JPH0968996A - Voice recognition method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To increase a recognition speed in voice recognition using a hidden Markov model as a standard pattern. SOLUTION: When output probability of an input voice characteristic vector is obtained from the hidden Markov model provided with a non-correlation mixture normal distribution, the largest output probability among the output probability obtained from respective normal distributions of the non-correlation mixture normal distribution is made the output probability of the input voice characteristic vector. When the output probability obtained from respective normal distributions are calculated, when the output probability obtained from a Q-th normal distribution of the non-correlation mixture normal distribution in an immediately before frame becomes the largest one, the probability that the output probability obtained from the Q-th normal distribution becomes the largest one is high even in a next frame also. Thus, the output probability obtained from the Q-th normal distribution is made a maximum value candidate, and the output probability obtained from respective normal distributions are compared with the maximum value candidate on the way of calculation, and the calculation of the output probability is interrupted according to the comparison result.

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.

[0002]

[Prior art]

For example: Seiichi Nakagawa "Speech recognition by stochastic model"
IEICE (1988) ISBN-4-885
As disclosed in 52-072-X, in speech recognition, a hidden Markov model (HiddenMarkov Model) is used as a speech standard pattern.
(M is called) is widely used. H, which is the standard voice pattern
The MM has several states, eg S ₀ -S ₃ , and states S _i.
Is represented by probability a _ij of transition from S _j to S _j and probability b _ij (V _t ) of outputting a voice symbol vector V _{t at} the time of the transition. The output probability b _ij (V _t ) is generally represented by an uncorrelated mixed normal distribution including a plurality of normal distributions.

In voice recognition using an HMM, an input voice feature vector x _t is extracted from an input voice signal for each frame of a voice section. Then using uncorrelated Gaussian mixture of HMM, the output probability b _ij (x of the input speech feature vector x _t
t ) = Σ {λ _ijm b _ijm (x _t )} is calculated. Where λ
_ijm represents the weight of the m-th normal distribution in the decorrelated mixed normal distribution, and b _ijm represents the unweighted probability of the input speech feature vector x _t obtained from the m-th normal distribution in the decorrelated mixed normal distribution.

Then, the likelihood between the time series of the input 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 (of each input voice feature vector x _t ). x _t ). Each HM
The likelihood is calculated for each M, and the category name given to the HMM for which the maximum likelihood is obtained is used as the recognition result of the input speech signal.

[0005]

However, as the output probability b _ij (x _t ) of the input speech feature vector x _t , Σ {λ
_ijm b _ijm (x _t )} is difficult to calculate the likelihood between the time series of the input speech feature vector x _t and the HMM because the calculation amount increases. Therefore, it has been desired to obtain the output probability b _ij (x _t ) more simply while suppressing the decrease in accuracy.

[0006]

In order to solve the above-mentioned problems, the speech recognition method of the present invention uses a hidden Markov model as a speech standard pattern, and the hidden Markov model has a plurality of normal distributions that are uncorrelated with each other. Having a non-correlated mixed normal distribution that represents the output probability of the voice symbol vector output from the model, the time series of the input voice feature vector extracted from the start frame to the end frame in the voice section, and a hidden Markov model Between the input speech signal of the input speech signal of the relevant speech section, and the category name given to the Hidden Markov Model that has obtained the maximum likelihood is calculated by using the logarithmic value of the output probability of each input speech feature vector. in the speech recognition method the recognition result for, b _ij (x _t): the hidden Markov models with uncorrelated Gaussian mixture having a total number of M normal distribution The t-th output probability of the input extracted by the frame speech feature vector x _t is output (1 ≦ t ≦ T. 1st frame start frame of the speech segment, and the T-th frame speech section , G _ijm (x _t ): Weighted probability of the input speech feature vector x _t calculated from the m-th (1 ≦ m ≦ M.) Normal distribution in the total number M of normal distributions. (However, g _ijm (x _t ) = λ _ijm b _ijm (x _t ), b _ijm (x _t ) =
^{(2π) -p / 2 | ρ} ijm | -1/2 exp {-D ijmt 2/2},
D _ijmt ² = (x _t −μ _ijm ) ′ ρ _ijm ⁻¹ (x _t −μ
_ijm ), λ _ijm : weight of the m-th normal distribution, b _ijm (x _t ): unweighted probability of the input speech feature vector x _t calculated from the m-th normal distribution, p: input speech feature vector x order of _t , ρ _ijm : variance-covariance matrix of the m-th normal distribution, μ _ijm : mean vector of the m-th normal distribution, D _ijmt : input speech feature vector x _t and the m-th normal distribution Mahalavis general distance that represents the distance between. _{), G ijm (x t)} : logarithm of weighted probabilities g _ijm (x _{t) (where, G ijm (x t) =} E ijm -D ijmt 2/2, E ijm = ln (λ ijm) + ln {( 2π) ^{-p / 2} ｜ ρ _ijm ｜
^-1/2 }. ), Input the maximum logarithmic value G _ijm (x _t ) among the logarithmic values G _ijm (x _t ) of the weighted probabilities g _ijm (x _t ) calculated from each M normal distribution. When calculating the likelihood with the hidden Markov model using the logarithmic value of the output probability b _ij (x _t ) of the speech feature vector x _t , when t ≧ 2, the maximum in the t-th frame is calculated. in either the maximum value candidate for detecting logarithm G _ijm (x _t), the normal distribution was obtained maximum logarithm G _ijm the (x _t) in the first t-1 th frame at t ≧ 2 is A reference information storage unit that stores an index indicating whether or not there is is provided, and at t = 1, the logarithmic value G _ijm (x _t ) is calculated for each normal distribution with respect to all normal distributions of the total number M, and the maximum value is calculated. detecting a logarithmic value G _ijm (x _t), outermost sized logarithm G _ijm the (x _t) of the input speech feature vector x _t in the first frame Storing an index corresponding to a normal distribution was obtained outermost sized logarithm G _ijm (x _t) with the logarithm of the power probability b _ij (x _t), the t ≧ 2, (1) first corresponding to the index Logarithmic value G calculated using the normal distribution
_ijm (x _t ) is stored as the maximum value candidate, and (2) -D is calculated in calculating the logarithmic value G _ijm (x _t ) using the remaining normal distribution that does not correspond to the index among the M normal distributions. _ijmt ²
The logarithmic value G _ijm (x _t ) in the process of calculation is compared with the maximum value candidate for each one or a plurality of calculation intervals for calculating the term of / 2, and (3-A) the logarithmic value in the process of calculation When G _ijm (x _t ) becomes smaller than the maximum value candidate, the calculation of the logarithmic value G _ijm (x _t ) is finished, and _thereafter , the logarithmic value G for the remaining next normal distribution G
The calculation of _ijm (x _t ) is started, and (3-B) the calculation of the logarithmic value G _ijm (x _t ) is completed without the logarithmic value G _ijm (x _t ) being calculated being smaller than the maximum value candidate. Then, the maximum value candidate is rewritten to the logarithmic value G _ijm (x _t ), and after that, calculation of the logarithmic value G _ijm (x _t ) is started for the remaining next normal distribution, and (4)
When the calculation of the logarithmic value G _ijm (x _t ) is completed for all the normal distributions of the total number M, the index of the reference information storage unit is set to the index corresponding to the normal distribution that has obtained the maximum value candidate stored at this time. In addition to rewriting, the maximum value candidate is used for the logarithmic value of the output probability b _ij (x _t ) to calculate the likelihood with the hidden Markov model.

According to such an invention, the maximum logarithmic value G _ijm (x _t ) among the logarithmic values of the weighting probability g _ijm (x _t ) calculated from each normal distribution of the total number M is input. The likelihood with respect to the hidden Markov model is calculated using the logarithmic value of the output probability b _ij (x _t ) of the speech feature vector x _t . This is the weighted probability g calculated from each of the M normal distributions.
_ijm maximum weighted probability g _ijm (x _t) among (x _t), to be used for the output probability b _ij of the input speech feature vector x _t (x _t), none other.

On the other hand, the output probability b _ij (x _t ) of the speech feature vector x _t , which is typically used in the past, is
Linear sum Σ {λ _ijm b of weighted probabilities g _ijm (x _t ) = λ _ijm b _ijm (x _t ) obtained from each normal distribution of uncorrelated mixed normal distribution
_ijm (x _t )}.

By the way, since the total number M of normal distributions included in the hidden Markov model are uncorrelated with each other, the distance between the normal distribution _whose weighting probability g _ijm (x _t ) does not become maximum and the input speech feature vector x _t. Is the weighted probability g
_ijm (x _t ) is longer than the distance from the maximum normal distribution.

Because of this, the maximum weighting probability g
_{Since ijm} (x _t ) is so small as to be negligible with respect to the maximum weighting probability g _ijm (x _t ), the maximum weighting output probability g _ijm (x _t ) in the present invention is set to the input speech feature vector x.
_As the output probability b _ij (x _t ) of _t , it is possible to obtain the output probability b _ij (x _t ) that is approximately equal to the conventional one.

The logarithmic value G of the weighting probability g _ijm (x _t )
_ijm (x _t) is expressed as _{G ijm (x t) = E} ijm -D ijmt 2/2, and in the m-th normal distribution, lambda _ijm and |
Since ρ _ijm | is constant and thus E _ijm is constant,
The logarithmic value G _ijm (x _t ) in the process of calculation has a peak at E _ijm −
_D ijmt ^2/2 of the slide into reduced per one operation interval of the operation. Here as one calculation interval calculation -D _ijm ^2/2 _is, -D ijmt ²
In the calculation process of / 2, it represents the interval from the start to the end of the calculation performed on one vector component of the input speech feature vector x _t .

[0012] This is because, the operation of -D _ijm ^2/2, the processing of each one or more of the operational interval, calculates developing the logarithm G _ijm the (x _t) as compared to the maximum value candidate (above (2) ), If the logarithmic value G _ijm (x _t ) _under calculation becomes smaller than the maximum value candidate,
By ending the calculation of the logarithmic value G _ijm (x _t ) during calculation (processing of (3-A) above), the maximum logarithmic value G is obtained.
_The amount of calculation required for _ijm (x _t ) detection can be reduced.

[0013] Moreover the input speech feature vector x _t-1 of the t-1 th frame and the input speech feature vector x _t of the t-th frame, since the temporal proximity, these vectors x _t and The components of x _t-1 are likely to be similar to each other.

Therefore, the weighting probability g _ijI (x _t-1 ) obtained from the I-th normal distribution in the ( _t-1 ) th frame
When the logarithmic value G _ijI (x _t-1 ) of is the maximum logarithmic value G _ijm (x _t-1 ), the weighting probability g obtained from the I-th normal distribution also in the next t-th frame. _ijI (x _t) of the logarithmic value G _ijI (x _t) is likely to be the largest logarithmic value G _ijm (x _t).

For this reason, the logarithmic value G _ijI (x _t ) obtained from the I-th normal distribution is used as the initial value of the maximum value candidate (processing of (1) above), and the logarithmic value G _ijm (x in the process of calculation is used. _{When t} ) becomes smaller than the maximum value candidate, the logarithmic value G
(process of (3-A)) by finish calculating the _ijm (x _t) in calculating developing, it is possible to reduce the amount of calculation required to detect the largest logarithm G _ijm (x _t).

[0016]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a functional block diagram schematically showing an example of the configuration of a voice recognition apparatus suitable for implementing the voice recognition method of the present invention.

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

The dictionary unit 12 stores the hidden Markov model as a voice standard pattern. The hidden Markov model has a plurality of normal distributions that are uncorrelated with each other and has a non-correlated mixed normal distribution that represents the output probability of a speech symbol vector output from the model.

The sound processing unit 14 extracts an input 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 HMM matching unit 18 calculates the likelihood between the time series of the input voice feature vector extracted from the start frame to the end frame of the voice section and the hidden Markov model,
The category name assigned to the hidden Markov model that has been calculated using the output probabilities of the input speech feature vectors and has the maximum likelihood is used as the recognition result for the input speech signal in the relevant speech section.

Here, b _ij (x _t ): an input speech feature vector x _t extracted in the t-th frame from a hidden Markov model having an uncorrelated mixed normal distribution having a total number M of normal distributions. Output probability (1 ≦ t ≦ T. The first frame represents the start frame of the voice section and the T-th frame represents the end frame of the voice section), g _ijm (x _t ): The weighted probability of the input speech feature vector x _t calculated from the m-th (1 ≦ m ≦ M.) Normal distribution in the total number M of normal distributions (where g _ijm (x _t ) = λ _ijm b _ijm ( x _t ), b _ijm (x _t ) =
^{(2π) -p / 2 | ρ} ijm | -1/2 exp {-D ijmt 2/2},
D _ijmt ² = (x _t −μ _ijm ) ′ ρ _ijm ⁻¹ (x _t −μ
_ijm ), λ _ijm : weight of the m-th normal distribution, b _ijm (x _t ): unweighted probability of the input speech feature vector x _t calculated from the m-th normal distribution, p: input speech feature vector x order of _t , ρ _ijm : variance-covariance matrix of the m-th normal distribution, μ _ijm : mean vector of the m-th normal distribution, D _ijmt : input speech feature vector x _t and the m-th normal distribution Mahalavis general distance that represents the distance between. ), (X _t −μ _ijm ) ′: (x _t −μ _ijm ) ′ is (x _t −μ _ijm ) '
represents the transposed matrix of _{ijm), G ijm (x t} ): logarithm of weighted probabilities g _ijm (x _{t) (where, G ijm (x t) =} E ijm -D ijmt 2/2, E ijm = ln ( λ _ijm ) + ln {(2π) ^{-p / 2} _{│ρ ijm} │
^-1/2 }. ), The weighting calculated from each of the M normal distributions as the logarithmic value of the output probability b _ij (x _t ) of the input speech feature vector x _t used for the likelihood calculation with the hidden Markov model. using the probability g _ijm maximum logarithm value G _ijm among (x _t) of the logarithmic value _{G ijm (x t) (x} t).

The reference information storage unit 20 stores the maximum value candidate for detecting the maximum logarithmic value G _ijm (x _t ) in the t-th frame when t ≧ 2, and the t-th candidate when t ≧ 2. The index indicating which is the normal distribution that has the largest logarithmic value G _ijm (x _t ) in the _−1st frame is stored.

The HMM matching unit 18 then calculates the likelihood between the time series of the input speech feature vector x _t output from the start frame to the end frame and the hidden Markov model,
Do as follows.

That is, when t = 1, the logarithmic value G _ijm (x _t ) is calculated for each normal distribution with respect to the total number M of normal distributions, and the maximum logarithmic value G _ijm (x _t ) is detected. Then, the maximum logarithmic value G _ijm (x _t ) is set as the logarithmic value of the output probability b _ij (x _t ) of the input speech feature vector x _t in the first frame, and the maximum logarithmic value G _ijm (x Store the index corresponding to the normal distribution for which _t ) was obtained.

When t ≧ 2, (1) First, the logarithmic value G calculated using the normal distribution corresponding to the index
_ijm (x _t ) is stored as the maximum value candidate, and (2) -D is calculated in calculating the logarithmic value G _ijm (x _t ) using the remaining normal distribution that does not correspond to the index among the M normal distributions. _ijmt ²
The logarithmic value G _ijm (x _t ) in the process of calculation is compared with the maximum value candidate for each one or a plurality of calculation intervals for calculating the term of / 2, and (3-A) the logarithmic value in the process of calculation When G _ijm (x _t ) becomes smaller than the maximum value candidate, the calculation of the logarithmic value G _ijm (x _t ) is finished, and _thereafter , the logarithmic value G for the remaining next normal distribution G
The calculation of _ijm (x _t ) is started, and (3-B) the calculation of the logarithmic value G _ijm (x _t ) is completed without the logarithmic value G _ijm (x _t ) being calculated being smaller than the maximum value candidate. Then, the maximum value candidate is rewritten to the logarithmic value G _ijm (x _t ), and after that, calculation of the logarithmic value G _ijm (x _t ) is started for the remaining next normal distribution, and (4)
When the calculation of the logarithmic value G _ijm (x _t ) is completed for all the normal distributions of the total number M, the index of the reference information storage unit 20 is added to the index corresponding to the normal distribution that has obtained the maximum value candidate stored at this time. And the maximum value candidate is used as the logarithmic value of the output probability b _ij (x _t ) to calculate the likelihood with the hidden Markov model.

FIG. 2 is a diagram for explaining the hidden Markov model used for the voice standard pattern. The Hidden Markov Model (hereinafter, HMM) used for the voice standard pattern is a voice signal for one unit of voice recognition, here, a voice signal for one word, and represents a voice signal to which a category z is added. A plurality of HMMs are prepared individually for each category, and the HMMs and categories z are stored in the dictionary unit 12 in association with each other.

The HMM has a set 1 of states consisting of a total of I states S _{1 to} S _I, a set 2 of speech symbol vectors x, a set 3 of state transition probabilities a _ij , and an output probability b _ij (x )
A set 4, a set 5 of the initial state probability .PHI _i, final state F
And the set 6 of

[0028]

[Equation 1]

For example, in the example of FIG. 2, a ₁₂ is the state S ₁
The probability and b ₁₂ (x) is the voice symbol vector x when a transition from the state S ₁ to state S ₂ transitions to a state S ₂ is output from, also a ₂₂ transitions from state S ₂ to state S ₂ probability and b ₂₂ (x) represents the probability that speech symbol 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 of different generation methods are collected.
Sets 1 to 6 expressing the statistical properties of these audio signals are obtained.

The output probabilities b _ij (x) are represented by an uncorrelated mixed normal distribution (uncorrelated continuous probability density distribution) consisting of a plurality of normal distributions that are uncorrelated with each other and are functions of the speech symbol vector x. It The decorrelated mixed normal distribution has the advantage of being easy to handle mathematically and having high expressiveness.

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

The sound processing section 14 extracts the input voice feature vector x _t for each frame from the input voice signal. The frame number t given to the input speech feature vector x _t at this point is a number sequentially given with the frame at the start of the acoustic processing as the t = 1th frame, and this frame number t will be described later. In the HMM matching unit 18, the start frame of the voice section is rewritten as the first (t = 1) frame into numbers sequentially assigned from the start frame to the end frame of the voice section.

The input speech feature vector x _t is x _t = (x
_t 1, x _t 2, ..., X _t p). p represents the order of the input speech feature vector x _t , and x _{t1 to} x _t p represent vector components of the input speech feature vector x _t .

As the vector component of the input speech feature vector x _t , for example, one obtained from each filter output when an input speech signal is input to a bandpass filter group consisting of a plurality of bandpass filters having different center frequencies, or an input A power spectrum component obtained by Fourier analysis of an audio signal, or a linear predictive analysis or LPC of an input audio signal
The LPC cepstrum coefficient determined by analysis can be used. Here, an example in which the input speech feature vector x _t is extracted using a band filter group will be described.

The acoustic processing unit 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, the components of the input speech feature vector x _t x _t 1
Extract as Similarly, the remaining signal components x2 to x
From p, it extracts the component x _t 2~x _t p of the input speech feature vector x _t.

Next, the voice section detecting section 16 includes the sound processing section 1
Based on the input 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 including 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.

The HMM collation unit 18 inputs the section information and the input voice feature vector x _t from the voice section detection unit 16 and extracts the input voice feature vector x _t from the start frame to the end frame of the voice section. Time series x _{1 to} x _T
Generate Here, the frame number t is rewritten into a number sequentially assigned from the start frame to the end frame of the voice section with the start frame of the voice section as the first (t = 1) frame.

Then, the HMM matching unit 18 finds the likelihood ln {P (x _{1 to} x _T )} between the vector time series x _{1 to} x _T and the HMM stored in the dictionary unit 12, and determines the maximum likelihood. The category z assigned to the obtained HMM is output as the recognition result.

Here, P (x _{1 to} x _T ) is the probability that the vector time series x _{1 to} x _T will appear in the HMM, and
It is expressed as the following equation (1).

[0041]

[Equation 2]

In the equation (1), * i is i (the number i given to the state S _i belonging to the final state F) that satisfies S _i εF, and therefore the forward probability c that i = * i. _iT
Among them, the maximum forward probability c _iT is _defined as the appearance probability P (x ₁ ~
x _T ).

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

[0044]

(Equation 3)

Here, ln (a _ij ) = A _ij , ln {b
_ij (x _t )} = B _ij (x _t ), ln (c _it ) = C _it (hereinafter, transition logarithmic value A _ij , output logarithmic value B _ij , forward logarithmic value C
_iT ), and the equations (1) to (3) are modified to obtain the likelihood ln
Expressions (4) to (6) regarding the calculation of {P (x _{1 to} x _t )} are obtained.

[0046]

(Equation 4)

Since the equations (5) to (6) are recurrence equations of t, the forward logarithmic values C _iT when t = 1, 2, ..., T are calculated sequentially as the following equation. it can.

[0048]

(Equation 5)

In the HMM, there is one or a plurality of transition paths from the initial state to generate the vector series x _{1 to} x _t and reach the state S _i . In most cases, there are a plurality of transition paths. When there are a plurality of transition paths, the forward logarithmic value C _iT is obtained for each transition path, and therefore a plurality of forward logarithmic values C _iT corresponding to the respective transition paths are obtained.

The HMM matching unit 18 obtains the forward logarithmic value C _iT in the HMM assigned with the category z, and i = *
The maximum forward logarithmic value C _iT among the forward logarithmic values C _iT for i is obtained as the likelihood ln {P (x _{1 to} x _T )} between the vector time series x _{1 to} x _T and the HMM. . Then, for all HMMs stored in the dictionary unit 12,
For each MM, the likelihood ln {P (x _{1 to} x _T )} is calculated, and the category z assigned to the HMM that has the maximum likelihood ln {P (x _{1 to} x _T )} is calculated at vector time. The sequence x _{1 to} x _T is output as the recognition result of the input speech signal.

In the process of calculating the above-mentioned likelihood ln {P (x _{1 to} x _T )}, the most complicated operation is the output logarithmic value B _ij.
This is an operation for obtaining (x _t ). In order to perform this calculation at high speed, the output probability b _ij (x _t ) is defined by the following equation (12).
The output probability b _ij (x _t ) is the probability that the input speech feature vector x _t will be output from the hidden Markov model having a decorrelation mixed normal distribution having a total number M of normal distributions.

[0052]

(Equation 6)

G _ijm (x _t ) in the equation (12) is the input speech feature vector x _t calculated from the m-th normal distribution in the uncorrelated mixed normal distribution consisting of a total of M normal distributions.
Is a weighted probability of and can be expressed using the following equations (13) to (15).

[0054]

(Equation 7)

In equation (13), λ _ijm is the weight of the m-th normal distribution, and b _ijm (x _t ) is the unweighted probability of the input speech feature vector x _t calculated from the m-th normal distribution. is there. The unweighted probability b _ijm (x _t ) is expressed by Expression (14), and p in Expression (14) is the order of the input speech feature vector x _t , ρ
_ijm is the variance and covariance matrix of the mth normal distribution, and D
_ijmt is a Mahalabis general distance that represents the distance between the input speech feature vector x _t and the m-th normal distribution. The Mahalavis general distance D _ijmt is expressed by Equation (15), and Equation (15)
_Where μ _ijm is the mean vector of the m-th normal distribution, (x
_t −μ _ijm ) ′ is the transposed matrix of (x _t −μ _ijm ).

[0056] (12) is weighted probability g _ijm (x having the maximum of the weighted probability g _ijm which the uncorrelated Gaussian Mixture consisting of the total number of M normal distributions obtained from the individual normal distribution (x _t) the _t), representative of the detected as the output probability b _ij of the input speech feature vector x _t (x _t).

The conventional typical output probability b _ij (x _t ) is expressed as a linear sum of the weighting probabilities g _ijm (x _t ).
Even if the maximum weighted probability g _ijm (x _t ) is used as the output probability b _ij (x _t ) as in the equation (12), the conventional output probability b _ij (x _t )
Output probabilities b _ij (x _t ) approximately equal to can be obtained. In the uncorrelated mixed normal distribution, the total number M of normal distributions are uncorrelated with each other, so that the weighting probability g _ijm (x _t ) that is not the maximum is _smaller than the maximum weighting probability g _ijm (x _t ). This is because it is considered to be a very small value.

[0058] and the logarithm B _ij of the output probability _{b ij (x t) (x} t)
(Hereinafter, the output logarithmic value B _ij (x _t )) can be expressed by the following expression (16) using the expression (12).

[0059]

(Equation 8)

Weighted logarithmic value G _ijm (x _t ) in the equation (16)
Is a logarithmic value of the weighted probability g _ijm (x _t ), and
It can be expressed by the following equation (17) using 3) to (15).

[0061]

[Equation 9]

_Attention is now paid to the weighted logarithmic value G _ijm (x _t ). Since all M normal distributions that form the HMM uncorrelated mixed normal distribution are uncorrelated, the variance / covariance matrix ρ _ijm of each normal distribution is a diagonal matrix.

The element of the r-th row and the s-th column of the variance- _sub- dispersion matrix ρ _ijm (m-th ρ _ijm of the uncorrelated mixed normal distribution) is A
_ijm rs, _represent the r-th component of the input speech feature vector x _t as Br, and the r-th component of the mean vector μ _ijm (m-th μ _ijm of the uncorrelated mixed normal distribution) as C _ijm r For example, the equation (15) can be transformed into the following equation (18).

[0064]

(Equation 10)

Since the variance / _sub- dispersion matrix ρ _ijm is a diagonal matrix, A _ijm rs = 0 when r ≠ s. Therefore, the equation (18) can be transformed into the following equation (19).

[0066]

[Equation 11]

Moreover, since the variance / _sub- dispersion matrix ρ _ijm is an inverse correlation matrix, A _ijm rr ≧ 0 holds. Therefore, (19)
_{Each term} of A _ijm rr · (Br −C _ijm r) ^{2 in} the _equation is non-negative and therefore D _ijmt ² ≧ 0.

Therefore, in the equation (17), E _ijm is a constant value determined for each normal distribution and D _ijmt ² ≧ 0. Therefore, the weighted logarithmic value G _ijm (x _t ) in the calculation is E
_{From ijm} , A _ijm rr · (Br −C _ijm r) ^{2 in the} equation (19)
The value is obtained by sequentially subtracting each term of. In other words, the value of G _ijm (x _t ) in the process of calculation is A _ijm rr ·, which is performed for one component of the input speech feature vector x _t , with E _ijm as the peak.
The value of (Br-C _ijm r) ² decreases with each calculation interval.

Next, the flow of likelihood calculation performed by the HMM matching unit 18 will be described with reference to FIGS. 3 and 4. FIG.
Is an operation flow for calculating the maximum weighted logarithmic value G _ijm (x _t ) when t = 1, and FIG. 4 is an operation for calculating the maximum weighted logarithmic value G _ijm (x _t ) when t ≧ 2. The flow is shown.

First, the HMM matching unit 18 sets an initial value C _i0 of the forward logarithmic value. Next, the HMM matching unit 18 sets t =
The forward logarithmic value C _it when _it is 1, that is, the forward logarithmic value C _it for the input speech feature vector x _t of the start frame (first frame) is obtained.

Therefore, the HMM matching unit 18 searches the HMM for a voice symbol vector corresponding to the input voice feature vector x _t . The uncorrelated mixed normal distribution representing the output probability of the corresponding speech symbol vector is used as the uncorrelated mixed normal distribution representing the output probability b _ij (x _t ) of the input speech feature vector x _t. calculating a weighted logarithmic value G _ijm (x _t) from each normal distribution, and detects the maximum weighted logarithmic value G _ijm (x _t) (S1 in FIG. 3), and the maximum of the weighted logarithmic value G _ijm (x _t ) _Is stored as the output logarithmic value B _ij (x _t ) of the input speech feature vector x _t , and the number m of the normal distribution that obtains the maximum weighted logarithmic value G _ijm (x _t ) is stored as the index Q _ij . (S2 in FIG. 3). When there are multiple state transitions that output the corresponding speech symbol vector, there is an uncorrelated mixed normal distribution that represents the output probability of the speech symbol vector for each state transition. The output logarithmic value B _ij (x _t ) and the index Q _ij are individually obtained and stored for each state transition by using the uncorrelated mixed normal distribution of the input speech feature vector x _t .

Next, the HMM matching unit 18 calculates the forward logarithm value C _it when t = 1 by using the calculated output logarithm value B _ij (x _t ).

Next, the HMM matching unit 18 calculates the forward logarithmic value C _it of the input speech feature vector x _t when t ≧ 2.

Therefore, the HMM collation unit 18 searches the HMM for a voice symbol vector corresponding to the input voice feature vector x _t . The uncorrelated mixed normal distribution representing the output probability of the corresponding speech symbol vector is used as the uncorrelated mixed normal distribution representing the output probability b _ij (x _t ) of the input speech feature vector x _t. The Q _ij -th normal distribution corresponding to the index Q _ij is searched from among the normal distributions, and the weighted logarithmic value G _ijm (x _t ) is calculated from this normal distribution. Then, the calculated weighted logarithmic value G _ijm (x _t ) is stored as the maximum value candidate G _ijQ (x _t ), and then the number m of the normal distribution is initialized to m = 1 (S in FIG. 4).
1).

Next, the number m of the normal distribution is the index Q.
_It is determined whether it is equal to _ij (S2 in FIG. 4).

If the number m is not the index Q _ij in S2 of FIG. 4, the calculation of the weighted logarithmic value G _ijm (x _t ) is started using the m-th normal distribution (S3 of FIG. 4). First, the term of E _ijm of G _ijm (x _t ) is calculated (S4 in FIG. 4), and then the operation of the term of D _ijmt ² of G _ijm (x _t ) is performed by one operation interval or multiple operation intervals. , (S5 in FIG. 4). One calculation interval is a calculation interval performed for one component of the input speech feature vector x _t . Next, it is determined whether or not the weighted logarithmic value G _ijm (x _t ) being calculated is larger than the maximum value candidate G _ijQ (x _t ) (S6 in FIG. 4).

If G _ijm (x _t )> G _ijQ (x _t ) in S6 of FIG. 4, it is determined whether the calculation of D _ijmt ² has been completed for all the components of the input speech feature vector x _t. (S in FIG. 4
7), _if the calculation of D _ijmt ² is not completed, the process returns to S5. When the operation of D _ijmt ² is completed, the weighted logarithmic value G _ijm (x _t ) for which the operation is completed is set to the maximum value candidate G.
_ijQ (x _t ) is rewritten, and at the same time, the number m of the normal distribution that has obtained the weighted logarithmic value G _ijm (x _t ) for which the calculation has been completed is rewritten as an index Q _ij (S8). After that, it is determined whether or not the processing has been completed for all of the M normal distributions (S9), and if not completed, 1 is added to the normal distribution number m (S11). Returning to the processing of S2, if completed, the maximum value candidate G _ijQ (x _t ) stored at this time is stored as the output logarithmic value B _ij (x _t ) of the input voice signal x _t (S10). In step S6 of FIG. 4, G _ijm (x _t ) ≦ G
If it is _ijQ (x _t ), S7 to S8 are not processed and S
9 is performed.

If m = Q _ij in S2, S3 to S8
The process of S9 is performed without performing the process of.

When there are a plurality of state transitions that output the corresponding speech symbol vector, since there is a decorrelation mixture normal distribution that represents the output probability of the speech symbol vector for each state transition, each of these decorrelation mixture The normal distribution is used for the uncorrelated mixed normal distribution of the input speech feature vector x _t , and the processes of S1 to S11 of FIG. 4 are individually performed for each state transition.

Each input speech feature vector x _{t of} t = 2 to _T
Each time the output logarithmic value B _ij (x _t ) is obtained, the forward logarithmic value C
_It is obtained, and the finally obtained forward logarithmic value C _iT is obtained as the likelihood between the input speech feature vectors x _{1 to} x _T and the HMM.

As described above, G in the process of calculation
The value of _ijm (x _t ) is A _ijm rr ·, which is performed for one component of the input speech feature vector x _t , with E _ijm as a peak.
Since the calculation of (Br-C _ijm r) ² decreases at each calculation interval, G _ijm (x _t )> G in the determination of S6 in FIG.
when the _ijQ (x _t), by ending the calculation of the calculation course of G _ijm (x _t), can be improved calculation speed by omitting wasteful operations.

Further, the number m of the normal distribution which has obtained the maximum weighted logarithmic value G _ijm (x _t ) in the immediately preceding frame, that is, the index Q _ij is stored, and is _calculated from the normal distribution corresponding to the index Q _ij in the next frame. Weighted logarithmic value G
_{By using ijm} (x _t ) as the maximum value candidate, unnecessary calculation can be omitted and the calculation speed can be improved. This is because the input speech feature vector x _t is similar between the previous frame and the next frame, and therefore the index Q _ij is also used in the next frame.
Weighted logarithmic value G obtained from the normal distribution corresponding to
_This is because there is a high possibility that _ijm (x _t ) will be the maximum.

[0083]

As is clear from the above description, according to the speech recognition method of the present invention, since the total number M of normal distributions included in the hidden Markov model are uncorrelated with each other, the weighting probability g _ijm ( The distance between the normal distribution in which x _t ) is not the maximum and the input speech feature vector x _t is longer than the distance with the normal distribution in which the weighting probability g _ijm (x _t ) is the maximum. Because of this, the maximum weighting probability g
_{Since ijm} (x _t ) is so small as to be negligible with respect to the maximum weighting probability g _ijm (x _t ), the maximum weighting output probability g _ijm (x _t ) in the present invention is set to the input speech feature vector x.
_As the output probability b _ij (x _t ) of _t , it is possible to obtain the output probability b _ij (x _t ) that is approximately equal to the conventional one.

The logarithmic value G of the weighting probability g _ijm (x _t )
_ijm (x _t) is expressed as _{G ijm (x t) = E} ijm -D ijmt 2/2, and in the m-th normal distribution, since E _ijm is constant, calculating developing logarithmic values G _ijm ( x _t) is slide into reduced per one operation interval of the operation of _-D ijmt ^2/2 the E _ijm a peak.

[0085] This is because, -D _ijm ^2/2 operations for each one or a plurality of calculation intervals, calculates developing logarithm G _ijm the (x _t) is compared with the maximum value candidate calculation developing logarithmic values G _{When ijm} (x _t ) becomes smaller than the maximum value candidate, the relevant logarithmic value G _ijm (x _t )
The calculation amount required for detecting the maximum logarithmic value G _ijm (x _t ) can be reduced by terminating the calculation of the calculation.

[0086] Moreover, the input speech feature vector of the immediately preceding frame x _t-1 and the input speech feature of the next frame vector x _t
And are close in time, these vectors x _t
And the components of x _t-1 are likely to be similar to each other. Therefore, the logarithmic value G _ijI (x of the weighting probability g _ijI (x _t-1 ) obtained from the I-th normal distribution in the t-1 th frame
_t-1 ) becomes the maximum logarithmic value G _ijm (x _t-1 ), the pair of weighting probabilities g _ijI (x _t ) obtained from the I-th normal distribution is calculated in the next t-th frame. It is highly possible that the numerical value G _ijI (x _t ) becomes the maximum logarithmic value G _ijm (x _t ).

For this reason, the logarithmic value G _ijI (x _t ) obtained from the I-th normal distribution is set as the initial value of the maximum value candidate.
When calculating the course of the logarithmic value G _ijm (x _t) is smaller than the maximum value candidate, by ending the calculation of the logarithmic value G _ijm (x _t) in calculating developing, the largest logarithm G _ijm (x _t ) The amount of calculation required for detection can be reduced.

[Brief description of drawings]

FIG. 1 is a diagram schematically showing a configuration of a voice recognition device suitable for implementing the present invention.

FIG. 2 is a diagram for explaining an HMM.

FIG. 3 is an operation flow for calculating a weighted logarithmic value G _ijm (x _t ) when t = 1.

FIG. 4 is an operation flow for calculating a weighted logarithmic value G _ijm (x _t ) when t ≧ 2.

[Explanation of symbols]

 10: Speech recognition device 12: Dictionary section 14: Sound processing section 16: Speech section detection section 18: HMM collation section 20: Reference information storage section

Claims (1)

[Claims] 1. A hidden Markov model is used as a voice standard pattern, and the hidden Markov model has a plurality of normal distributions that are uncorrelated with each other and represents a non-correlated mixed normal representing an output probability of a voice symbol vector output from the model. With distribution,
The likelihood between the time series of the input speech feature vector extracted from the start frame to the end frame in the speech section and the hidden Markov model is calculated using the logarithmic value of the output probability of each input speech feature vector. , _Ij (x _t ): the total number of M normal _patterns in the speech recognition method in which the category name given to the hidden Markov model that has the maximum likelihood is used as the recognition result for the input speech signal in the speech section. The output probability (1 ≦ t ≦ T. The output probability of the input speech feature vector x _t extracted at the t-th frame is output from the hidden Markov model having the uncorrelated mixed normal distribution having the distribution. The beginning frame of the voice section and the T-th frame represent the end frame of the voice section.), G _ijm (x _t ): The m-th (1 ≦ m ≦ M.) In the total number M of normal distributions. Regular minutes of Weighting the probability of the input speech feature vector x _t calculated from _{(where, g ijm (x t) =} λ ijm b ijm (x t), b ijm (x t) = (2π) -p / 2 | ρ ijm | ^-1/2 exp {-D
^{_{ijmt 2/2}, D ijmt}} 2 = (x t -μ ijm) 'ρ ijm -1 (x t -μ
_ijm ), λ _ijm : weight of the m-th normal distribution, b _ijm (x _t ): unweighted probability of the input speech feature vector x _t calculated from the m-th normal distribution, p: input speech feature vector x order of _t , ρ _ijm : variance-covariance matrix of the m-th normal distribution, μ _ijm : mean vector of the m-th normal distribution, D _ijmt : input speech feature vector x _t and the m-th normal distribution Mahalavis general distance that represents the distance between. _{), G ijm (x t)} : logarithm of weighted probabilities g _ijm (x _{t) (where, G ijm (x t) =} E ijm -D ijmt 2/2, E ijm = ln (λ ijm) + ln {( 2π) ^{-p / 2} ｜ ρ _ijm ｜
^-1/2 }. ), Input the maximum logarithmic value G _ijm (x _t ) among the logarithmic values G _ijm (x _t ) of the weighted probabilities g _ijm (x _t ) calculated from each of the M normal distributions. In calculating the likelihood with the hidden Markov model by using the logarithmic value of the output probability b _ij (x _t ) of the speech feature vector x _t , when t ≧ 2, the maximum value in the t-th frame is calculated. The maximum value candidate for detecting the logarithmic value G _ijm (x _t ) and the maximum logarithmic value G in the (t−1) th frame when t ≧ 2.
_ijm (x _t ) is provided with a reference information storage unit that stores an index that indicates which normal distribution is obtained, and at t = 1, for all normal distributions of the total number M, logarithmic values for each normal distribution calculates the G _ijm (x _t), detects the maximum logarithmic value G _ijm (x _t), outermost sized logarithm G _ijm the (x _t) of the input speech feature vector x _t in the first frame The index corresponding to the normal distribution that obtains the _maximum logarithmic value G _ijm (x _t ) is stored as the logarithmic value of the output probability b _ij (x _t ), and when t ≧ 2, (1) corresponds to the index first. The logarithmic value G _ijm (x _t ) calculated using the normal distribution is stored as the maximum value candidate, and (2) the logarithmic value G using the remaining normal distribution that does not correspond to the index among the total number M of normal distributions.
In the calculation of _ijm (x _t), for each one or a plurality of calculation intervals of the operation for calculating the term of _-D ijmt ^2/2, calculate developing logarithm G _ijm the (x _t), the maximum value candidate In comparison, (3-A) When the logarithmic value G _ijm (x _t ) in the process of calculation becomes smaller than the maximum value candidate, the calculation of the logarithmic value G _ijm (x _t ) ends, and then the remaining next value. It starts the calculation of the normal distribution per logarithmic value _{G ijm (x t), (} 3-B) calculating developing logarithmic values G _ijm (x _t) without becomes smaller than the maximum value candidate, the logarithmic value G _ijm ( x _t ), the maximum value candidate is set to the logarithmic value G _ijm.
Rewriting to (x _t ), after that, calculation of the logarithmic value G _ijm (x _t ) is started for the remaining next normal distribution, and (4) logarithmic value G _ijm (x _When the calculation of ( _t ) is completed, the index of the reference information storage unit is rewritten to the index corresponding to the normal distribution for which the maximum value candidate stored at this time is obtained, and the maximum value candidate is
A speech recognition method characterized by calculating a likelihood with a hidden Markov model using a logarithmic value of an output probability b _ij (x _t ).