JPH07160288A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To improve recognizing performance by making a code vector suitable to an input voice. CONSTITUTION:This device is provided with a code book storage means 406, a fuzzy vector quantization means 405 which converts each vector of a feature vector group to a group (vector of attribution factor) corresponding to each label by a code book and converts the feature vector group to an attribution factor vector group, a HMM storage means 407 which stores HMM in which occurrence probability of a label defined is defined for each state, a feature vector group occurrence rate calculating means which calculates occurrence rate from the HMM of the feature vector group by label occurrence probability and an attribution factor vector, and a code book correcting means 408 which corrects each code vector.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention relates to pattern recognition, and in particular to a time series standard model using Hidden Markov Model (HMM: Hidd).
An apparatus for correcting a finite number of representative points in a vector space (hereinafter, simply referred to as a code vector) and an apparatus for normalizing an input feature vector in speech recognition using en Markov Model).

The present invention is applicable to general time-series signals, but for convenience of explanation, the prior art and the present invention will be described below by taking speech recognition as an example.

[0003]

BACKGROUND OF THE INVENTION In general, speech recognizers are configured to convert an unknown speech signal into a sequence of defined feature vectors and compare the feature vectors with a previously stored identified reference model. ing. As a result of the comparison, the speech signal is identified as the best matching (maximum likelihood) reference model according to defined recognition criteria. At present, the reference model that is considered to have the best performance is an HMM that uses a set of states and state transitions based on statistical estimation.
To use.

In HMM-based voice recognition, the following operations are performed in order to calculate the likelihood. First, unknown speech signals are subjected to linear predictive coding (LPC).
The feature vector is converted into a series of feature vectors (feature vector series) by using a well-known method such as oding analysis, and the feature vector is then labeled (labeled to the code vector indicating the code vector with the shortest inter-vector distance). Number or symbol), the feature vector series becomes a label series (label series), and the HMM is created and stored in advance.
Calculates the probability (likelihood) of generating the label sequence, and sets the HMM with the maximum likelihood as the recognition result.

Here, the code vector is a vector representing a finite number of representative points created in advance by a well-known method such as the LBG algorithm in a multidimensional space of feature vectors, and can be searched by a label. Remembered

Next, the HMM will be described. HMM
Is used to evaluate an observation label sequence O = o ₁ , o ₂ , ..., O _T such that each observation is one of a finite number of M labels. FIG. 1 is a diagram for explaining such an HMM.

In FIG. 1, the number of states N = 3 and the finite number of labels M = 4 are taken as an example. The transition between states 1, 2, and 3 is represented as a state transition probability matrix A = [a _ij ], and the state transition probability a _ij is the probability that a transition will occur next in state j when in state i. The label occurrence probability from the HMM is represented by a label occurrence probability matrix B = [b _ij (k)], and the label occurrence probability b _ij (k) generates a label k when the state i transits to the state j. It is a probability. One HMM is created for each vocabulary and used to classify the label series based on the probability (likelihood) of each HMM generating a label series.

Feature vector sequence Y = y of unknown voice signal
₁ , y ₂ , ..., Y _t , ..., y _T are label sequences obtained for O = o ₁ , o ₂ , ..., O _T , and an arbitrary state sequence of length T capable of generating HMMλ is S = When s ₁ , s ₂ , ..., S _T , the probability (likelihood) that the HMMλ generates the label sequence O
Is expressed as in (Equation 1).

[0009]

[Equation 1]

The above is the feature vector y_tThe only label
o_tThis is an explanation of the case of converting to
Chi feature vector y_tIn the vicinity of K (distance between vectors)
A set of labels (limited to K from the closest one)_t1，
o_t2,, o_tKAnd the feature vector y _tIn the label set of
Feature vector for each code vector searched
y_tThe degree of membership u_t1, U_t2, ..., u_tKUsing the
Le vector o_t= (O_t1, O_t2,, o_tK) And the degree of membership
Cutle u_t= (U_t1, U_t2, ..., u_tK) To convert
There is a method called Z vector quantization. The fuzzy
The vector quantization is performed by labeling the feature vector sequence Y
Le series O = o₁, O₂,, o_TAnd membership vector series U
= U₁, U₂, ..., u_THMMλ is converted to
The probability (likelihood) of generating a Kuttle sequence is the label occurrence probability.
b_imInstead of, the degree of occurrence of the feature vector ω _iusing (t)
Is shown as (Equation 2).

[0011]

[Equation 2]

Here, the possible value of K is an integer from 1 to M, and is limited to a small value K (K = 3 is appropriate) in order to reduce the calculation amount. The case of 1 corresponds to the case where the unique label is generated.

Recognition is performed by comparing the likelihoods (probabilities that HMMs of each vocabulary occur in the feature vector series Y) of the HMMs obtained as described above and selecting the one having the maximum likelihood. .

For example, if the number of vocabularies is ^W , and the likelihood of HMMλ ^w of the w-th vocabulary with respect to the unknown feature vector sequence Y is L (Y | λ ^w ), the recognition result w is (Equation 3).
Is.

[0015]

[Equation 3]

Therefore, the recognition result is obtained by the relative comparison of the obtained likelihoods. FIG. 2 shows an example of the structure of the codebook. The codebook is constructed in such a form that a label is stored in the first column of each row and the value of the codevector is stored thereafter, that is, the codevector can be searched by the label. ing.

A block diagram of the speech recognition apparatus having the above-described structure is shown in FIG. A feature extraction unit 301 converts an unknown speech signal into a feature vector at regular time intervals using a well-known method such as LPC analysis, and outputs a sequence of feature vectors Y = y ₁ , y ₂ , ..., Y. _{Get t} , ..., y _T.
Here, T is the length of the feature vector sequence Y with respect to the unknown voice signal.

A codebook storage unit 302 stores a code vector in a searchable form by a label attached to the code vector.

Reference numeral 303 is a fuzzy vector quantizer.
The feature vector extracted by the feature extraction unit 301.
Ry_tAnd stored in the codebook storage unit 302
K in order of the distance between the vector and the code vector
Labels and feature vector y _tSearch with the label set of
Feature vector y for each code vector_t
Label vector o_t= (O_t1, O
_t2,, o_tK) And the membership vector u_t= (U_t1, U_t2，
…, U_tK), And the feature vector y_tThe series Y of
Label vector series O = o₁, O₂,, o_TAnd degree of membership
Kuturu series U = u₁, U ₂, ..., u_TTo convert to
It

An HMM storage unit 304 stores the state transition probability matrix A and the label occurrence probability matrix B for each vocabulary for recognizing already created HMMλ ^w (w = 1 to W). Just remember. Therefore, the wth HMM
Is expressed as λ ^w = {A ^w , B ^w } w = 1 to W.

Reference numeral 305 is a feature vector sequence occurrence degree calculation unit.
And is obtained by the fuzzy vector quantizer 303.
The label vector series O and the degree of membership vector series
U and the w-th stored in the HMM storage unit 304
Vocabulary label occurrence probability matrix B ^wUsing HMMλ^wAgainst
Occurrence matrix of feature vector sequence Ω^w= {Ω^w _it}
It is calculated according to (Equation 4) or (Equation 5).

[0022]

[Equation 4]

[0023]

[Equation 5]

Here, the label occurrence probability b _i o _tk is the label occurrence probability that the k-th label o _tk when the feature vector y _t at time t is fuzzy vector quantized occurs from the state i of the HMM.

A likelihood calculation unit 306 calculates the HMMλ ^w calculated by the feature vector sequence generation degree calculation unit 305.
Using a state transition probability matrix A ^w of the feature vector series occurrence rate matrix Omega ^w wherein stored in the HMM storage unit 304 and HMMramuda ^w relative likelihood L | and calculates the (Y λ ^w).

A likelihood storage unit 307 stores the likelihood L (Y | λ ^w ) of each word HMMλ ^w for the feature vector series Y calculated by the likelihood calculation unit 306 for comparison.

A comparison / determination unit 308 determines, as a recognition result, a vocabulary corresponding to the HMM that gives the maximum value among the likelihoods of the HMMs stored in the likelihood storage unit 307.

305 to 307 are HMMλ of each vocabulary
^{It is} performed once for ^w and repeated from w = 1 to W, and the result is evaluated by the comparison and determination unit 308.

The above is the configuration of the speech recognition apparatus using the conventional codebook and HMM. In order to perform the voice recognition using the HMM as described above, an HMM creating procedure using a large amount of HMM creating voice data (hereinafter referred to as HMM learning data), called HMM learning, is required. .

When the HMM learning data is input by the same speaker even when it is recognized by one specific speaker, it is called specific speaker voice recognition, and HMM learning is performed by using an unspecified number of speakers. If the speaker at the time of recognition is unknown, it is called unspecified speaker voice recognition.

[0031]

In the case of specific speaker voice recognition, performance is good, but enormous amount of HMM learning data needs to be collected from one speaker, which is not practical.

In the case of unspecified speaker speech recognition, it is easy to collect the learning data of the HMM, and the more data, the higher the statistical reliability of the HMM. However, the performance of the specific speaker is not exceeded, and in the HMM, there is a specific speaker whose recognition performance is extremely poor.

Further, when the HMM is stored in units smaller than words such as syllables and phonemes, there is a difference in context (order of syllables, phonemes, etc.) between the learning vocabulary and the recognition vocabulary. Poor performance occurs due to the effect.

Also, when the surrounding environment at the time of recognition is different from the case where the learning data is recorded, the difference causes deterioration in performance.

As described above, in the conventional speech recognition, there is a problem in that the performance is degraded due to the difference between the speakers at the time of learning and the recognition and the difference in the context.

[0036]

[Means for Solving the Problems]

(1) In order to solve this problem, the present invention provides a finite number of representatives of the feature vector space when the utterance content of the correction voice is known to the system side (what kind of word is uttered is known). Codebook storage means for storing points (code vectors) in a searchable form by labels attached to the points, and a set of membership degrees (membership vectors) for each vector of the feature vector series by the codebook Fuzzy vector quantization means for converting the feature vector sequence into a membership vector sequence, HMM storage means for storing an HMM in which the label occurrence probability (label occurrence probability) is defined for each state, Feature vector sequence occurrence degree calculation means for calculating the occurrence degree of the feature vector series from the HMM based on the label occurrence probability and the belonging degree vector. Codebook modifying means for modifying each of the codevectors, the codebook modifying means including modification vector calculating means for modifying the codevectors so as to maximize the degree of occurrence of the feature vector sequence from the HMM, It is configured to modify the code vector. (2) In order to solve this problem, the present invention provides a finite number of representatives of the feature vector space when the utterance content of the correction voice is known to the system side (which word is known). Codebook storage means for storing points (code vectors) in a searchable form by labels attached to the points, and a set of membership degrees (membership vectors) for each vector of the feature vector series by the codebook Fuzzy vector quantization means for converting the feature vector sequence into a membership vector sequence, HMM storage means for storing an HMM in which the label occurrence probability (label occurrence probability) is defined for each state, Feature vector sequence occurrence degree calculation means for calculating the occurrence degree of the feature vector series from the HMM based on the label occurrence probability and the belonging degree vector. A correction vector calculating means for correcting the feature vector, the feature vector correcting means including a correction vector calculating means for correcting the feature vector so as to maximize the degree of occurrence of the feature vector series from the HMM, It is configured to modify the feature vector.

[0037]

[Action]

(1) The effect of the present invention is that, when the utterance content of the correction voice is known to the system side (which word is uttered is known), the likelihood of the HMM corresponding to the voice is maximized. Thus, the correction vector of the code vector is calculated, and the code vector is corrected, thereby creating a code vector suitable for the correction voice and improving the recognition performance. . (2) The effect of the present invention is that, when the utterance content of the correction voice is known to the system side (which word is uttered is known), the likelihood of the HMM corresponding to the voice is maximized. Thus, the correction vector of the feature vector is calculated, and at the time of recognition, the correction vector is used to normalize the feature vector sequence of the unknown input voice, thereby improving the recognition performance.

[0038]

BEST MODE FOR CARRYING OUT THE INVENTION The present invention can be applied to the correction of the difference in condition between the time of learning and the time of recognition of HMM.
When the speaker who uttered the MM learning data is different from the speaker who uses the recognition device, the correction of the difference between the speakers is taken as an example.
The present invention will be described.

The limit range for calculating the sum of products or the exponentiation product of the feature vector sequence generation degree calculation formula can be an integer value from 1 to a value M equal to the codebook size, but it is small to reduce the calculation amount. In many cases, a value is used, and in this embodiment, the limit range is represented by the letter K.

FIG. 4 shows a codebook correction device according to the present invention.
It is a block diagram showing the outline of an arrangement. 401 is a correction voice
It's a memory section, and I want to modify the codebook for that person
Person (hereinafter referred to as a modified speaker), that is, a voice recognition system
Known content spoken by the speaker using
Voice content of the device is known in advance)
^r(Hereinafter, referred to as correction voice) is the number of utterances R (r = 1 to 1)
R) I remember it and use it to modify the codebook after that.
Be done. If the utterance content is known, what is the correction voice?
You can use words and sentences.

Reference numeral 402 denotes a feature extraction unit, which uses a technique similar to the feature extraction technique used in the recognition apparatus shown in FIG. 3 and outputs the correction speech S ^r at a fixed time interval Y ^r = y. _{^{1 r, y 2 r, ...}} , y t r, ..., it is converted to y _T ^r. Here, T ^r is the number of data frames when the correction voice S ^r is converted into a feature vector sequence.

A correction feature vector storage unit 403 stores a feature vector sequence Y ^r obtained by feature extraction of the correction voice signal S ^r stored in the 401 by the feature extraction unit 402 for r = 1 to R. is doing.

A data control unit 404 controls the following processes by using the r-th utterance of the R utterances at present or by using the content of the r-th utterance. Here, word (r) indicates the r-th utterance content (HMM number w of the utterance content).

Reference numeral 405 is a fuzzy vector quantizer.
The vocalization number r sent from the data control unit 404.
Using the correction feature vector storage unit 403,
Cuttle series Y^rAnd the feature vector y at each time t_t
^rIn the codebook storage unit 406 described later.
Is C₁~ C_MDistance between code vector and
Labels from 1st to Kth in order of closest separation, and feature vectors
Ry_tEach code vector searched for in the label set of
Feature vector y for_tReplaced with the degree of
Le vector o_t ^r= (O_t1, O_t2,, o_tK) And the degree of membership
Cutle u_t ^r= (U _t1, U_t2, ..., u_tK) To the above
Feature vector y_t ^rSeries Y^rLabel vector series O^r=
o₁ ^r, O₂ ^r,, o_T ^rAnd membership vector series U^r=
u₁ ^r, U₂ ^r, ..., u_T ^rIs to be converted to.

A codebook storage unit 406 stores the code vector C _m in a searchable form by the label m attached thereto, and is used by the fuzzy vector quantization unit 405 during vector quantization.

Reference numeral 407 denotes an HMM storage unit, which stores the state transition probability matrix A and the label occurrence probability matrix B for each vocabulary of which the number of vocabularies is W as shown in FIG. I'll do it. Therefore, the wth HMMλ
^w is expressed as λ ^w = {A ^w , B ^w }.

Reference numeral 408 denotes a codebook correction unit, which is a code vector C _m in the codebook storage unit 406.
Value of the correction voice and the probability that the HMM generates the correction voice by using the HMM corresponding to the utterance content of the correction voice stored in the HMM storage unit 407.
The code vector value is modified so that the maximum value of C becomes the maximum, and the modified new code vector C ^′ _m is transferred to the codebook storage unit 406.

Reference numeral 409 denotes a correction convergence determination unit, which determines the convergence status when the code vector is corrected using the correction voice. If the predetermined convergence condition is satisfied, the correction operation is terminated, If not satisfied, until satisfied,
Repeat the modification of the code vector.

A feature of the present invention is that the codebook correction unit 4
In the configuration of No. 08, the code vector is modified so that the probability (likelihood) that the HMM corresponding to the utterance content generates the voice is maximized, provided that the utterance content is known. . FIG. 5 is a block diagram showing a specific configuration of the codebook correction unit.

Each terminal 1-8 is connected to FIG.
Terminals 1 and 6 are connected to the codebook storage unit 406.
And terminal 1 receives codebook C,
6 is the modified codebook C^'To send. Terminals 4, 5
Is connected to the HMM storage unit 407, and the terminal 4 is
State transition probability matrix A of HMM corresponding to r-th speech
^{word (r)}And the terminal 5 has the same label occurrence probability matrix B
^{word (r)}To receive. Terminals 2 and 3 are the fuzzy vector
Connected to the audio quantizer 405,
Label vector series O^rAnd membership vector series U^rTo
To receive. The terminal 7 is connected to the correction convergence determination unit 409.
And the average likelihood L used for the convergence determination_{av e}Send
Believe. The terminal 8 is connected to the data control unit 404.
Information that the current correction voice data is the rth
When r = R is received, the correction vector ΔC and the average
Likelihood L_aveTo calculate.

While exchanging the information as described above, the codebook correction unit 408, that is, FIG. 5 is executed.

Numeral 501 is a feature vector sequence occurrence degree calculation unit, which is based on the attribution degree and the label occurrence probability based on the label occurrence probability matrix, the label vector series, and the attribution degree vector series received from the terminals 2, 3, and 5. The vector generation degree ω _i (t) is calculated for all times t and all states i of the HMM to obtain the feature vector generation degree matrix Ω. The correction vector calculation formula, which will be described later, differs depending on how the calculation formula of the feature vector generation degree is given.

Reference numeral 502 denotes a route probability calculating unit, which calculates a route probability γ _i (t) existing in a state i of the HMM at a certain time t for all times t and all states i of the HMM to obtain a route probability matrix. Find Γ. Also, the probability (likelihood) L that the HMM corresponding to the r-th utterance content generates the voice is L.
(O ^r , U ^r | λ ^{word (r)} ) is calculated and sent to a likelihood storage unit described later for convergence determination.

Reference numeral 503 is a correction vector denominator / numerator calculation unit that calculates the denominator and numerator of the correction vector calculation formula corresponding to the calculation formula of the feature vector generation ratio ω _i (t) in the feature vector sequence generation ratio calculation unit 501. To do.

A correction vector denominator / numerator storage unit 504 stores the denominator and numerator values of the correction vector calculation formula calculated by the correction vector denominator / numerator calculation unit 503 for use in the correction vector calculation unit described later.

A likelihood storage unit 505 stores R words for all correction words of the likelihood L (Y | λ ^w ) sent from the route probability storage unit 502.

After the above operation is performed for R correction voices (the signal from the terminal 7 becomes R), the operation described later is performed.

A correction vector calculation unit 506 stores the correction vector denominator / numerator memory based on a correction vector calculation formula corresponding to the calculation formula of the feature vector generation degree ω _i (t) in the feature vector sequence generation degree calculation unit 501. Part 50
A correction vector set ΔC is obtained from the denominator and numerator of the correction vector stored in 4.

Reference numeral 507 denotes a corrected code vector calculation unit, which uses the code vector value of the uncorrected codebook C received from the terminal 1 and the correction vector set ΔC obtained by the correction vector calculation unit 507. The code vector value of the corrected code book C ′ is calculated and transmitted from the terminal 6 to the code vector storage unit 406.

An average likelihood calculator 508 averages all likelihoods to be transmitted to the modified convergence determiner 409 through the terminal 7, and calculates an average likelihood L _ave .

The above is the configuration of the codebook correction unit in the present invention, but there are roughly two cases in this configuration. One is a method of individually obtaining the correction vector between the code vectors before and after the correction when correcting the code vector of the codebook, and the other is a method of obtaining the correction vector common to all the clusters.

First, the case where the former correction vector is individually obtained for each cluster will be described with reference to the flowcharts of FIG. 6 to FIG.

At S 601, a voice S whose voicing content of the modified speaker is known
It is confirmed whether ^r is stored in the correction voice storage unit 401. If it is stored, the process proceeds to the next step, and if not, a correction voice is uttered and stored as indicated by 602. 60
3 corresponds to converting the correction speech S ^r into a feature vector Y ^r by the feature extraction unit 402 using a well-known feature vector extraction means, which is executed for r = 1 to R and is shown at 604. Thus, they are stored in the correction feature vector storage unit 403.

The following operation is repeated until it is determined that the correction has converged. First, in 605, the buffers for the denominator and the numerator of the correction vector are cleared to zero, and the subsequent preparations are made. At 606 or 607, the feature vector series Y ^r of the correction voice data is read, and at 608, fuzzy vector quantization is executed in the fuzzy vector quantization unit 405 and the code vector storage unit 406 by a well-known method, and the degree of membership is calculated. The vector series U ^r and the label vector series O ^r are calculated.

609, that is, in FIG. 7, the calculation of the characteristic vector generation degree ω _i (t) in the characteristic vector sequence generation degree calculation unit 501 is performed by the calculation formula (7) for t = 1 to ^Tr , i = 1 to I. Perform according to equation 4).

At 610, the route probability γ _i (t) of the route probability calculating unit 502 is calculated by using a well-known forward / backward algorithm.

611, that is, in FIG. 8, the operation of the correction vector denominator / numerator calculation unit 503 is performed until the frame length T ^r of the r-th feature vector sequence Y ^r ends at 803 and at the r-th utterance content at 806. Until all the code vectors C _m (m =
1 to M) (satisfies the condition of 809), 810 calculates the denominator and the numerator of the correction vector calculation formula by (Equation 6) for the denominator and (Equation 7) for the numerator.
However, (Formula 6) and (Formula 7) are denominator and numerator formulas of the correction vector calculation formula (Formula 8) for each label m corresponding to the calculation formula (Formula 4) of the feature vector generation degree.

[0068]

[Equation 6]

[0069]

[Equation 7]

[0070]

[Equation 8]

Here, _{ΔC m} ^r _{_denom} and _{ΔC m} ^r _{_numer}
Represents the denominator and numerator of the calculation formula of the correction vector ΔC _m for the rth word and the mth cluster, respectively.

When all T ^r , I have been completed (806, 80
The condition of 3) is satisfied), and in 811 the values of the denominator and the numerator of the correction vector calculation formula are normalized by likelihood.

Until the condition that the last feature vector sequence Y ^R is selected in 612 is satisfied, that is, the operations from 608 to 611 are repeated for all feature vector sequences Y ^r (r = 1 to R), When the denominator and the numerator of the correction vector calculation formula for each code vector for the correction voice of are finished, 613 to 615, that is, the set of correction vectors ΔC = {ΔC ₁ , ΔC ₂ , ..., In FIG.
Calculation of ΔC _M ) and the corrected codebook C ^′ = {C ₁ ^′ , C
₂ ^' , ..., _CM ^' ) is calculated.

As shown in 903, the correction vector ΔC _m for each cluster is obtained by 613, that is, 904 according to (Equation 9) using the denominator and numerator of the correction vector calculation formula for all code vectors.

[0075]

[Equation 9]

Once the correction vector set ΔC is obtained,
After addition to the code vector of the codebook C before modify it (614 ie 905), replaces the codebook C ^'after correction as the new codebook C (6
15 or 906).

At 616, it is determined whether or not the correction has converged to a predetermined convergence condition. If it is determined that the correction has converged, the process ends, and the codebook at that time is set as the codebook for the corrected speaker. Further, when it is determined that it has not converged, the process returns to 605 and is repeated until it converges.

Next, in the latter case, that is, in the case of obtaining the correction vector commonly to all the clusters, FIG. 6, FIG. 7 and FIG.
This will be described with reference to the flowchart of FIG.

At S 601, the voice S whose utterance content of the modified speaker is known
It is confirmed whether ^r is stored in the correction voice storage unit 401. If it is stored, the process proceeds to the next step, and if not, a correction voice is uttered and stored as indicated by 602. 60
3 corresponds to converting the correction speech S ^r into a feature vector Y ^r by the feature extraction unit 402 using a well-known feature vector extraction means, which is executed for r = 1 to R and is shown at 604. Thus, they are stored in the correction feature vector storage unit 403.

The following operation is repeated until it is determined that the correction has converged. First, in 605, the buffers for the denominator and the numerator of the correction vector are cleared to zero, and the subsequent preparations are made. At 606 or 607, the feature vector series Y ^r of the correction voice data is read, and at 608, fuzzy vector quantization is executed in the fuzzy vector quantization unit 405 and the code vector storage unit 406 by a well-known method, and the degree of membership is calculated. The vector series U ^r and the label vector series O ^r are calculated.

609, that is, in FIG. 7, the calculation of the characteristic vector generation degree ω _i (t) in the characteristic vector sequence generation degree calculation unit 501 is performed at 707 for t = 1 to ^Tr , i = 1 to I. Perform according to equation 4).

At 610, the route probability γ _i (t) of the route probability calculating unit 502 is calculated using a well-known forward / backward algorithm.

In 611, that is, in FIG. 10, the operation of the correction vector denominator / numerator calculation unit 503 is performed until the frame length T ^r of the r-th feature vector Y ^r ends at 1003, and the r-th speech content at 1006. All the code vectors C _m (m
= 1 to M) (satisfies the condition of 1009), 101
0 for the denominator (Equation 10) and for the numerator (Equation 1)
In 1), the denominator and numerator of the correction vector calculation formula are calculated. However, (Formula 10) and (Formula 11) are denominator and numerator formulas of the correction vector calculation formula (Formula 12) common to all the labels corresponding to the calculation formula (Formula 4) of the feature vector generation degree.

[0084]

[Equation 10]

[0085]

[Equation 11]

[0086]

[Equation 12]

Here, _ΔC ^r _{_denom} and _ΔC ^r _{_numer} represent the denominator and the numerator of the calculation formula of the correction vector _ΔC common to all clusters of the r-th word.

When all T ^r , I have finished (1006, 1
(The condition of 003 is satisfied.) At 1011, the values of the denominator and the numerator of the correction vector calculation formula are normalized by the likelihood.

Until the condition that the last feature vector sequence Y ^R has been selected in 612 is satisfied, that is, 608 to 6 for all feature vectors Y ^r (r = 1 to R).
11 is repeated, and when the denominator and the numerator of the correction vector calculation formula common to all code vectors for all the correction voices are calculated, 613 to 615, that is, FIG.
In, the common correction vector ΔC and the corrected codebook C ^′ = {C ₁ ^′ , C ₂ ^′ , ..., C _M ^′ ) are calculated.

Using the denominator and the numerator of the correction vector calculation formula, the correction vector ΔC common to all clusters is obtained at 613, that is, 1101 according to (Equation 13).

[0091]

[Equation 13]

[0092] If the correction vector △ C is obtained, after the addition to the codebook C before modify it (614 i.e. 1105), replaces the codebook C ^'after correction as the new codebook C (615 ie 110
6).

At 616, it is determined whether or not the correction has converged to a predetermined convergence condition. If it is determined that the correction has converged, the process ends, and the codebook at that time is set as the codebook for the corrected speaker. Further, when it is determined that it has not converged, the process returns to 605 and is repeated until it converges.

In the above, the calculation of the characteristic vector generation degree ω _i (t) in the characteristic vector sequence generation degree calculation unit 501 of FIG. 5 and 609 of FIG. 6 and 707 of FIG. 7 is performed by the calculation formula (Equation 4). According to the above equation, the correction vector calculation formula is derived using the well-known Baum-Welch algorithm that is often used for deriving the estimation formula of the HMM under the condition of likelihood maximization, and the calculation is performed for each cluster as described above. In (Equation 8), if all clusters are common, the above (Equation 12)
When given in.

As a matter of course, in deriving the formula for calculating the correction vector, under the condition of maximum likelihood, Baum-
It is also possible to use the steepest descent method without using the Welch's algorithm, and it can be considered under the condition that the code vector distortion is minimum, not the maximum likelihood.

Next, the calculation of the feature vector occurrence degree ω _i (t) in the feature vector sequence occurrence degree calculation unit 501 in FIG. 5 and the calculation result in 609 in FIG. Without changing the configuration of
By modifying the configuration of one part, it is possible to provide a correction device that gives different effects.

FIG. 12 shows the operations of the characteristic vector sequence generation degree calculation unit 501 and the flowchart 609 of FIG. 6 when the calculation of the characteristic vector generation degree ω _i (t) is performed by the calculation formula (Equation 5). 9 is a flowchart and corresponds to FIG. 7 in the case where the calculation of the feature vector generation degree ω _i (t) is performed by the calculation formula (Equation 4). The basic operation is the same as that of FIG. 7, and the calculation formula in 1207 is different.

Also in this case, the correction vector may be obtained for each code vector or may be common to all code vectors, and the correction vector calculation formulas may be set under the likelihood maximization condition. When calculating for each cluster (Equation 1
4), and when all clusters are common, (Equation 15) is obtained.

[0099]

[Equation 14]

[0100]

[Equation 15]

By changing the calculation formula of the correction vector, only the structure of the codebook correction unit 408, especially the structure of the correction vector denominator / numerator calculation unit 503, that is, the procedure of 611 in the flowchart of FIG. 6 is changed. The change in the configuration will be described separately for the case where the correction vector is obtained for each cluster and the case where it is obtained for all clusters.

First, when the correction vector is individually obtained for each cluster, the correction vector denominator / numerator 50 is used.
According to the correction vector calculation formula (Formula 14), the denominator and the numerator calculation formula of the correction vector calculation formula individual to each cluster in 810 of FIG. 8 are as follows. (Equation 17)

[0103]

[Equation 16]

[0104]

[Equation 17]

When the denominator and the numerator of the correction vector calculation formula unique to each cluster are obtained, the correction vector ΔC _m unique to each cluster is calculated 613, that is, 904, according to (Equation 11) for all code vectors. And add it to the codebook C before modification (614, ie 9
05), the corrected codebook C ^′ is replaced with a new codebook C (615 or 906).

Next, when a correction vector common to all clusters is to be obtained, the correction vector denominator / numerator calculation unit 503.
Is performed according to the modified vector calculation formula (Equation 15),
The denominator and numerator calculation formulas of the correction vector calculation formula common to all clusters in 1010 of FIG. 10 are (numerical formula 18) and numerator (numeral 19), respectively.

[0107]

[Equation 18]

[0108]

[Formula 19]

If the denominator and numerator of the correction vector calculation formula common to all clusters are obtained, 61 according to (Equation 13) above.
3 or 1101 obtains the correction vector ΔC common to all clusters, adds it to the codebook C before correction (614 or 1105), and replaces the corrected codebook C ^′ as a new codebook C (615).
That is, 1106).

In the above embodiment, once the mapping vector of the codebook before and after the modification called the modification vector is obtained,
Although the corrected codebook is sought, it goes without saying that it is also possible to directly obtain the code vector of the corrected codebook so as to maximize the likelihood of the HMM for the correction speech. is there.

The voice recognition can be executed only by replacing the corrected codebook obtained in the above embodiment with the value in the codebook storage device 302 of the voice recognition device as shown in FIG.

The above is also a feature of the present invention, in which the codebook is changed with the maximum likelihood of the HMM.
The correspondence with the label occurrence probability of M is retained. For example, if you modify only the codebook under different conditions,
The correspondence between the codebook and the label occurrence probability may be broken, and the recognition rate may be reduced, but the present invention does not pose a problem.

In the above embodiment, the correction vector ΔC is obtained regardless of whether the calculation of the feature vector generation degree ω _i (t) is given by the calculation formula (Formula 4) or the calculation formula (Formula 5). Was added to the code vector C to update the code vector, and a constant vector ΔH (hereinafter referred to as a normalization vector) obtained based on the correction vector ΔC was input. By subtracting from the feature vector y _t of the person's voice as in (Equation 20), the speaker can be normalized by removing the difference in the input voice depending on the speaker.

[0114]

[Equation 20]

In this case, the normalization vector may be stored with the code vector value as it is, and the normalization vector may be subtracted from the feature vector of each frame of the feature vector series at the time of recognition. FIG. 13 is a block diagram of an apparatus for creating a normalization vector ΔH) for such speaker normalization.

Reference numeral 1301 denotes a correction voice storage unit, which is a known speaker (hereinafter referred to as a corrected speaker) who wants to obtain a normalized vector, that is, a known voice uttered by a speaker who needs a voice recognition system (in the feature vector normalization device). The number of occurrences R (r = 1 to R) of a voice S ^r (hereinafter referred to as a correction voice) of which the content of the generation is known) is stored and used for the subsequent correction.

Reference numeral 1302 denotes a feature extraction unit, which uses a technique similar to the feature extraction technique used in the recognition apparatus shown in FIG. 3 to output the correction speech S ^r at a constant time interval Y ^r = y. _{^{1 r, y 2 r, ...}} , y t r, ..., it is converted to y _T ^r. Here, T ^r is the number of data frames when the correction voice S ^r is converted into a feature vector sequence.

Reference numeral 1303 denotes a correction feature vector storage unit, which is the correction voice signal S ^r stored in the above 1301.
The feature vector series Y ^r extracted by the feature extraction unit 1302 is stored for r = 1 to R.

A data control unit 1304 controls the following processing by using whether the r-th utterance of the R utterances is currently handled or what the r-th utterance content is. Here, word (r) indicates the r-th utterance content (HMM number w of the utterance content).

Reference numeral 1305 denotes a normalized vector storage unit,
A normalization vector for correcting the feature vector is stored.

Reference numeral 1306 is a feature vector normalization unit,
The positive vector stored in the normalized vector storage unit 1305.
Using the normalization vector ΔH, at each time t of the feature vector
Value y _tTo obtain the modified feature vector by modifying
It is something.

Reference numeral 1307 denotes a fuzzy vector quantizer.
Voice number r sent from the data control unit 1304.
By using the correction feature vector storage unit 1303.
Characteristic vector series Y^rAnd read the characteristic vector at each time t
Ry_t ^rIs stored in the normalized vector storage unit 1305.
Using the normalized vector
The feature vector is modified by the digitization unit, and the resulting modified feature is
The codebook storage unit 13 to be described later with respect to the characteristic vector.
C stored in 08₁~ C_MWith code vector
Labels from 1st to Kth in order of the closest distance between the coutres,
Feature vector y _tEach label found in the label set of
Feature vector y for the code vector_tThe degree of belonging
Instead, the label vector o_t ^r= (O_t1, O_t2,, o_tK)
And the membership vector u_t ^r= (U_t1, U_t2, ..., u_tK)
In other words, the modified feature vector y _t ^r’Series Y^r’
Label vector series O^r= O₁ ^r, O₂ ^r,, o_T ^rAnd attribution
Degree vector series U^r= U₁ ^r, U₂ ^r, ..., u_T ^rConvert to
It is a thing. Where K is 1 to codebook size M
And the feature vector voicing degree described later.
Is to limit the number of calculation times.

A codebook storage unit 1308 stores the code vector C _m in a searchable form by the label m attached thereto, and is used by the fuzzy vector quantization unit 1307 during vector quantization.

Reference numeral 1309 denotes an HMM storage unit, which has a state transition probability matrix A and a label occurrence probability matrix B as shown in FIG.
Is stored for the number of vocabulary W. Therefore, the wth HMM
Is expressed as λ ^w = {A ^w , B ^w }.

Reference numeral 1310 is a normalization vector adjustment unit,
The value of the normalization vector ΔH in the normalization vector storage unit 1305 is used as the correction voice and the HMM storage unit 1309.
Corresponding to the utterance content of the correction voice stored in
Using M, the normalization vector is adjusted so that the probability (likelihood) that the HMM generates the correction speech is maximized, and a new adjusted normalization vector ΔH ^′ is stored in the normalization vector storage unit. It is transferred to 1305.

Reference numeral 1311 denotes a correction convergence determination unit which determines the convergence status when the normalization vector is corrected using the correction voice, and ends the correction operation when the predetermined convergence condition is satisfied. If not satisfied, the normalization vector is updated and the correction is repeated using the same correction voice until the satisfaction is satisfied.

The feature of the present invention resides in the configuration of the normalization vector adjusting unit 1310, and on the condition that the utterance content is known, the utterance content for the modified feature vector obtained by modifying the feature vector with the normalization vector. Is to adjust the normalization vector so that the likelihood by the HMM corresponding to

FIG. 14 is a block diagram showing the specific structure of the normalization vector adjustment unit. Each of terminals 1 to 9 is connected to FIG. 13 and terminal 1 is connected to the code vector storage unit 1308 to receive the codebook C. . The terminals 4 and 5 are connected to the HMM storage unit 1309. The terminal 4 outputs the state transition probability matrix A ^{word (r)} of the HMM corresponding to the r-th speech, and the terminal 5 also outputs the label occurrence probability matrix B ^{word ( r). r)} is received. Terminals 2 and 3 are connected to the fuzzy vector quantizer 1307, and r
The label vector sequence O ^r and the membership vector sequence U ^r for the th speech are received. The terminals 6 and 9 are connected to the normalized vector storage unit 1305, the terminal 6 receives the normalized vector ΔH, and the terminal 9 transmits the corrected normalized vector ΔH ′. The terminal 7 is the correction convergence determination unit 1
It is connected to 311 and transmits the average likelihood L _ave used for the determination of convergence. The terminal 8 is the data control unit 13
04, and receives information as to whether the current correction speech data is the r-th, and when r = R, calculates a normalized vector ΔH and an average likelihood L _ave .

While exchanging the information as described above, the codebook correction unit 1310, that is, FIG.
4 is executed.

Reference numeral 1401 is a feature vector sequence occurrence degree calculation unit, which is based on the attribution degree and the label occurrence probability based on the label occurrence probability matrix, label vector series, and attribution degree vector series received from terminals 2, 3, and 5. The vector generation degree ω _i (t) is calculated for all times t and all states i of the HMM to obtain the feature vector generation degree matrix Ω. The correction vector calculation formula, which will be described later, differs depending on how the calculation formula of the feature vector generation degree is given.

Reference numeral 1402 denotes a path probability calculating unit which calculates a path probability γ _i (t) existing in a state i of the HMM at a time t for all times t and all states i of the HMM.
Obtain the path probability matrix Γ. Further, the probability (likelihood) L (O) that the HMM corresponding to the r-th utterance content generates the voice is
^r , U ^r | λ ^{word (r)} ) is calculated and sent to a likelihood storage unit described later for convergence determination.

A correction vector denominator / numerator 1403 calculates the denominator and numerator of the correction vector calculation formula corresponding to the characteristic vector generation degree ω _i (t) in the characteristic vector sequence generation degree calculation section 1401.

A correction vector denominator / numerator storage unit 1404 stores the values of the denominator and the numerator of the correction vector calculation formula calculated by the correction vector denominator / numerator calculation unit 1403 for use in the correction vector calculation unit described later.

A likelihood storage unit 1405 stores R words for all correction words of the likelihood L (Y | λ ^w ) sent from the path probability storage unit 1402.

After the above operation is performed for R correction voices (the signal from the terminal 7 becomes R), the operation described later is performed.

Reference numeral 1406 denotes a correction vector calculation unit, which is based on a correction vector calculation formula corresponding to the generation degree ω _i (t) in the feature vector sequence generation degree calculation unit 1401.
The correction vector value ΔC is obtained from the denominator / numerator of the correction vector stored in the correction vector denominator / numerator storage unit 1404, and is sent to the correction vector storage unit 1305 from the terminal 9.

Reference numeral 1407 is a normalized vector calculation unit,
The correction vector ΔC obtained by the correction vector calculation unit
A new normalized vector ΔH ′ can be obtained from the uncorrected normalized vector ΔH received at the terminal 6.

Reference numeral 1408 denotes an average likelihood calculating section, which averages all likelihoods to send it to the correction convergence determining section 1311 through the terminal 7 to calculate an average likelihood L _ave .

In the above-described feature vector normalizing apparatus, the occurrence degree ω _i (t) is also the same as in the codebook correcting apparatus.
May be defined by the above (Formula 4) or may be defined by the above (Formula 5). The correction vector calculation formula is the above (Formula 12) and the latter case is the above (Equation 1
Same as 5).

The above is an embodiment of the codebook correction device and the feature vector normalization device of the present invention. In the above case, the correction voice is supposed to be uttered in advance, but it is considered from the viewpoint that the utterance content is known. When the reliability of the recognition result is high, the recognition result can be considered as the utterance content, and the speaker using the voice recognition system does not need to utter the correction voice in advance.

In this case, the reliability of the recognition result is high if the likelihood itself is large or the likelihood difference between the first candidate and the second candidate is large, and the reliability of the recognition result is otherwise. Since it is considered that the reliability of the code is low, an appropriate threshold value is set for each, and if the threshold value is exceeded, the codebook is corrected, and if not, the correction is not performed. Even if it is unknown, the codebook can be modified by using the recognition result as the utterance content.

A block diagram of such a voice recognition device will be described with reference to FIG. A feature extraction unit 1501 converts an unknown voice signal into a feature vector at regular time intervals using a well-known method such as LPC analysis, and outputs a sequence of feature vectors Y = y ₁ , y ₂ , ..., Y. _{Get t} , ..., y _T.
Here, T is the length of the feature vector sequence Y with respect to the unknown voice signal.

Reference numeral 1502 denotes a codebook storage unit which stores the code vector in a searchable form by the label attached to the code vector.

Reference numeral 1503 denotes a fuzzy vector quantizing unit, which is arranged in the order of closest inter-vector distance between the feature vector y _t extracted by the feature extracting unit 1501 and the code vector stored in the codebook storage unit 1502. and the K label, replaces the membership of the feature vector y _t for each of the code vectors searched by the label of the set of feature vectors y _t, label vector o _t = (o _t1,
o _t2 , ..., o _tK ) and the membership vector u _t = (u _t1 ,
u _t2, ..., is converted into u _tK), the feature vector y series Y label vector series _{t O = o 1, o 2} , ..., o T a membership vector series _{U = u 1, u 2,} ... , U _T.

Reference numeral 1504 denotes an HMM storage unit, which has the state transition probability matrix A and the label occurrence probability matrix B for each vocabulary to recognize the already created HMMλ ^w (w = 1 to W).
Is stored for the number of vocabulary W. Therefore, the wth HMM
Is expressed as λ ^w = {A ^w , B ^w } w = 1 to W.

Reference numeral 1505 is a feature vector sequence generation degree calculation
And the fuzzy vector quantizer 1503
The label vector series O and the degree of membership vector
Sequence U and w stored in the HMM storage unit 1504
Label occurrence probability matrix B of the th vocabulary^wUsing the HMM
λ^wThe occurrence degree matrix Ω of the feature vector sequence for^w= {Ω ^w
_itIs calculated according to the above (Equation 4) or the above (Equation 5)
Of.

Here, the label occurrence probability b _i o _tk is the label occurrence probability that the k-th label o _tk when the feature vector y _t at time t is fuzzy vector quantized occurs from the state i of the HMM.

Reference numeral 1506 denotes a likelihood calculating unit, which is the HMM calculated by the feature vector sequence generation degree calculating unit 1505.
using the feature vector series occurrence rate matrix Omega ^w a state transition probability matrix A ^w of HMMramuda ^w stored in the HMM storage section 1504 for lambda ^w, likelihood L | and calculates the (Y lambda ^w) .

[0149] 1507 is a likelihood storage unit, the likelihood L of each word HMMramuda ^w for the feature vector sequence Y calculated by the likelihood calculation unit 1506 | stores (Y lambda ^w) for comparison.

A comparison / determination unit 1508 determines the vocabulary corresponding to the HMM giving the maximum value among the likelihoods of each HMM stored in the likelihood storage unit 1507 as a recognition candidate.

The above 1505 to 1507 are HM of each vocabulary
It is performed once for each Mλ ^w , repeated from w = 1 to W, and the result is evaluated by the comparison / determination unit 1508.

Reference numeral 1509 denotes a recognition candidate reliability calculation unit,
The reliability of the recognition candidate selected by the comparison determination unit 1508 is calculated using the likelihood of the recognition candidate stored in the likelihood storage unit 1507.

Reference numeral 1510 is a codebook correction execution determination unit, and if the reliability of the recognition candidate obtained from the recognition candidate reliability calculation unit 1509 is a predetermined threshold value or more, a codebook correction signal will be described later. Send to the codebook modification section to execute the codebook modification.

Reference numeral 1511 denotes a codebook correction unit, which receives the codebook correction signal from the codebook correction execution determination unit and receives the codebook and the fuzzy vector quantization stored in the codebook storage unit 1502. The codebook is corrected using the label vector series O, the belonging degree vector series U, and the HMM corresponding to the recognition candidate stored in the HMM storage unit obtained by the unit 1503. The codebook is sent to the codebook storage unit.

Similarly, if the normalization vector is adjusted when the threshold value is exceeded and adjustment is not made otherwise, the recognition result is used as the utterance content even if the utterance content is unknown. The normalization vector can be modified.

A block diagram of such a voice recognition device will be described with reference to FIG. Reference numeral 1601 denotes a feature extraction unit that converts an unknown voice signal into a feature vector at regular time intervals using a well-known method such as LPC analysis, and outputs a sequence of feature vectors Y = y ₁ , y ₂ , ..., Y. _{Get t} , ..., y _T.
Here, T is the length of the feature vector sequence Y with respect to the unknown voice signal.

Reference numeral 1602 denotes a normalized vector storage section,
A normalization vector for normalizing the feature vector is stored.

Reference numeral 1603 is a feature vector normalization unit,
The feature vector is further normalized by the normalization vector.

A codebook storage unit 1604 stores the code vector in a searchable form by the label attached to the code vector.

Reference numeral 1605 is a fuzzy vector quantizer.
After the normalization by the feature vector normalization unit
Feature vector y'of_tAnd the codebook storage unit 1604
Of the distance between the vector and the code vector stored in
K labels in the closest order and the feature vector after normalization
y '_tEach code vector searched for in the label set of
The normalized feature vector y ′ for_tReplaced with
E, label vector o _t= (O_t1, O_t2,, o_tK)When
Membership vector u_t= (U_t1, U_t2, ..., u_tK)Conversion to
Then, the normalized feature vector y ′_tThe series Y'of
Le vector series O = o₁, O₂,, o_TAnd degree of membership
Le series U = u₁, U₂, ..., u_TIs to be converted to.

Reference numeral 1606 denotes an HMM storage unit, which has the state transition probability matrix A and the label occurrence probability matrix B for each vocabulary to recognize the already created HMMλ ^w (w = 1 to W).
Is stored for the number of vocabulary W. Therefore, the wth HMM
Is expressed as λ ^w = {A ^w , B ^w } w = 1 to W.

Reference numeral 1607 denotes a feature vector sequence generation degree calculation unit, which is stored in the HMM storage unit 1606 and the label vector sequence O obtained by the fuzzy vector quantization unit 1605 and the membership degree vector sequence U.
Using the label occurrence probability matrix B ^w of the th vocabulary, the HMM
The occurrence degree matrix Ω ^w = {ω ^w _it } of the normalized feature vector series for λ ^w is calculated according to the (Equation 4) or the (Equation 5).

Here, the label occurrence probability b _i o _tk is the label occurrence probability that the k-th label o _tk when the feature vector y ′ _t at time t is fuzzy vector quantized occurs from the state i of the HMM. .

Reference numeral 1608 denotes a likelihood calculating unit, which is the HMM calculated by the feature vector sequence generation degree calculating unit 1607.
calculates a ^{| (λ w Y ') λ} w using a state transition probability matrix A ^w of the feature vector series occurrence rate matrix Omega ^w wherein stored in the HMM storage unit 1606 and HMMramuda ^w relative likelihood L is there.

A likelihood storage unit 1609 stores the likelihood L (Y '| λ ^w ) of each word HMMλ ^w for the feature vector series Y calculated by the likelihood calculation unit 1608 for comparison.

A comparison / determination unit 1610 determines the vocabulary corresponding to the HMM giving the maximum value among the likelihoods of each HMM stored in the likelihood storage unit 1609 as a recognition candidate.

Numerals 1607 to 1609 are HM of each vocabulary
It is performed once for each Mλ ^w , repeated from w = 1 to W, and the result is evaluated by the comparison / determination unit 1610.

Reference numeral 1611 is a recognition candidate reliability calculation unit,
The reliability of the recognition candidate selected by the comparison determination unit 1610 is calculated using the likelihood of the recognition candidate stored in the likelihood storage unit 1609.

Reference numeral 1612 denotes a normalization vector adjustment execution determination unit, which outputs a normalization vector adjustment signal if the reliability of the recognition candidate obtained from the recognition candidate reliability calculation unit 1611 is a predetermined threshold value or more. It is sent to the normalization vector adjustment unit described later to execute the normalization vector adjustment.

Reference numeral 1613 is a normalization vector adjustment unit,
Upon receiving the normalization vector adjustment signal from the normalization vector adjustment execution determination unit, the normalization vector storage unit 1
The normalization vector stored in 602, the label vector sequence O obtained by the fuzzy vector quantization unit 1605, the membership vector sequence U, and the HMM.
HMM corresponding to the recognition candidate stored in the storage unit
And are used to adjust the normalization vector, and the corrected normalization vector is sent to the normalization vector storage unit.

[0171]

【The invention's effect】

(1) According to the present invention, a voice whose utterance content is known is used,
By correcting the codebook using the correction vector obtained so as to maximize the likelihood of the HMM for the voice, the difference between the environments at the time of creating the HMM and at the time of recognition is corrected, and the voice recognition device is resistant to the changes in the environment. Can be configured. (2) According to the present invention, a voice whose utterance content is known is used,
By correcting the feature vector using the normalized vector obtained so as to maximize the likelihood of the HMM for the voice, the difference in the environment between when the HMM is created and when the recognition is performed is corrected, and the voice recognition is robust against the change in the environment. It is possible to configure the device.

[Brief description of drawings]

FIG. 1 is a block diagram of an HMM for explaining a Hidden Markov Model (HMM).

FIG. 2 is a diagram illustrating the configuration of a codebook.

FIG. 3 is a block diagram illustrating a conventional example of a voice recognition device.

FIG. 4 is a block diagram illustrating an example of an outline of a codebook correction device according to the present invention.

FIG. 5 is a block diagram illustrating an embodiment of a codebook correction unit which is a main part of the codebook correction device of the present invention.

FIG. 6 is a flowchart illustrating the operation of the present invention.

FIG. 7 is a flowchart for explaining the operation when the occurrence degree calculation formula in FIG. 6 is expressed by (Equation 4).

FIG. 8 is a flowchart illustrating an operation of calculating a correction vector denominator / numerator when the correction vector in FIG. 6 is obtained for each cluster.

9 is a flowchart illustrating an operation when a correction vector in FIG. 6 is obtained for each cluster.

FIG. 10 is a flowchart illustrating an operation of calculating a correction vector denominator / numerator when the correction vector in FIG. 6 is commonly found in all clusters.

FIG. 11 is a flowchart for explaining the operation when the correction vector in FIG. 6 is obtained commonly to all clusters.

FIG. 12 is a flowchart for explaining the operation when the expression for calculating the degree of occurrence in FIG. 6 is expressed by (Equation 5).

FIG. 13 is a block diagram illustrating an example of an outline of a feature vector normalization device of the present invention.

FIG. 14 is a block diagram illustrating an embodiment of a correction vector correction unit which is a main part of the feature vector normalization device of the present invention.

FIG. 15 is a block diagram illustrating an embodiment of a voice recognition device incorporating a codebook normalizing means.

FIG. 16 is a block diagram illustrating an embodiment of a voice recognition device incorporating a normalization vector adjustment means.

Claims (12)

[Claims] 1. A codebook storage means for storing a finite number of representative points (code vectors) in a feature vector space in a searchable form by labels attached thereto, and each vector of a feature vector series by the codebook. Fuzzy vector quantizing means for converting to a set of degree of membership corresponding to a label (degree of membership vector) and converting the feature vector series into a degree of membership vector series, and the probability of occurrence of the label (probability of occurrence of label) for each state. HMM storage means for storing the defined HMM, feature vector series occurrence degree calculation means for calculating the occurrence degree of the feature vector series from the HMM based on the label occurrence probability and the degree of membership vector, and each of the code vectors A codebook modifying means for modifying, wherein the codebook modifying means uses the HM as the feature vector sequence; It includes modification vector calculation means for modifying the code vector to maximize the degree generated from codebook adjustment device also being configured to modify the code vector. 2. The codebook correction device according to claim 1, wherein the correction vector calculation means individually calculates the value of the correction vector for each code vector of the codebook. 3. The codebook correction device according to claim 1, wherein the correction vector calculation means calculates the value of the correction vector in common to all codevectors of the codebook. 4. The feature vector sequence occurrence degree is defined as a power product of the label occurrence probability and the membership degree, or the logarithm of the occurrence degree is defined by a sum of products of the logarithm of the label occurrence probability and the membership degree. Codebook correction device described. 5. The codebook correction device according to claim 1, wherein the degree of occurrence of the feature vector series is defined by the sum of products of the label occurrence probability and the degree of membership. 6. A codebook storage means for storing a finite number of representative points (code vectors) in a feature vector space in a searchable form by labels attached thereto, and each vector of the feature vector series by the codebook. Fuzzy vector quantizing means for converting to a set of degree of membership corresponding to a label (degree of membership vector) and converting the feature vector series into a degree of membership vector series, and the probability of occurrence of the label (probability of occurrence of label) for each state. HMM storage means for storing the defined HMM, feature vector series occurrence degree calculation means for calculating the occurrence degree of the feature vector series from the HMM based on the label occurrence probability and the membership degree vector, and the feature vector is modified And a normalization vector for calculating a normalization vector for correcting the feature vector And a correction vector calculating means for correcting the feature vector so as to maximize the degree of occurrence of the feature vector sequence from the HMM. A feature vector normalization device having the following configuration. 7. The degree of occurrence of the feature vector sequence is characterized in that the power product of the label occurrence probability and the degree of membership or the logarithm of the degree of occurrence is defined by the sum of products of the log of the label occurrence probability and the degree of membership. The described feature vector normalizer. 8. The feature vector normalization apparatus according to claim 6, wherein the occurrence degree of the feature vector series is defined by the sum of products of the label occurrence probability and the membership degree. 9. The codebook correction device according to claim 4 or 5, wherein the range of exponentiation products or the range of product sums of the feature vector sequence generation degree calculation means is limited. 10. The feature vector normalizing apparatus according to claim 7 or 8, wherein the range of exponentiation products or the range of product sums of the feature vector sequence generation degree calculating means is limited. . 11. A codebook storage means for storing a finite number of representative points (code vectors) in a feature vector space in a searchable form by labels attached thereto, and each vector of the feature vector series by the codebook. Fuzzy vector quantizing means for converting to a set of degree of membership corresponding to a label (degree of membership vector) and converting the feature vector series into a degree of membership vector series, and the probability of occurrence of the label (probability of occurrence of label) for each state. HMM storage means for storing the defined HMM, feature vector series occurrence degree calculation means for calculating the occurrence degree of the feature vector series from the HMM based on the label occurrence probability and the membership degree vector, and the feature vector series H of each vocabulary
Likelihood calculation means for calculating the likelihood of MM, comparison judgment means for judging the recognition result, recognition candidate reliability calculation means for calculating the reliability of the recognition candidates obtained from the comparison judgment means,
If the reliability of the recognition candidate exceeds a certain threshold, a codebook correction execution determination means for issuing an instruction to execute a codebook correction and a codebook correction means for correcting each code vector are provided. The means includes correction vector calculation means for correcting the code vector so as to maximize the degree of occurrence of the feature vector sequence from the HMM corresponding to the recognition candidate, and the recognition candidate is uttered when the utterance content is unknown in advance. A voice recognition device characterized in that the code vector is modified by the following. 12. A codebook storage means for storing a finite number of representative points (code vectors) in a feature vector space in a searchable form by labels attached thereto, and each vector of the feature vector series by the codebook. Fuzzy vector quantizing means for converting to a set of degree of membership corresponding to a label (degree of membership vector) and converting the feature vector series into a degree of membership vector series, and the probability of occurrence of the label (probability of occurrence of label) for each state. HMM storage means for storing the defined HMM, feature vector series occurrence degree calculation means for calculating the occurrence degree of the feature vector series from the HMM based on the label occurrence probability and the membership degree vector, and the feature vector series H of each vocabulary
Likelihood calculation means for calculating the likelihood of MM, comparison judgment means for judging the recognition result, recognition candidate reliability calculation means for calculating the reliability of the recognition candidates obtained from the comparison judgment means,
If the reliability of the recognition candidate exceeds a certain threshold, a codebook correction execution determining means for issuing an instruction to execute a codebook correction, a feature vector correcting means for correcting the feature vector, and the feature vector are corrected. A normalization vector adjustment means for calculating a normalization vector for
The normalization vector adjustment means includes a correction vector calculation means for correcting the feature vector so as to maximize the degree of occurrence of the feature vector sequence from the HMM, and when the utterance content is previously unknown, the recognition candidate is uttered. The voice recognition device is configured to correct the feature vector according to the above.