JPH10319986A - Acoustic model creating method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make learning voice data a model capable of a recognition having flexibility without being influenced by the distinctiveness of the learning voice data being origins of an acoustic model (HMM). SOLUTION: Voice data for learning 17 are converted into a featured parameter series (18) and this series is subjected to a voice recognition in which free connections are allowed in a phoneme unit and a restriction of vocabulary is not present (32) and not only a correct interpretation but also plural contrast candidates near this are calculated from this recognition. Then, parameters of an initial acoustic model 31 are corrected so that errors are decreased by using probabilities of them (33) and a voice recognition is performed again by using these corrected acoustic model 34 and the same things are repeated, in short, a recognition learning is performed by an error minimizing method (MCE).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of creating a model used in a pattern recognition method for recognizing input data by modeling a feature amount of each category and obtaining a probability of each model with respect to an input feature amount sequence. .

[0002]

2. Description of the Related Art A method using a probability model based on probability and statistics is a useful technique in pattern recognition of voice, characters, figures, and the like. In the following, the Hidden Markov Model (HMM)
) Will be described. The hidden Markov model is described in, for example, Seiichi Nakagawa, “Speech Recognition by Stochastic Model,” edited by the Institute of Electronics, Information and Communication Engineers (1988).

In a conventional speech recognition apparatus, a method of modeling a certain speech unit (phonemes, syllables, words, etc.) for each category by using an HMM has a high performance and is currently the mainstream. FIG. 4 shows a functional configuration example of a conventional speech recognition apparatus using an HMM. The audio input from the input terminal 11 is converted into a digital signal by the A / D converter 12. A voice feature parameter extraction unit 13 extracts voice feature parameters from the digital signal. The HMM created in advance for each voice unit is read from the model parameter memory 14, and the model probability calculation unit 15 calculates the probability of each model with respect to the input voice. The speech unit represented by the model having the highest probability is output from the recognition result output unit 16 as a recognition result.

[0004] H as an acoustic model often used at present
MM is a three-state three-loop. An HMM is created for each voice unit (generally, a word, a phoneme, a syllable, and the like). Each state is provided with a statistical probability distribution of a voice feature parameter. In the current mainstream, phonemes or syllables are used as speech units instead of words, and these HMMs are connected and used according to the vocabulary to be recognized. Before using in the recognition device, an acoustic model is generated using the acoustic model learning speech data 17. The learning data from the database 17 is converted into a feature parameter by a speech feature parameter extraction unit 18, and the obtained model is used in an acoustic model parameter learning unit 19 based on the initial model obtained by the initial acoustic model generation unit 21. To learn. The model parameters obtained here are used by the recognition device.

In the field of pattern recognition, there is a method of representing each category by a probability distribution and estimating the probability distribution so as to reduce an identification error in order to improve an identification rate. In the field of speech recognition, in the HMM, an MCE method has been proposed as one method for improving recognition accuracy by learning between error-prone models so as to minimize identification errors. This is described in Chin-Hui Lee, Frank K. Soong and Kuldip K. Pa
liwal, “Automatic Speech and Speaker Recognition─
"Advanced Topics" (Kluwer Academic Publishers, 1
996), Chapter 5, B.-H. Juang, W. Chou and C.-H. Lee,
“Statistical and Discriminative Methods for ASR”
Are outlined in

In such discrimination learning, when there is only one model in the recognition target, the correct category and the opposing category may be simply learned so as to reduce errors. When there are a plurality of modeled models in the recognition target, segmentation is performed using the model, and learning is performed to minimize the discrimination error between the correct category and the opposing category in the section.

[0007]

By using this discriminative learning, word HMMs and subword HMMs in units smaller than words (such as phonemes and syllables in Japanese) can be created to improve recognition accuracy. However, in the word HMM, the vocabulary that can be recognized is fixed. In addition, by simply recognizing the learning speech even in subword units and using the recognition error as an alternative candidate, the word is fixed, and the task-dependent model, that is, the training speech data at the time of model creation consists only of the city name In this case, the model depends on the learning data, and when only the numeric voice data is used, the model depends on the learning data. In this case, if the recognition task is different from the task of the learning data, improvement in accuracy cannot be guaranteed. In particular, in a general vocabulary-based speech recognition method, even if the alignment between categories is not accurately matched, if there is a part with a high probability, a partly high score becomes dominant and the candidate becomes a candidate. Becomes dominant and often causes recognition errors. If such a thing is discriminated and learned as an opposition candidate, there is a possibility that a partially distorted segment may be learned. In a recognition task other than the learning task, the accuracy is rather lower than before the recognition learning.

The present invention has been made in view of the above-described drawbacks of the conventional method, and has as its object to provide a method of creating a general-purpose acoustic model by discriminating learning without depending on a learning task. .

[0009]

According to the present invention, in order to create a speech model that can be used for general purposes using discrimination learning, the recognition result to be discriminated has no vocabulary restriction, that is, no phoneme or acoustic unit. And a recognition result generated by a speech recognition system that allows a free chain. The speech recognition system without vocabulary here means that it has no vocabulary,
In this method, arbitrary chains such as phonemes or syllables are allowed so that any word or sentence can be recognized.

At present, the recognition accuracy of such a voice recognition method is generally low. However, partial errors include:
It is considered that there is often a bias. If this general error tendency can be identified and learned, the obtained speech model will be a speech model that can be used for general purposes. It is also conceivable that the learning section includes only the error section as a learning method.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS As an embodiment of the present invention, an example in which discrimination learning is MCE will be described below. First, MC
The general formulation of E will be described. Here, reference C.-
S.Liu, C.-H.Lee, W.Chou, B.-H.Juang and AERosenber
g, “A study on minimum erroe discriminative traini
ng for speaker recognition, "J.Acoust.Soc.Am.97
(1), pp. 637-648, January 1995, will be described along with those formulated in APPENDIX.

In the parameter set P of the acoustic model, the log likelihood for the class (model) k of the observed vector sequence X is defined as an identification function g _k (X, （), and the identification error function is defined as d _k (X, P) = −. g _k (X, P) + G _k (X, P) (1) Here, G _k (X, P) = log [1 / (K−1) {exp {eg _j (X, Λ)}] ^{1 / e} (2)} is the value of each value of j other than j = k. This is an addition, and corresponds to the likelihood of an alternative candidate (error-prone model) for the correct class (model) k. Where K is the number of opposing candidates, e
Is a constant. The loss function L is defined as follows using a sigmoid function.

L _k (X; P) = L (d _k ) = 1 / {1 + exp (−a (d _k + b))} (3) where a and b are constants. The parameter set P includes a transition probability A, a weight C of the mixture distribution, an average value μ, and a variance U. To avoid becoming negative, transition probabilities, weights, and variances are formulated logarithmically, and the parameter set is expressed as P ′
And The parameters of the acoustic model are updated by the following equation.

P _{t + 1} ′ = P _t ′ − _t V _t ∇L _k (X; P ′) | _{P ′ = P ′ t} (4) where _t is a small positive real number with a learning step size, V _t Is a positive definite matrix. Minute variation of each parameter ΔL _k (X;
P | controls the ') P' _{= P't} in ∈ _t and V _t, to update the parameters P '. The optimum point is found by gradually moving the parameter P '.

Each class (model) C _j , j = 1,..., K
∇L _k (X; P ′) is expressed by the following equation. _{∇ rj L k (X; P} ') = (∂L k / ∂d k) (∂k k / ∂g i) · ∇ rj [g j (X; P')] (5) where, L _k Is defined by equation (3). Each _{term, ∂L k / ∂d k = a} · L k (1-L k) (6) ∂d k / ∂g j = -1, when j = k = [Σexp {eg n (X , P)}] ^-1 · exp {eg _j (X, P)}, j ≠ k (7)} is an addition for each value of n other than n = k. Here, when j = k is the weight for the correct answer (candidate),
When j ≠ k, the weight is given to the conflict candidate in accordance with the magnitude of the likelihood. As can be seen from equation (4), when the answer is correct, the second term on the right side has a positive sign, the whole parameter approaches the parameter estimated by learning, and in the case of an alternative candidate, it has a negative sign. , Will move in the opposite direction from the estimated parameters.

Consider a case where the log likelihood g _j (X, P) has a Gaussian mixture distribution. For ∇g _j , the transition probability, average, variance, and weight coefficient are calculated as follows. For transition probabilities a _{s1, s2} from the state S1 of class j to state S2 _{is, ∂g i / ∂a 's1,} s2 = Σ t = 1 T δ (q t-1 -S1) δ (q t -S2 , S1 = 1,..., N, S2 = S1 or S1 + 1.
_{s1m = {μ s1m, c s1m} , u s1m} _{respect, ∂g i / ∂B s1m = Σ} t = 1 T (1 / b qt (x t)) · (∂b qt (x t) / ∂ _{B s1m) δ (q t -S1} ) (9) for the average value mu _{S1M is, ∂b qt (t) / ∂μ} s1m = c s1m N (x t | μ s1m, U s1m) (x t -μ s1m ) / U _s1m (10) For the variance U _s1m ′, ∂b _qt (x _t ) / _{∂U s1 m} ′ = c _s1m N (x _t | μ _s1m , U _s1m ) · (1/2) ｛(x _{t ^{-μ s1m) 2 / U s1m -1}} } (11) weighting coefficients c _S1M the mixture distribution _{'for, ∂b qt (x t) /} ∂c s1m' = c s1m N (x t | μ s1m, U s1m (12) As described above, the acoustic model is created by the identification error learning. When the task-dependent acoustic model is created at this time, any acoustic model that matches the task or vocabulary may be used. When creating an independent acoustic model, Use the results of the unconstrained by the speech recognition system.

FIG. 1 shows a procedure of discriminative learning according to the present invention. The feature parameter of the learning speech data 17 is extracted by the extraction unit 18, and the feature parameter sequence of the learning speech data 17 is recognized by the speech recognition unit 32 without vocabulary restriction using the initial acoustic model parameters 31. Here, a plurality of candidates and their probabilities are obtained. The identification error learning acoustic parameter estimating unit 33 performs parameter estimation to minimize the above-described identification error d _k or L _k , and updates the acoustic model parameters. Using the updated acoustic model parameters 34, the vocabulary-free speech recognition unit 32 recognizes a feature parameter sequence of the learning acoustic data. By repeating this, an optimal acoustic model can be obtained. Also, among the acoustic model parameters 34, the variance and the weighting factor become values biased toward the learning data when the amount of the learning data is small, and conversely there is a possibility that the accuracy decreases. If the amount of learning data is small, only the most effective average value can be learned for other parameters.

The speech section used for learning, that is, parameter estimation, differs depending on how the observation vector sequence X is defined in the equation (1), and the characteristics of the obtained acoustic model also differ. In the following description, a phoneme is used as a segment unit. However, when an HMM is actually used, the segment unit is in a state. If one phoneme HMM has three states, one phoneme section is segmented into three state sequences. In other words, the correct phoneme and the wrong phoneme (alternative candidate)
Are updated such that errors between each of the corresponding states are reduced.

(A) In the invention according to claim 4, equation (1) uses the observation vector sequence X as the entire input speech, and uses a plurality of vocabulary-free speech recognition candidates as the opposition candidates. As shown in FIG. 2A, parameters are estimated for all segments obtained from each candidate. In this case, there is a possibility that a correct part is also included in the segment, and when there are many correct sections with respect to the error section, the error is not so much reflected in the learning.

(B) In the invention according to claim 5, a plurality of candidates in speech recognition without vocabulary restriction are used as opposing candidates, and as shown in FIG. The parameter is estimated using only the section subjected to. Multiple candidates obtained by the vocabulary-free recognition system are:
Since there are many cases where a part is incorrect, it is possible to emphasize and emphasize a part that is likely to be wrong. In this case, the equation (1) is transformed into the following equation.

[0021] _{d k (X kj, r k} , r j) = - g k (X kj, r k) + g j (X kj, r j), j ≠ k (13) Here, r _k: the correct answer Parameter set for partial class (model) k, r _j : Parameter set for partial class (model) j in which a section different from the correct answer of the opposing candidate exists, and is a subset of P. Correct answer subclass k
X _kj : a segment section different from the correct answer in the _alternative candidate, and corresponds to class j. That is, equation (1)
Is replaced with X _kj , and d _k considers only the section in which the error occurs. For example, in the part of the correct answer "0", only the first part is updated.

(C) As a comparison method, a case will be described in which a speech recognition system without vocabulary restriction is not used, but each correct phoneme segment is replaced with another phoneme as it is. FIG. 2 (c)
As shown in (1), each correct phoneme section corresponds to X, and the alternative candidate is a phoneme having a similar likelihood other than the correct answer obtained for each section. In this method, learning can be easily performed without using the result of a recognition system having no vocabulary constraint. That is, for example, the correct answer i is recognized and recognized by the class of the contending candidate a, and then the embodiment showing the effect of the present invention is shown. A score is obtained, and the correct answer i is similarly recognized in all the other classes and viewed, and an alternative candidate is selected in descending order of the score of the recognition result, and parameter estimation is performed using the parameters of the alternative candidate.

The experimental conditions include a 12 kHz sampling frequency, a frame length of 32 ms, and a frame period of 8 ms.
And the feature amount is a 16th-order selected linear prediction cepstrum,
A 16th order ΔCepstrum and ΔPower were used. The initial model contains about 450 states of the phoneme environment-dependent HMM.
It is a 27 phoneme model of a mixture distribution. The training data of the initial model is the ATR database A set phoneme balance 21
6 words and 5240 important words were 10 males and 10 females, respectively, and the Acoustic Society of Japan database 503 sentences were 3 males
A general-purpose acoustic model was created using 9,600 sentences in total of 0 and 34 women.

For discriminative learning, 216 words from the ATR database C set were used by 10 males and 10 females. In order to create a general-purpose model, the training data needs to include a wide range of phonemes and phoneme environments. Therefore, 216 words with well-balanced phonemes were used. The evaluation was performed using (a) a learning task, (b) an unlearned task, (c) a speech recognition system without vocabulary constraints, and (d) a word speech recognition system (speech recognition system with vocabulary constraints). Evaluation data is
As the evaluation for the learning task, an unlearned speaker of 216 words in the same C set, five men, and five women were used. As evaluations for unlearned tasks, 100 city names, 5 males, and 4 females recorded in different environments were used. In the word-speech recognition experiment, evaluation was performed using 1002 word names and vocabulary of 1202 words in order to increase the difficulty of the recognition task for 100 city name data.

FIG. 3 shows the results. "No long vowel" in the table
Indicates that the answer was correct regardless of whether it was a long vowel in the evaluation, and "has a long vowel" indicates that only the correct long vowel and single vowel were correct. For both the learning task (216 words) and the unlearned task (100 city names), the accuracy of speech recognition without vocabulary restriction can be improved. In the learning task, both “a” and “b” were about 2% in “no long vowel”
As for "long vowels", about 3% improvement is seen.
In (c), the degree of improvement is slightly small. In the unlearned task, (a) is 13.1% in “with long vowel”,
(B) shows an improvement of 7.4%, and (c) shows an improvement of 11.7%. In particular, it can be seen that the improvement of the long vowel part is large.

As word recognition, for a learning task,
Almost the same or slightly upward trend is seen. For the unlearned task, all three methods have about 2% improvement. That is, it can be seen that an acoustic model independent of the task has been created.
Comparison of the methods (a), (b) and (c) shows that the method (a) is relatively good in any evaluation. Note that the discrimination learning is not limited to the example described above, and a candidate that is likely to be erroneous may be searched for, and the parameter may be updated so that errors with the candidate are reduced.

[0027]

As described above, according to the present invention, a general-purpose acoustic model can be created without depending on the accuracy of the acoustic model depending on the learning task.

[Brief description of the drawings]

FIG. 1 is a functional configuration diagram showing a flow of identification learning using a speech recognition system without vocabulary restriction according to the present invention.

FIG. 2 is a diagram showing, for each method, a difference in a learning section targeted for identification learning of an acoustic model according to the present invention;

FIG. 3 is a diagram showing an experimental result as an effect of the present invention.

FIG. 4 is a block diagram showing a functional configuration of the speech recognition device.

Claims (5)

[Claims] 1. A method for creating a general-purpose acoustic model from an initial probabilistic acoustic model, comprising: extracting speech characteristic parameters corresponding to the probabilistic acoustic model from learning speech data; A process of performing vocabulary-free speech recognition that allows free chaining in units smaller than words, phonemes or syllables, and using the recognition result of the speech to find an alternative candidate that differs from the correct recognition result And updating the initial acoustic model by discriminating learning to reduce the acoustic model. 2. The method according to claim 1, wherein the stochastic acoustic model is a hidden Markov model. 3. The method according to claim 2, wherein the discrimination learning is error minimization learning (MC
3. The acoustic model creation method according to claim 2, wherein E: Minimum Classification Error). 4. The sound according to claim 1, wherein the discriminative learning learns parameters of the acoustic model by using respective results of a correct answer and an alternative candidate in all segment sections. Model creation method. 5. The method according to claim 1, wherein the discriminative learning learns the parameters of the acoustic model from the results of the opposing candidates only in the segment sections different from the correct answer using the results. Acoustic model creation method.