the first embodiment of the invention of claim 7.

The voice recognition device 34 shown in FIG.
6, sound processing unit 38, voice section detection unit 40, collation unit 42
And a reference information storage unit 44.

The dictionary unit 36 stores a plurality of Hidden Markov models prepared for each category as standard patterns for recognition and matching. The reference information storage unit 44 stores the forward probability reference frame number qc and the output probability reference frame number qs.
And the reference probability b _ji are stored.

The sound processing unit 38 extracts a voice feature vector from the input voice signal for each frame of a fixed time width.
The voice section detection unit 40 detects a voice section from the input voice signal.

The collation unit 42 implements the first embodiment of the invention of claim 7, and is the time series x ₁ , x _{2 of} the voice feature vectors extracted from the start frame to the end frame of the voice section. , ..., The likelihood ln {P (x ₁ , x ₂ , ..., x _T )} between x _T and the Hidden Markov model is
The category given to the Hidden Markov model that has been obtained using the following equations (1) to (3) and has the maximum likelihood is used as the recognition result for the voice signal in the voice section.

[0311]

(Equation 24)

However, i: i = 1, 2, ..., I j: j = 1, 2, ..., J Φ _i : Probability that the initial state is S _i in the Hidden Markov model a _ji : Hidden Markov model At the state S _j to the state S _i in the above, x _t : the speech feature vector extracted in the t-th frame in the speech section (1 ≦ t ≦ T, and the first frame corresponds to the speech section). The start frame and the T-th frame represent the end frame of the speech section) b _ji (x _t ): speech feature vector x output when transitioning from state S _j to state S _i in the Hidden Markov model
Output probability of _t c _it : Time series of speech feature vector x ₁ , x ₂ , ..., Which starts transition from the initial state in Hidden Markov model
Forward probability of outputting x _t to the state S _i * i: State number i assigned to the state S _i that is the final state in the Hidden Markov model When storing the likelihood, it is stored in the reference information storage unit 44. Using the given reference probability b _ji , the forward probability c _it for each case of t = 1, 2, ..., T is sequentially obtained as follows.

(1). When t = 1, the forward probability reference frame number qc and the output probability reference frame number qs are initialized to 1, and all j and i are
Writing output probability b _ji (x _t) the probability that output determined from Hidden Markov Models b _ji the (x _t) as the initial value of the reference probability b _ji, reads out the reference probability b _ji after completion of writing of the reference probability b _ji Processing (3A) for obtaining the forward probability c _it . After the processing (3A) is completed, processing (3B) for adding 1 to the current frame number t is performed.

(2). 2 ≦ t when the ≦ T, the distance between the speech feature vector x _qc of forward probabilities reference frame number qc and speech feature vectors x _t of the current frame number t dtc
Is compared with the threshold value DTC (3C), and this processing (3C)
When the comparison result of C) is dtc ≦ DTC, the forward probability c _it is regarded as equal to the forward probability c _{i (t-1)} of the immediately preceding frame, and the process of terminating the forward probability c _it is finished (3D). , The comparison result of this process (3C) is dtc> D
If it becomes TC, the forward probability reference frame number qc is rewritten to the current frame number t (3E).

[0315] Then after the processing (3E), the distance dts between the speech feature vector x _qs output probabilities reference frame number qs with speech feature vectors x _t of the current frame number t is compared with a threshold DTS, the comparison When the result is dts> DTS, the output probability reference frame number qs is rewritten to the frame number t, and the output probabilities b _ji (x _t ) are obtained from the Hidden Markov model for all j and i, and the reference probabilities b _{ji are obtained.} To the output probability b _ji (x _t ), read the reference probabilities b _ji after the reference probabilities b _ji have been rewritten, obtain the forward probabilities c _it, and when the comparison result is dts ≦ DTS, the reference probabilities are without rewriting b _ji reads each reference probability b _ji seek forward probability c _it processes (3
F).

After the processing (3D) or (3F), the processing (3G) of adding 1 to the current frame number t is performed.

FIG. 16 is a diagram for explaining the Hidden Markov Model. Hidden Markov Model (HMM) stored in the dictionary unit 36
Represents a voice signal for one unit of voice recognition. One unit of speech recognition can be a word unit, a phoneme unit, or another unit, but here, it is a word unit. Each category z
A plurality of HMMs are prepared for each, and the HMM and the category z are associated with each other and stored in the dictionary unit 36.

The HMM has a set 1 of states consisting of a total of I states S _{1 to} S _I and a set 2 of speech feature vectors x.
, A set 3 of state transition probabilities a _ji, a set 4 of output probabilities b _ji (x), a set 5 of initial state probabilities Φ _i , and a set 6 of final states F. However,

[0319]

(Equation 25)

I: i = 1, 2, ..., I j: j = 1, 2, ..., J a_ji: State S_j To state S_i Probability of transition to b_ji(x): State S_j To state S_i When switching to
Probability that characteristic vector x is output Ф_i : Initial state is S_i For example, in the example of FIG. 14, a₁₂Is state S₁ To state S
_Two Probability of transition to and b₁₂(x) is state S₁ To state S_Two 
The probability that the speech feature vector x is output when the transition to
Also a_{twenty two}Is state S_Two To state S_Two Probability of transition to and b
_{twenty two}(x) is state S _Two To state S_Two Voice features when transitioning to
It represents the probability that the vector x will be output.

Sets 1 to 6 for defining the HMM are individually obtained for each category z by a statistical method. That is, various voice signals are collected as voice signals corresponding to the category z, for example, voice signals are collected by age or sex, or voice signals having different voicing methods are collected.
Sets 1 to 6 expressing the statistical properties of these audio signals are obtained.

The output probability b _ji (x) is expressed by using a non-correlated mixed normal distribution consisting of a plurality of normal distributions that are uncorrelated with each other, and each of these normal distributions is a function of the speech feature vector x. There is. The decorrelated mixed normal distribution has the advantage of being easy to handle mathematically and having high expressiveness.

Next, the operation flow of the voice recognition device 34 will be described, and the flow of processing of the voice recognition method of this embodiment will be specifically described.

The sound processing unit 38 uses the input voice signal to output a voice feature vector x _t = (x _t 1, x _t 2, ...) For each frame.
,, x _t p) are extracted. Where p is the voice feature vector x
order and x _t 1 to x _t p to _t represents a vector component of the speech feature vector x _t. t is a number given to the frame from which the voice feature vector x _t is extracted. HM described later
At the stage of matching with M, the frames are rewritten in ascending order with the frame number t of the starting frame of the voice section as 1, but at the time of the acoustic processing, the frame number t may be added so that each frame can be identified.

The vector component of the voice feature vector x _t is, for example, one obtained from each filter output when an input voice signal is input to a band filter group consisting of a plurality of band pass filters having different center frequencies, or input voice signals. Power spectrum component obtained by Fourier analysis of the signal,
Alternatively, the LPC cepstrum coefficient obtained by the linear prediction analysis, that is, the LPC analysis of the input speech signal can be used. Here, an example of extracting the voice feature vector x _t using a bandpass filter group will be described.

The acoustic processing unit 38 converts the input audio signal from an analog signal into a digital signal, and outputs the converted input audio signal through a band filter group to a frequency band (channel) corresponding to each band pass filter. Separated into signal components,
The total number p of signal components x1 ...
Get xp. Next, the acoustic processing unit 38 rectifies the signal component x1 and obtains an average value of the rectified signal component x1 (absolute value of the signal component x1) in frame units. This average value is obtained by dividing the rectified signal component x1 by the time width of one frame. The average value of the signal component x1 obtained in the t-th frame is extracted as component x _t 1 of the audio feature vector x _t. Similarly, from the remaining signal components x2 ～Xp, it extracts the component x _t 2~x _t p of a speech feature vector x _t.

Next, the voice section detecting section 40 includes the sound processing section 3
Based on the voice feature vector x _t from 8, the start frame and the end frame of the voice section are detected, and the section information indicating which frame is the start frame and the end frame of the voice section is generated. The voice section is a section in which a voice signal for one unit of voice recognition is included here.

The collation unit 42 inputs the section information and the voice feature vector x _t from the voice section detection unit 40, and the time series x of the voice feature vector x _t extracted from the start frame to the end frame of the voice section. Generate ₁ , x ₂ , ..., X _T. At this time, the frame number t of the start frame is set to 1, and the frame numbers t from the start frame to the end frame of the voice section are rewritten in ascending order.

The collating unit 42 then calculates the vector time series x ₁ ,
x ₂ , ..., X _T and the HMM stored in the dictionary unit 36
And the likelihood ln {P (x ₁ , x ₂ , ..., X _T )} between
The category z given to each HMM of the dictionary unit 36 is individually obtained and the maximum likelihood is obtained, and the category z is output as the recognition result.

Here, P (x ₁ , x represented by the equation (1)
₂ , ..., x _T ) is the vector time series x in the HMM.
It is the probability that ₁ , x ₂ , ..., x _T will appear.

[0331]

(Equation 26)

C _iT in the equation (1) is a vector time series x ₁ , x ₂ , ...
..., the forward probability of outputting x _T to reach the state S _i , * i is i (the number _i assigned to the state S _i belonging to the final state F) that satisfies S _i εF, and thus i = The maximum forward probability c _iT among the forward probabilities c _iT of * i is _defined as the appearance probability P (x ₁ , x ₂ , ..., X _T ).

The forward probability c _iT is approximately obtained by the Viterbi algorithm using the recurrence formulas shown in the equations (2) to (3). c _i0 = Ф _i (2)

[0334]

[Equation 27]

In the HMM, there are one or more state transitions that output the voice feature vector x _t . Therefore, there is one or a plurality of transition paths from the initial state to output the vector series x _{1 to} x _t to reach the state S _i , and in most cases, there are a plurality of transition paths. Therefore, as shown in Expression (3), the maximum c _{j (t-1)} a _ji b _ji (x of the c _{j (t-1)} a _ji b _ji (x _t ) calculated for each transition path is obtained. Let _t ) be the forward probability c _it . This calculation method is called the Viterbi method.

The output probability b _ji (x _t ) in the equation (3) is defined as in the following equation (4).

[0337]

[Equation 28]

However, m = 1, 2, ..., M g _jim (x _t ): A speech feature vector calculated from the m-th normal distribution in the uncorrelated mixed normal distribution consisting of M normal distributions. Weighting probability of x _t The weighting probability g _jim (x _t ) in the expression (4) is expressed by the following expression (5)-
It is expressed using (7).

G _jim (x _t ) = λ _jim b _jim (x _t ) ... (5) b _jim (x _t ) = (2π) ^{−p / 2} _{│ρ jim} │ ^-1/2 exp {−D _jimt ^{2/2} ...... (6)} D jimt 2 = (x t -μ jim) 'ρ jim -1 (x t -μ jim) ...... (7) λ jim: weight b _jim of the m-th normal distribution (x _t ): _Unweighted probability of the speech feature vector x _t calculated from the m-th normal distribution ρ _jim : _Covariance / covariance matrix of the m-th normal distribution μ _jim : _{Mean of} the m-th normal distribution Vector D _jimt : Mahalabis's general distance (x _t −μ _jim ) ′: transposed matrix of (x _t −μ _jim ), which represents the distance between the voice feature vector x _t and the m-th normal distribution. Various types of b _ji (x _t ) can be used, and in addition to the formula (4), for example, a formula defined as the following formula (8) may be used. Equation (8) is the total number M
Number of regular weights probability in the distribution uncorrelated Gaussian Mixture consisting calculated from individual normal distribution g _jim largest weighted probability g _jim (x _t) of the (x _t), the output probability b _ji (x _t)
It means to detect as.

[0340]

(Equation 29)

Further logarithmic transition probability A _ji = ln
(A _ji ), logarithmic output probability B _ji (x _t ) = ln {b
_ji (x _t )} and the logarithmic forward probability C _it = ln
When expressed as (c _it ), the likelihoods are modified by modifying equations (1) to (3).
Equations (9) to (11) relating to the calculation of ln {P (x ₁ , x ₂ , ..., X _t )} are obtained.

[0342]

[Equation 30]

Since the expressions (9) to (11) are recurrence expressions of t, the logarithmized forward probability C _it when t = 1, 2, ..., T is expressed by the following expressions (12) to (12). It can be calculated sequentially as in 16).

[0344]

(Equation 31)

The HMM matching unit 42 uses i = 1, 2, ...,
Logarithmic forward probability C _{iT of} t = T for all i in I
, The maximum C _iT among the logarithmic forward probabilities C _iT with i = * i is _calculated as the likelihood ln {P (x ₁ , x ₂ , ..., X
_T )}. For all HMMs stored in the dictionary unit 36, the likelihood ln {P (x ₁ ,
x ₂ , ..., x _T )} is obtained and the maximum likelihood is obtained.
The category z given to the time series x ₁ , x
₂ , ..., x _T is output as the recognition result for the input voice signal.

Next, in the first embodiment of the invention as claimed in claim 7, time series x ₁ , x ₂ , ... Of HMM and voice feature vector.
,, the flow of the process of obtaining the likelihood between x _T and 1
The flow of processing focused on individual HMMs will be described. 17 to 19 are diagrams showing the flow of processing focusing on this one HMM. In this example, the output probability b _ji (x _t ), the forward probability c _it, and the reference probability b _ji are logarithmized output probability B _ji (x _t ), logarithmic forward probability C _it, and logarithmic reference probability, respectively. B _ji and i = j = 1, 2, ..., I.

When the section information and the speech feature vector x _t are inputted from the speech section detecting section 40, the collating section 42 receives i =
For all i of 1, 2, ..., I, the initial value C _i0 of the forward probability logarithmized is set according to the equation (10) (S).
1).

Next, the collation unit 42 initializes the current frame number t to t = 1 in order to process the start frame of the voice section (S2).

Next, the collation unit 42 determines that j = 1, 2, ..., J
And i = 1, 2, ..., I, logarithmized output probabilities B _ji (x ₁ ) are obtained according to equations (4) to (7) (S3), and the output probabilities B _{ji are obtained.} (x ₁ ) is written as the initial value of the logarithmic reference probability B _ji (S4).

In the reference information storage section 44, j = 1, 2, ...
, J and i = 1, 2, ..., I are respectively provided with storage areas save B _ji for storing the reference probabilities B _ji . Therefore, the reference information storage unit 44 stores the reference probabilities B ₁₁ , B
₁₂ , ……, B _1I , B ₂₁ , B ₂₂ , ……, B _2I , ……,
B _J1 , B _J2 , ..., B _JI are stored separately J ×
It has I storage areas. Therefore, in the figure, the process of storing the initial value of the reference probability B _ji is performed as save B _ji = B _ji (x ₁ )
It is expressed as

Next, the collation unit 42 initializes the forward probability reference frame number qc and the output probability reference frame number qs to the current frame number 1 (S5).

After that, all i in i = 1, 2, ..., I
The logarithmic forward probability C _i1 is _{calculated according} to the equation (11) (S6).

Next, the collation unit 42 adds 1 to the current frame number t in order to process the next frame in the voice section (S7), and thereafter, the current frame number t and the frame number T of the end frame are added. Are compared to determine whether the processing has been completed for all the frames in the voice section (S8).

(3-1A: When t ≦ T in S8) If the current frame number t is equal to or smaller than the end frame number T in S8, the processing has not been completed for all the frames in the voice section, and thus the comparison is performed. part 42 a distance dtc between the speech feature vector x _qc of forward probabilities reference frame number qc and speech feature vectors x _t of the current frame number t, the following equation (19)
(S9).

[0355]

(Equation 32)

However, x _t k: vector component of the voice feature vector x _t of the current frame number t x _qc k: vector component of the voice feature vector x _qc of the forward probability reference frame number qc Next, the matching unit 42 determines the distance dtc. And the threshold value DTC are compared to determine whether or not these vectors x _t and x _qc are approximately equal (S10).

If the distance dtc is less than or equal to the threshold value DTC in S10, the voice feature vector x _t of the current frame number _t.
Is close to the voice feature vector x _qc of the forward probability reference frame number qc, and therefore the voice feature vector x _t of the current frame number t has a small change from the voice feature vector x _qc of the forward probability reference frame number qc. t
The forward probability C _{it of} is the forward probability C _{i (t-1)} of the immediately preceding frame.
Can be approximated by Therefore, the forward probability C of the current frame number t
_It is assumed that _it is equal to the forward probability C _{i (t-1)} of the immediately preceding frame, and the calculation for obtaining the forward probability C _it is completed (S1).
1). After that, the process returns to S7 so as to perform the process for the next frame of the voice section.

When the distance dtc exceeds the threshold value DTC in S10, the voice feature vector x _t of the current frame number t is the voice feature vector x _t of the forward probability reference frame number qc.
Since the speech feature vector x _t approximation without therefore current frame number t in _qc large change from the speech feature vector x _qc of forward probabilities reference frame number qc, the current frame number t
The forward probability C _{it of} is the forward probability C _{i (t-1)} of the immediately preceding frame.
Cannot be approximated with. Therefore, the forward probability reference frame number q
c is rewritten to the current frame number t (S12).

[0359] Then, after completion of S12, the distance between the matching unit 42 speech feature vectors x _t and the output probability reference frame number qs speech feature vector x _qs of the current frame number t d
ts is calculated according to the following equation (17) (S13).

[0360]

[Equation 33]

However, x _t k: vector component of the voice feature vector x _t of the current frame number t x _qs k: voice feature vector x _qs of the reference frame number _qs
Next, the matching unit 42 compares the distance dts with the threshold value DTS and determines whether or not these vectors x _t and x _qs are approximately equal (S14).

If the distance dts exceeds the threshold value DTS in S14, the voice feature vector x _t of the current frame number t is the voice feature vector x qs of the output probability reference frame number _qs.
Therefore, since the voice feature vector x _t of the current frame number t has a large change from the voice feature vector x _qs of the output probability reference frame number qs, refer to the output probability B _ji (x _t ) of the current frame number t. It cannot be approximated with the probability B _ji . Therefore, the output probability reference frame number qs is rewritten to the current frame number t (S15). After that, logarithmized output probabilities B _ji (x _t ) for all j and i of j = 1, 2, ..., J and i = 1, 2 ,. _Obtained according to (7), the reference probability B _ji is rewritten to the output probability B _ji (x _t ) (S16). After the rewriting of the reference probabilities B _ji, the reference probabilities B _ji are read out, and the forward probabilities C _it for all i of i = 1, 2, ..., I are calculated according to the equation (11) (S13). After that, the process returns to S7 so as to perform the process for the next frame of the voice section. The process of rewriting the reference probability B _{ji in} S16 is save B _ji =
It is expressed as B _ji (x _t ).

Reference probability B read in S17 in this case
_ji is the output probability B _ji (x _t ) of the current frame number t obtained in S16. Therefore, in S17 of this case, the forward probability C is calculated using the output probability B _ji (x _t ) of the current frame number t. _It will ask for it.

If the distance dts is less than or equal to the threshold value DTS in S14, the voice feature vector x _t of the current frame number t is approximately equal to the voice feature vector x _qs of the output probability reference frame number qs, and therefore the current frame number. Since the voice feature vector x _t of _t has a small change from the voice feature vector x _qs of the output probability reference frame number qs, the output probability B _ji (x _t ) of the current frame number t can be approximated by the reference probability B _ji . Therefore, each reference probability B _ji is read out without calculating the output probability B _ji (x _t ) using the equations (4) to (7), and i = 1,
The logarithmic forward probability C _it is calculated for all i of 2, ..., I according to the equation (11) (S17). After that, S is executed to process the next frame of the voice section.
It returns to the process of 7.

Reference probability B read in S17 in this case
_ji is the output probability B _ji (x _qs ) obtained in the frame of the output probability reference frame number qs, and thus S17 in this case.
Then, the output probability B _ji (x of the output probability reference frame number qs
_The forward probability C _it is calculated using _qs ).

(3-1B: When t> T in S8) When the current frame number t is larger than the frame number T of the end frame in S8, i = 1, 2, ... since finished sought forward probability C _iT, wherein i = * i maximum forward probability C _iT of the forward probability C _iT comprising the following (9), the speech feature vector time series _{x 1, x 2, ......,} x
Likelihood ln {P (x ₁ , x ₂ , ..., X between _T and HMM
_T )}, and after that, the process of calculating the likelihood for the HMM is ended (end).

The collation unit 42, for all the HMMs stored in the dictionary unit 36, is shown in FIGS.
, The likelihood (forward probability C _iT ) is obtained by performing the processes of S1 to S17, and the category of the HMM for which the maximum likelihood is obtained is the time series x ₁ , x ₂ , ... Of the speech feature vector.
, X _T are output to a device (not shown) in the next stage as a recognition result for the extracted input voice signal.

As described above, the likelihood ln {P (x ₁ , x ₂ , ...
, X _T )} = C _{iT In the} process of _obtaining C _iT , when the distance dtc related to the forward probability C _it becomes the threshold value DTC, the output probability B _ji (x _t ) is calculated from the equations (4) to (7). even without also calculation for obtaining the forward probability C _it from equation (3) or formula (11), the forward probability C _it obtains the forward probability C _it as equal to forward immediately preceding frame probability C _{i (t-1)} The calculation ends. If the further distance dts is equal to or less than the threshold DTS associated with the output probability B _ji (x _t), without calculation for obtaining the output probabilities B _ji the (x _t) from equation (4) to (7), the reference probability B _Since the forward probability C _it is calculated using _ji , the amount of calculation can be significantly reduced. Moreover, such simplification of the calculation is performed when the distance dtc related to the forward probability C _it is less than or equal to the threshold DTC or when the distance dts related to the output probability B _ji (x _t ) is less than or equal to the threshold DTS. Even if is simplified, the error of the forward probability C _it can be reduced.

[0369] According to the inventor of the simulation results of this application, as the amount of calculation for obtaining the forward probability C _it is approximately 1/2 of the case of not performing the simplification of calculation, the forward probability C _it defining a threshold DTC involved, and the output probability B _ji (x _t) calculation amount for obtaining the found to be about 1/5 of the case of not performing the simplification of the operation, the output probability B _ji (x _t)
Even if the threshold value DTS related to is determined, the accuracy of voice recognition is hardly decreased.

<Second Embodiment of the Invention of Claim 7> As a voice recognition apparatus suitable for carrying out the second embodiment of the invention of claim 7, the collation unit 42 is configured as follows. The voice recognition device 34 having the same configuration as the above can be used.

That is, the collation unit 42 uses the reference probability b _ji stored in the reference information storage unit 44 when obtaining the likelihood, and the forward probability in each case of t = 1, 2, ..., T. c
_It is calculated sequentially as follows.

(1). When t = 1, the forward probability reference frame number qc and the output probability reference frame number qs are set to 1, respectively, and the forward probability skip number skipc,
Is initialized to output probability skip number skips to 0 respectively, all j, for i, the output probability b _ji (x _t) from hidden Markov model determined reference probability b _ji the output probability b _ji (x _t) The reference probability b is written as the initial value.
After the writing of _ji is finished, the reference probability b _ji is read out and the forward probability c _it is calculated (3A). After the processing (3A) is completed, processing (3B) for adding 1 to the current frame number t is performed.

(2). When the 2 ≦ t ≦ T, the forward probability skip number skipc with is compared with a threshold value NSKIPC, the distance between the speech feature vector x _qc of forward probabilities reference frame number qc and speech feature vectors x _t of the current frame number t Process of comparing dtc with threshold DTC (3C)
And the comparison result of this processing (3C) is skipc ≦ NSK
Forward probability c when IPC and dtc ≦ DTC
_It is assumed that _it is equal to the forward probability c _{i (t−1)} of the immediately preceding frame, and the calculation of the forward probability c _it is completed, and the forward probability skip number skippc and the output probability skip number sk
When the comparison result of the process (3D) of adding 1 to ips and the process (3C) is skipc> NSKIPC or dtc> DTC, the forward probability skip number skippc is initialized to 0, and the forward probability. A process (3E) of rewriting the reference frame number qc to the current frame number t is performed.

[0374] Then after the processing (3E), together with comparing the output probability skip number skips a threshold NSKIPS, the speech feature vector x _t and the output probability reference frame number qs of the current frame number t of the audio feature vector x _qs The distance dts between them is compared with the threshold value DTS, and the comparison result is s.
When kips> NSKIPS or dts> DTS, the output probability skip number skips is initialized to 0, the output probability reference frame number qs is rewritten to the current frame number t, and the output probability b is set for all j and i. _ji (x _t) rewriting the reference probability b _ji determined from hidden Markov model to the output probability b _ji (x _t), the forward probability c _it reads each reference probability b _ji after rewriting completion of the reference probability b _ji And the comparison result is skips ≦ N.
When SKIPS and dts ≦ DTS are satisfied, 1 is added to the output probability skip number skips, and each reference probability b _ji is read without rewriting the reference probability b _ji (x _t ) to obtain the forward probability c _it . Process (3F) is performed.

After the processing (3D) or (3F), the processing (3G) of adding 1 to the current frame number t is performed.

Next, in the second embodiment of the invention of claim 7, the time series x ₁ , x ₂ , ... Of the HMM and the voice feature vector.
,, the flow of the process of obtaining the likelihood between x _T and 1
The flow of processing focused on individual HMMs will be described. 20 to 22 are diagrams showing the flow of processing focusing on this one HMM. In this example, the output probability b _ji (x _t ), the forward probability c _it, and the reference probability b _ji are logarithmized output probability B _ji (x _t ), logarithmic forward probability C _it, and logarithmic reference probability, respectively. B _ji and i = j = 1, 2, ..., I.

When the section information and the speech feature vector x _t are input from the speech section detecting section 40, the collating section 42 receives i =
For all i of 1, 2, ..., I, the initial value C _i0 of the forward probability logarithmized is set according to the equation (10) (S).
1).

Next, the collating unit 42 initializes the current frame number t to t = 1 in order to process the start frame of the voice section (S2).

Next, the collation unit 42 determines that j = 1, 2, ..., J
And i = 1, 2, ..., I, logarithmized output probabilities B _ji (x ₁ ) are obtained according to equations (4) to (7) (S3), and the output probabilities B _{ji are obtained.} (x ₁ ) is written as the initial value of the logarithmic output probability B _ji (S4).

In the reference information storage section 44, j = 1, 2, ...
, J and i = 1, 2, ..., I, respectively, save B _ji is provided as a storage area for storing the reference probability B _ji for each j and i. Therefore, the reference information storage unit 44 determines that the output reference probability B
₁₁ , B ₁₂ , ..., B _1I , B ₂₁ , B ₂₂ , ..., B _2I , ...
, B _J1 , B _J2 , ..., B _JI are respectively stored in J × I storage areas. Therefore, in the figure, the process of storing the initial value of the reference probability B _ji is performed as save B _ji = B
It is expressed as _ji (x ₁ ).

Next, the collation unit 42 initializes the forward probability reference frame number qc and the output probability reference frame number qs to the current frame number 1, and sets the forward probability skip number skipc and the output probability skip number skips to 0, respectively. Initialize (S5). After that, the collating unit 42
_{Calculates the} logarithmic forward probability C _{i1 according} to the equation (11) for all i = 1, 2, ..., I (S
6).

Next, the collation unit 42 adds 1 to the current frame number t in order to process the next frame of the voice section (S7), and thereafter, the current frame number t and the frame number T of the end frame are compared. Are compared to determine whether the processing has been completed for all the frames in the voice section (S8).

(3-2A: When t ≦ T in S8) If the current frame number t is equal to or less than the end frame number T in S8, the processing has not been completed for all the frames in the voice section, so that it is forward-looking. Probability skip number skipc and threshold N
A comparison determination with SKIPC is performed (S9).

If the number of forward probability skips skippc exceeds the threshold value NSKIPC in S9, the forward probability C _it of the current frame number t is approximated by the forward probability C _{i (t-1)} of the immediately preceding frame to obtain the forward probability C _it . Number of times the calculation is completed s
Since kipc exceeds the threshold value NSKIPC, the time gap between the current frame number t and the forward probability reference frame number qc becomes large, and thus the error is likely to increase.
Therefore, since the reference probability B _ji is read to obtain the forward probability C _it , the forward probability skip number skipc is initialized to 0 and the forward probability reference frame number q is set.
c is rewritten to the current frame number t (S10).

Further, in S9, the number of forward probability skips skip
If c is less than or equal to the threshold value NSKIPC, the matching unit 42
Is determined according to equation (19) the distance dtc between the speech feature vector x _qc of forward probabilities reference frame number qc and speech feature vectors x _t of the current frame number t (S11),
The obtained distance dtc is compared with a threshold value DTC to determine whether or not these vectors x _t and x _qc are approximately equal (S
12).

If the distance dtc exceeds the threshold value DTC in S12, the voice feature vector x _t of the current frame number t is the voice feature vector x qc of the forward probability reference frame number _qc.
Therefore, since the voice feature vector x _t of the current frame number t has a large change from the voice feature vector x _qc of the forward probability reference frame number qc, the forward probability C _it of the current frame number t is the forward probability of the immediately preceding frame. It cannot be approximated by C _{i (t-1)} . Therefore, since the reference probability B _ji is read to obtain the forward probability C _it , the forward probability skip number skipc is initialized to 0 and the forward probability reference frame number qc is rewritten to the current frame number t (S
10).

If the distance dtc is less than or equal to the threshold DTC in S12, skippc≤NSKIPC and dtc≤
This is the case of DTC. For skipc ≦ NSKIPC, since the number of times Skipc completing the calculation of forward probabilities C _it is approximated by the forward probability C _{i (t-1)} of the immediately preceding frame forward probability C _it the current frame number t exceeds the threshold NSKIPC , The time gap between the current frame number t and the forward probability reference frame number qc is small, and therefore the error is unlikely to increase. Moreover, when dtc ≦ DTC, the voice feature vector x _t of the current frame number t is approximately equal to the voice feature vector x _qc of the forward probability reference frame number qc, and thus the voice feature vector x t of the current frame number _t is the forward probability reference. Voice feature vector x _qc of frame number _qc
Since the change from is small, the forward probability C _it of the current frame number t can be approximated by the forward probability C _{i (t-1)} of the immediately preceding frame. Therefore, neither the reference probability B _ji is read out to obtain the forward probability C _it nor the reference probability B _ji is rewritten. Therefore, the forward probability C _it of the current frame number t is assumed to be equal to the forward probability C _{i (t-1)} of the immediately preceding frame, and the calculation for obtaining the forward probability C _it of the current frame number t is completed, and the forward probability skip number skippc, Output probability skip number s
These skip numbers sk are obtained by adding 1 to each of the skips.
Count up ipc and skips respectively (S
13). After that, the process returns to S7 so as to perform the process for the next frame of the voice section.

The number of forward probability skips skipc is a threshold N.
Exceeds SKIPC or distance dtc is threshold DTC
If S10 is performed when the value exceeds the threshold, then the output probability skip number skips and the threshold value NSKIPS are compared and determined (S14).

If the output probability skip number skips exceeds the threshold value NSKIPS in S14, the number of times skips at which the reference probability B _ji is not rewritten is the threshold value NSKIPS.
Since the current frame number t exceeds the output probability reference frame number qs, there is a high possibility that the error increases. Therefore, in order to reduce the error, the reference probability B _ji is rewritten. Therefore, the output probability skip number skips is initialized to 0 and the output probability reference frame number qs is rewritten to the current frame number t (S15). Then, j = 1, 2, ..., J and i =
For all j, i of 1, 2, ..., I, logarithmic output probabilities B _ji (x _t ) are obtained according to equations (4) to (7), and reference probabilities B _ji are output probabilities B _ji ( x _t ) (S1
6). Then, after the reference probabilities B _ji have been rewritten, the reference probabilities B _ji are read out, and all i _values of i = 1, 2, ..., I are read.
Then, the forward probability C _it is calculated according to the equation (11) (S17). After that, the process returns to S7 so as to perform the process for the next frame of the voice section. In the figure,
The process of rewriting the reference probability B _{ji in} S16 is save B _ji = B
_{Expressed as ji} (x _t ).

Reference probability B read in S17 in this case
_ji is the output probability B _ji (x _t ) of the current frame number t obtained in S16. Therefore, in S17 in this case, the forward probability C _it is used by using the output probability B _ji (x _t ) of the current frame number t.
Will be asked.

[0391] If S14 in output probability skip number skips is equal to or less than the threshold NSKIPS is, the matching unit 42, between the speech feature vector x _qs speech feature vector x _t and the output probability reference frame number qs of the current frame number t Of the calculated distance dts (S18)
It is determined by comparing with S whether these vectors x _t and x _qs are approximately equal (S19).

When the distance dts exceeds the threshold DTS in S19, the voice feature vector x _t of the current frame number t is the voice feature vector x qs of the output probability reference frame number _qs.
Therefore, since the voice feature vector x _t of the current frame number t has a large change from the voice feature vector x _qs of the output probability reference frame number qs, the output probability B _ji (x _t ) of the current frame number t is referred to. It cannot be approximated with the probability B _ji . Therefore, the reference probability B _ji is rewritten. So S
The processes of 15 to S17 are performed, and thereafter, the process returns to the process of S7 to perform the process for the next frame of the voice section.

If the distance dts is less than or equal to the threshold value DTS in S19, the voice feature vector x _t of the current frame number _t.
Is the voice feature vector x of the output probability reference frame number qs
Since the speech feature vector x _t of approximately equal therefore current frame number t to _qs small changes from speech feature vector x _qs output probabilities reference frame number qs, output probability B _ji of the current frame number t (x _t) Can be approximated by the reference probability B _ji .
Therefore, the reference probability B _ji is not rewritten.
Therefore, 1 is added to the output probability skip number skips to count up the output probability skip number skips (S
20). After that, the output probability B _ji (x _t ) is _calculated by the equations (4) to (7).
The reference probability B _ji is read out without calculation using
The logarithmic forward probability C _it is calculated for all i of 1, 2, ..., I according to the equation (11) (S17). After that, the process returns to S7 so as to perform the process for the next frame of the voice section.

Reference probability B read in S17 in this case
_ji is the output probability B _ji (x _t ) obtained in the frame of the output probability reference frame number qs. Therefore, in S17 in this case, the output probability B _ji (x _qs ) of the output probability reference frame number qs is obtained.
Then, the forward probability C _it is obtained.

(3-2B: When t> T in S8) When the current frame number t is larger than the frame number T of the end frame in S8, i = 1, 2, ... since finished sought forward probability C _iT, the maximum forward probability C _iT of i = * i consisting forward probability C _iT according to equation (9), the time series x _1, x ₂ of the speech feature vector, ...,
Likelihood ln {P (x ₁ , x ₂ , ..., Between x _T and HMM
x _T )}, and after that, the process of calculating the likelihood for the HMM is terminated (end).

The collating unit 42, for all HMMs stored in the dictionary unit 36, for each HMM, see FIG.
The likelihood (forward probability C _iT ) is obtained by performing the processes of S1 to S20 shown in FIG. 2, and the maximum likelihood is detected from the obtained likelihoods. The largest category of likelihood the resulting HMM, time series x _1, x ₂ of the audio feature vector, ..., as the recognition result for the input speech signal obtained by extracting the x _T, without the next stage of the device (shown ).

As described above, the likelihood ln {P (x ₁ , x ₂ , ...
, X _T )} = C _{iT In the} process of obtaining C _iT , the number of skips skipc related to the forward probability C _it is a threshold value NSKIPC.
When the distance dtc is equal to or less than and the distance dtc is equal to or less than the threshold DTS, the calculation of obtaining the output probability B _ji (x _t ) from the equations (4) to (7) is also performed using the forward probability C _it from the equation (3) or the equation (11). Without performing the calculation, the forward probability C _it is assumed to be equal to the forward probability C _{i (t-1)} of the immediately preceding frame, and the calculation of the forward probability C _it ends. Further, when the skip number skips related to the output probability B _ji (x _t ) is less than or equal to the threshold value NSKIPS and the distance dts is less than or equal to the threshold value DTS, the output probability B _ji (x _t ) is obtained from the equations (4) to (7). Since the forward probability C _it is calculated without performing calculation, the amount of calculation can be significantly reduced. Moreover, such a simplification of the calculation is performed by setting the skip number skipc related to the forward probability C _it to be the threshold value NSK.
When the distance is less than IPC and the distance dtc is less than the threshold DTC, or the skip number skip related to the output probability B _ji (x _t ).
Since s is equal to or less than the threshold value NSKIPS and the distance dts is equal to or less than the threshold value DTS, the error of the forward probability C _it can be reduced even if the calculation is simplified.

The invention of claim 7 can be applied to all speech recognition apparatuses that perform matching processing in frame units.

[0399]

As is apparent from the above description, according to the voice recognition method of the invention of claim 1, the voice feature vector x _t of the current frame number t and the voice feature vector x qs of the reference frame number _qs are used. When the distance dts between them is equal to or smaller than the threshold value DTS (dts ≦ DTS), the output probability b of the current frame number t is calculated without rewriting the reference probability b _ji.
Since the reference probability b _ji is read out and the forward probability c _it of the current frame number t is calculated without performing the calculation for obtaining _ji (x _t ) from the Hidden Markov model, the amount of calculation can be greatly reduced.

[0400] Moreover if the dts ≦ DTS, since speech feature vector x _t of the current frame number t is the change in the speech feature vector x _qs reference frame number qs is small, the output probability b _ji the current frame number t ( x _t ) is the reference probability b
It can be approximated by _ji . Therefore, even if the forward probability c _it is calculated by simplifying the calculation when dts ≦ DTS,
The error of the forward probability c _it can be reduced.

Therefore, the likelihood ln when performing speech recognition
In the process of _obtaining {P (x ₁ , x ₂ , ..., X _T )} = C _iT , _it is possible to reduce the error of the forward probability c _it and simplify the calculation. Can perform voice recognition.

Further, according to the speech recognition method of the present invention, when the type s of the transition source S _j that gives the output probability b _ji (x _t ) of the current frame number t is a stationary part, the current frame number is If the distance dts between the voice feature vector x _{t of t} and the voice feature vector x _qs of the stationary part reference frame number qs is less than or equal to the threshold DTS (dts ≦ DTS), refer to j for which the type s is obtained. The reference probability b _ji is calculated without performing the calculation of the output probability b _ji (x _t ) of the current frame number t from the Hidden Markov model according to the rewriting of the probability b _ji.
ji is read to obtain the forward probability c _it of the current frame number t. Further, when the type s of the transition source S _j that gives the output probability b _ji (x _t ) of the current frame number t is the transition part, the voice feature vector x _t of the current frame number t and the voice of the transition part reference frame number qt If the distance dtt from the feature vector x _qt is less than or equal to the threshold value DTT (dtt ≦ DTT), the output probability of the current frame number t according to the reference probability b _ji is not rewritten for j that obtained the type s. The forward probability c _it of the current frame number t is read out by reading out the reference probability b _ji without performing the operation of obtaining b _ji (x _t ) from the Hidden Markov model.
Ask for. As described above, if dts ≦ DTS in the case of the stationary part and dtt ≦ DTT in the case of the transient part, the reference probability b
_{Since the} forward probability c _it is calculated without rewriting _ji ,
The amount of calculation can be greatly reduced.

If dts≤DTS, the voice feature vector x _t of the current frame number _t is the stationary part reference frame number q.
Since the change of the voice feature vector x of s from the voice feature vector x _qs is small, the output probability b _ji (x _t ) of the current frame number t can be approximated by the reference probability b _ji for j having obtained the type s. Also dtt
If ≦ DTT, since speech feature vector x _t of the current frame number t is small variation from the speech feature vector x _qt transient portion reference frame number qt, with respect to j that give the type s, the current frame number t The output probability b _ji (x _t ) can be approximated by the reference probability b _ji . Therefore, dts
Also it is determined ≦ DTS or forward probability c _it to simplify the operation in the case of dtt ≦ DTT, can reduce an error of the forward probability c _it.

Because of this, the likelihood ln at the time of performing speech recognition
In the process of _obtaining {P (x ₁ , x ₂ , ..., X _T )} = C _iT , _it is possible to reduce the error of the forward probability c _it and simplify the calculation. Can perform voice recognition.

[0405] Further according to the speech recognition method of the invention of claim 7, the following distance dtc threshold DTC between the speech feature vector x _qc of forward probabilities reference frame number qc and speech feature vectors x _t of the current frame number t (Dtc ≦ D
TC), the forward probability c of the current frame number t
_It is assumed that _it is equal to the forward probability c _{i (t-1)} of the immediately preceding frame, and the calculation for obtaining the forward probability c _it ends. In addition, the distance dtc exceeds the threshold value DTC (dtc> DTC).
If the following distance dts threshold DTS between speech feature vector x _qs speech feature vector x _t and the output probability reference frame number qs of the current frame number t (dts ≦ DTS)
If the reference probability b _ji the output probability b _ji the current frame number t in accordance without rewriting (x _t) of without operation for obtaining from Hidden Markov Models, reference probability b _ji reads in the current frame number t The forward probability c _{it of} is calculated. In this way, when dtc ≦ DTC or dts ≦ DTS, the forward probability c _it is obtained without rewriting the reference probability b _ji , so the amount of calculation can be greatly reduced.

[0406] Moreover when the dtc ≦ DTC, the change from speech feature vector x _qc speech feature vector x _t is the forward probability reference frame number qc of the current frame number t is small, the forward probability c of the current frame number t _It can be approximated by the forward probability c _{i (t-1)} of the immediately preceding frame. Also dt
When s ≦ DTS or less, since the voice feature vector x _t of the current frame number t has a small change from the voice feature vector x _qs of the reference frame number qs, the current frame number t
The output probability b _ji (x _t ) can be approximated by the reference probability b _ji . Therefore, even when the forward probability c _it is obtained by simplifying the calculation in the case of dtc ≦ DTC or dts ≦ DTS, the error of the forward probability c _it can be reduced.

Because of this, the likelihood ln when performing speech recognition
In the process of _obtaining {P (x ₁ , x ₂ , ..., X _T )} = C _iT , _it is possible to reduce the error of the forward probability c _it and simplify the calculation. Can perform voice recognition.

[Brief description of the drawings]

FIG. 1 is a diagram showing an example of a device configuration suitable for use in carrying out the invention of claim 1;

FIG. 2 is a diagram for explaining a Hidden Markov model.

FIG. 3 is a flowchart for explaining the first embodiment of the invention of claim 1;

FIG. 4 is a flowchart for explaining the first embodiment of the invention of claim 1;

FIG. 5 is a flowchart for explaining the second embodiment of the invention of claim 1;

FIG. 6 is a flowchart for explaining the second embodiment of the invention of claim 1;

FIG. 7 is a diagram showing an example of a device configuration suitable for use in carrying out the invention of claim 3;

FIG. 8 is a diagram for explaining a Hidden Markov model.

FIG. 9 is a flowchart for explaining the first embodiment of the invention of claim 3;

FIG. 10 is a flowchart for explaining the first embodiment of the invention of claim 3;

FIG. 11 is a flowchart for explaining the first embodiment of the invention of claim 3;

FIG. 12 is a flowchart for explaining the second embodiment of the invention of claim 3;

FIG. 13 is a flowchart for explaining the second embodiment of the invention of claim 3;

FIG. 14 is a flowchart for explaining the second embodiment of the invention of claim 3;

FIG. 15 is a diagram showing an example of a device configuration suitable for use in implementing the invention of claim 7;

FIG. 16 is a diagram for explaining a Hidden Markov model.

FIG. 17 is a flowchart for explaining the first embodiment of the invention of claim 7;

FIG. 18 is a flowchart for explaining the first embodiment of the invention of claim 7;

FIG. 19 is a flowchart for explaining the first embodiment of the invention of claim 7;

FIG. 20 is a flowchart for explaining the second embodiment of the invention of claim 7;

FIG. 21 is a flowchart for explaining the second embodiment of the invention of claim 7;

FIG. 22 is a flow chart for explanation of a second embodiment of the invention of claim 7;

[Explanation of symbols]

 10, 22, 34: Speech recognition device 12, 24, 36: Dictionary section 14, 26, 38: Sound processing section 16, 28, 40: Speech section detection section 18, 30, 42: Collation section 20, 32, 44: Reference information storage

Claims (8)

[Claims] 1. A time series x ₁ , x of voice feature vectors extracted from a start frame to an end frame of a voice section.
₂ , ……, the likelihood ln between x _T and Hidden Markov model
{P (x ₁ , x ₂ , ..., X _T )} is calculated, and the category assigned to the Hidden-Markov model that has the maximum likelihood is used as the recognition result for the speech signal in the speech section. In the recognition method, However, i: i = 1, 2, ..., I j: j = 1, 2, ..., J Φ _i : Probability that initial state is S _i in Hidden Markov model a _ji : State S in Hidden Markov model Probability of transition from _j to state S _i x _t : speech feature vector extracted in the t-th frame in the speech section (1 ≦ t ≦ T, the first frame is the start frame of the speech section) And the T-th frame represents the end frame of the speech section) b _ji (x _t ): Speech feature vector x output when transitioning from state S _j to state S _i in the Hidden Markov model
Output probability of _t c _it : Time series of speech feature vector x ₁ , x ₂ , ..., Which starts transition from the initial state in Hidden Markov model
Forward probability of outputting x _t and reaching state S _i * i: Likelihood ln {P (x (x)) is obtained by using each equation represented by state number i given to state S _i that is the final state in the Hidden Markov model. ₁ , x ₂ , ...
, X _T )}, a storage unit for storing the reference frame number qs and the reference probability b _ji is provided, and using the reference probability b _ji , t = 1, 2, ...
, The forward probability c _it in each case of T is sequentially obtained,
(1). When the t = 1, is initialized to 1 reference frame number qs, all j, for i, the output probability b _ji the (x _t) determined from the hidden Markov model the output probability b _ji the (x _t) writing the initial value of the reference probability b _ji, reads each reference probability b _ji after completion of writing of the reference probability b _ji seek forward probability c _it processing (1A), after the end of the process (1A), the current frame The process (1B) of adding 1 to the number t is performed, and (2). 2 ≦ t when the ≦ T, the distance between the speech feature vector x _qs speech feature vector x _t and the reference frame number qs of the current frame number t dts
Is compared with a threshold DTS, and when the comparison result is dts> DTS, the reference frame number qs is set to the current frame number t.
And the output probability b _ji (x _t ) for all j and i is calculated from the Hidden Markov model and the reference probability b
_ji is rewritten to the output probability b _ji (x _t ) and the reference probability b _ji
After the end of rewriting, each reference probability b _ji is read out to obtain the forward probability c _it , and when the comparison result is dts ≦ DTS, each reference probability b _ji is not rewritten.
Processing to read _ji and obtain forward probability c _it (1C)
And a process (1D) of adding 1 to the current frame number t after the process (1C) is completed. 2. The voice recognition method according to claim 1, wherein
(1). When t = 1, the reference frame number qs is set to 1 and the skip number ski
Initialize ps to 0, and for all j, i,
Writing output probability b _ji (x _t) the probability that output determined from Hidden Markov Models b _ji the (x _t) as the initial value of the reference probability b _ji, reads out the reference probability b _ji after completion of writing of the reference probability b _ji Processing (1A) for obtaining the forward probability c _it , and processing (1B) for adding 1 to the current frame number t after completion of the processing (1A), (2). 2 ≦ t when the ≦ T, together with comparing the number of skips skips a threshold NSKIPS, threshold DTS distance dts between the speech feature vector x _qs speech feature vector x _t and the reference frame number qs of the current frame number t And the comparison result is skips.
> NSKIPS or dts> DTS,
The skip number skips is initialized to 0, the reference frame number qs is rewritten to the current frame number t, and
For all j and i, the output probability b _ji (x _t ) is obtained from the Hidden Markov model, and the reference probability b _ji is set as the output probability b _ji.
rewritten (x _t), determine the forward probability c _it reads each reference probability b _ji after rewriting completion of the reference probability b _ji, if the comparison result is skips ≦ NSKIPS and dts ≦ DTS, skip number Skips in conjunction with adding 1, reference probability b _ji each reference probability b _ji without rewriting the
And a process (1C) for obtaining the forward probability c _it and a process (1D) for adding 1 to the current frame number t after completion of the process (1C). 3. A time series x ₁ , x of voice feature vectors extracted from the start frame to the end frame of a voice section.
₂ , ……, the likelihood ln between x _T and Hidden Markov model
{P (x ₁ , x ₂ , ..., X _T )} is calculated, and the category assigned to the Hidden-Markov model that has the maximum likelihood is used as the recognition result for the speech signal in the speech section. In the recognition method, However, i: i = 1, 2, ..., I j: j = 1, 2, ..., J Φ _i : Probability that initial state is S _i in Hidden Markov model a _ji : State S in Hidden Markov model Probability of transition from _j to state S _i x _t : speech feature vector extracted in the t-th frame in the speech section (1 ≦ t ≦ T, the first frame is the start frame of the speech section) And the T-th frame represents the end frame of the speech section) b _ji (x _t ): Speech feature vector x output when transitioning from state S _j to state S _i in the Hidden Markov model
Output probability of _t c _it : Time series of speech feature vector x ₁ , x ₂ , ..., Which starts transition from the initial state in Hidden Markov model
Forward probability of outputting x _t and reaching state S _i * i: Likelihood ln {P (x (x)) is obtained by using each equation represented by state number i given to state S _i that is the final state in the Hidden Markov model. ₁ , x ₂ , ...
,, x _T )}, the state S _j that is the transition source in the Hidden Markov model
Is provided with a stationary part or transient part type s, a stationary part reference frame number qs, a transient part reference frame number qt, and a reference probability b _ji are provided, and the reference probability b _ji is provided. , The forward probability c _it for each case of t = 1, 2, ..., T is sequentially obtained, and (1). When t = 1, the stationary part reference frame number qs and the transient part reference frame number qt are initialized to 1, and the output probabilities b _ji (x _t ) are calculated from the Hidden Markov model for all j and i. writing the output probability b _ji the (x _t) as the initial value of the reference probability b _ji, reads each reference probability b _ji after completion of writing of the reference probability b _ji seek forward probability c _it processes (2
A) and a process (2B) of adding 1 to the current frame number t after completion of the process (2A), (2). 2 ≦ t ≦ T when the distance d between the speech feature vector _qs speech feature vector x _t and the constant part reference frame number qs of the current frame number t
ts is compared with the threshold value DTS, and the comparison result is dts> DT
When the S, processing of rewriting the constant part reference frame number qs to the current frame number t and (2C), the speech feature vector x _qt speech feature vector x _t and the transient portion reference frame number qt the current frame number t The distance dtt between them is compared with the threshold value DTT, and the comparison result shows that dtt> D.
In the case of TT, the process (2D) of rewriting the transition part reference frame number qt to the current frame number t, and j = 1, 2, after the processes (2C) and (2D) are finished.
.., for each j of J, a process (2E) of determining the type s _assigned to the transition source S _j of the state transition that gives the output probability b _ji (x _t ), and the type of the process (2E) When the determination result is the stationary part and the comparison result of the process (2C) is dts> DTS, the output probabilities b _ji (x _t ) for all i with respect to j for which the type determination result is obtained. From the Hidden Markov model, the reference probability b _ji is rewritten to the output probability b _ji (x _t ), and when the type determination result of the process (2E) is a stationary part, the comparison result of the process (2C) Is dts ≦ DTS, the reference probability b _ji is not rewritten for j for which the type determination result is obtained, and if the type determination result of the process (2E) is a transient part, the process (2D If the comparison result of) is dtt> DTT, it is related to j that has obtained the type determination result. For all Te is the i, rewritten to the output probability b _ji (x _t) The hidden Markov model reference probability b _ji determined from the output probability b _ji (x _t), type determination result of the processing (2E) transient If the comparison result of the process (2D) is dtt ≦ DTT, the process (2F) in which the reference probability b _ji is not rewritten for j that has obtained the type determination result, and j = The above process (2F) for each j of 1, 2, ..., J
When the process (2F) is completed for all j, the reference probability b _ji is read to obtain the forward probability c _it (2G), and after the process (2G) is completed, the current frame number t is set. A voice recognition method characterized by performing a process of adding 1 (2H). 4. The voice recognition method according to claim 3, wherein the process (2E) is performed after the processes (2C) and (2D) are completed. 5. The voice recognition method according to claim 3,
(1). When t = 1, the number of skips in the steady part skips and the number of skips in the transient part ski
pt to 0 and the reference frame number q of the stationary part
s and the transition part reference frame number qt are initialized to 1, respectively, and the output probabilities b _ji (x _t ) for all j and i.
From the Hidden Markov model, the output probability b _ji (x _t )
_{Is written as} the initial value of the reference probability b _ji , and the reference probability b
After the writing of _ji , each reference probability b _ji is read out to obtain the forward probability c _it (2A), and after the completion of the process (2A), a process of adding 1 to the current frame number t (2B) is performed. , (2). When 2 ≦ t ≦ T, the constant part skip number skips is compared with the threshold value NSKIPS, and the voice feature vector x _t of the current frame number _t
And the voice feature vector x _{qs of the} stationary part reference frame number _qs
When the comparison result is skips> NSKIPS or dts> DTS, the constant part skip number skips is initialized to 0, and the constant part reference frame number qs is calculated. The frame number t is rewritten, and the comparison result is skips ≦ NSKIP.
When S and dts ≦ DTS, the process (2C) of adding 1 to the constant part skip number skips is compared with the transient part skip number skipt and the threshold NSKIPT, and the voice feature vector x _t of the current frame number _t
And the voice feature vector x _{qt of the} transition part reference frame number _qt
When the comparison result is skip> NSKIPT or dtt> DTT, the transition skip number skippt is initialized to 0, and the transition reference frame number qt is calculated. Rewrite to frame number t, and the comparison result is skipt ≦ NSKIP
When T and dtt ≦ DTT, the process (2D) of adding 1 to the transition part skip number skipt, and j = 1, 2, ... After the processes (2C) and (2D) are finished.
, Processing for determining the type s _assigned to the transition source S _j of the state transition that gives the output probability b _ji (x _t ) for each j of J, and type determination for this processing (2E) When the result is the stationary part, the comparison result of the process (2C) is skips> NSK.
If IPS or dts> DTS, for j for which the type determination result is obtained, output probability b for all i
_ji (x _t ) is calculated from Hidden Markov model and the reference probability is b _ji
To the output probability b _ji (x _t ), and if the type determination result of the process (2E) is the stationary part, the process (2
The comparison result of C) is skips ≦ NSKIPS and dts
If ≦ DTS, the reference probability b _ji is not rewritten for j for which the type determination result is obtained, and if the type determination result of the process (2E) is a transient part, the process (2
The comparison result of D) is skip> NSKIPT or d
If tt> DTT, the output probability b _ji (x _t ) is obtained from the Hidden-Markov model for all i for j for which the type determination result is obtained, and the reference probability b _{ji is set} for the output probability b _ji.
rewritten _ji (x _t), when the type determination result of the processing (2E) was a transient portion, the comparison result of the processing (2D) is s
If kipt ≦ NSKIPT and dtt ≦ DTT,
For j for which the type determination result has been obtained, the processing (2F) in which the reference probability b _ji is not rewritten and the processing (2F) is performed for each j of j = 1, 2, ... When the processing (2F) is completed for j,
It is characterized by performing processing (2G) for reading out each reference probability b _ji to obtain a forward probability c _it , and processing (2H) for adding 1 to the current frame number t after the processing (2G) is completed. Speech recognition method. 6. The voice recognition method according to claim 5, wherein the process (2E) is performed after the processes (2C) and (2D) are completed. 7. A time series x ₁ , x of voice feature vectors extracted from the start frame to the end frame of a voice section.
₂ , ……, the likelihood ln between x _T and Hidden Markov model
{P (x ₁ , x ₂ , ..., X _T )} is calculated, and the category assigned to the Hidden-Markov model that has the maximum likelihood is used as the recognition result for the speech signal in the speech section. In the recognition method, However, i: i = 1, 2, ..., I j: j = 1, 2, ..., J Φ _i : Probability that initial state is S _i in Hidden Markov model a _ji : State S in Hidden Markov model Probability of transition from _j to state S _i x _t : speech feature vector extracted in the t-th frame in the speech section (1 ≦ t ≦ T, the first frame is the start frame of the speech section) And the T-th frame represents the end frame of the speech section) b _ji (x _t ): Speech feature vector x output when transitioning from state S _j to state S _i in the Hidden Markov model
Output probability of _t c _it : Time series of speech feature vector x ₁ , x ₂ , ..., Which starts transition from the initial state in Hidden Markov model
Forward probability * i that outputs x _t and reaches the state S _i : Likelihood ln {P (x (x) is obtained by using each equation represented by the state number i assigned to the final state S _i in the Hidden Markov model. ₁ , x ₂ , ...
, X _T )}, a storage unit for storing the forward probability reference frame number qc, the output probability reference frame number qs, and the reference probability b _ji is provided, and using the reference probability b _ji , t = Forward probabilities c _it for each case of 1, 2, ..., T are sequentially obtained, and (1). t = 1
, The forward probability reference frame number qc and the output probability reference frame number qs are initialized to 1, and the output probabilities b _ji (x _t ) for all j and i are calculated from the Hidden Markov model. write b _ji the (x _t) as the initial value of the reference probability b _ji, reads each reference probability b _ji after completion of writing of the reference probability b _ji seek forward probability c _it processing (3A), the processing (3A After the end of (), the process of adding 1 to the current frame number t (3B) is performed, and (2). When the 2 ≦ t ≦ T, the process of comparing with a threshold DTC distance dtc between the speech feature vector x _qc of forward probabilities reference frame number qc and speech feature vectors x _t of the current frame number t (3C), When the comparison result of the processing (3C) is dtc ≦ DTC, the forward probability c _it is the forward probability c of the immediately preceding frame.
_When the comparison result of the processing (3D) for terminating the calculation of the forward probability c _it as equal to _{i (t-1)} and the processing (3C) is dtc> DTC, the forward probability reference frame number qc is set. and processing (3E) for rewriting the current frame number t, after the end of the process (3E), the distance between the speech feature vector x _qs speech feature vector x _t and the output probability reference frame number qs of the current frame number t dts Is compared with the threshold DTS, and when the comparison result is dts> DTS, the output probability reference frame number qs is rewritten to the current frame number t, and the output probabilities b _ji (x _t ) are calculated for all j and i. hidden rewritten reference probability b _ji determined from Markov model to the output probability b _ji (x _t), determine the forward probability c _it reads each reference probability b _ji after rewriting completion of the reference probability b _ji, the comparison If the result is dts ≦ DTS, without rewriting of the reference probability b _ji reads each reference probability b _ji seek forward probability c _it processing (3F), the processing (3D) or (3F) A voice recognition method characterized by performing a process (3G) of adding 1 to the current frame number t after the end. 8. The voice recognition method according to claim 7, wherein
(1). When t = 1, the forward probability reference frame number qc and the output probability reference frame number qs are initialized to 1, and the forward probability step number skipc and the output probability step number skips are initialized to 0, and all j are initialized. , I, the output probability b _ji (x _t ) is obtained from the Hidden Markov model, and the output probability b _ji (x _t ) is written as the initial value of the reference probability b _ji ,
And processing (3A) for obtaining the forward probability c _it reads each reference probability b _ji after completion of writing of the reference probability b _ji, after completion of the process (3A), the process of adding 1 to the current frame number t (3B ) And (2). When the 2 ≦ t ≦ T, the forward probability skip number skipc with is compared with a threshold value NSKIPC, the distance between the speech feature vector x _qc of forward probabilities reference frame number qc and speech feature vectors x _t of the current frame number t The process (3C) of comparing dtc with the threshold value DTC and the comparison result of the process (3C) are skipc ≦ NSKIPC.
When dtc ≦ DTC, the forward probability c _it is assumed to be equal to the forward probability c _{i (t-1)} of the immediately preceding frame, and the calculation of the forward probability c _it is completed, and the forward probability skip number skippc, the output probability Skip number skip
The comparison result of the process (3D) of adding 1 to s and the process (3C) is skippc> NSKIPC.
Alternatively, when dtc> DTC, the forward probability skip number skippc is initialized to 0, and the forward probability reference frame number qc is rewritten to the current frame number t (3E), and after the process (3E), Output probability skip number skip
s with a comparison with a threshold NSKIPS, the distance dts between the speech feature vector x _qs output probabilities reference frame number qs with speech feature vectors x _t of the current frame number t is compared with a threshold DTS, the comparison result is skips ＞ NSKI
When PS or dts> DTS, the output probability skip number skips is initialized to 0, the output probability reference frame number qs is rewritten to the current frame number t, and the output probability b _ji (for all j and i). x _t ) is calculated from the Hidden Markov model, the reference probability b _ji is rewritten to the output probability b _ji (x _t ), and each reference probability b _ji is read out after the reference probability b _ji is rewritten to obtain the forward probability c _it . ,
The comparison result is skips ≦ NSKIPS and dts ≦ D.
If TS, the output probability skip count is 1 in skips
Is added and the reference probability b _ji (x _t ) is not rewritten and each reference probability b _ji is read to obtain the forward probability c _it (3F), and the process (3D) or (3F) ends. Then, a process (3G) of adding 1 to the current frame number t is performed.