JPH08248974A - Speech recognition device and standard pattern learning method - Google Patents

Abstract

(57) [Summary] [Purpose] The object of the present invention is to measure the spread of the distribution of each standard pattern at each stage of learning of the standard patterns, and configure each standard pattern to have the same spread of distribution. , To realize highly accurate voice recognition. [Structure] In the standard learning process, standard pattern learning is performed using a conventional standard pattern learning method. For each standard pattern created, measure the spread of the distribution. In the process of determining the balance as a standard pattern set, the balance as a set of standard patterns is determined according to the measurement result of the spread of the distribution. If it is determined that the balance is unbalanced, a standard pattern with a large distribution spread is selected from the standard pattern set, and a plurality of standard patterns are selected so that the distribution spread of each standard pattern does not become large.
The above processing is repeated until it is determined that the balance of the set of standard patterns is good.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus that uses a basic unit in a phonetic language expression such as a syllable or a phoneme (consonant or vowel) as a standard pattern, and a method for creating the standard pattern.

[0002]

2. Description of the Related Art Several units such as a word unit, a syllable unit, a phonological unit (a vowel and a consonant subdivided unit) can be considered as a unit of a standard pattern in a speech recognition apparatus. What kind of standard pattern is used in the speech recognition device depends on the size of the target vocabulary to be recognized, the ability to express speech phenomena in the standard pattern, the amount of learning voice data that can be used for learning the standard pattern, etc. Make a decision after considering.

First, consideration will be given from the viewpoint of the number of standard patterns. For example, in a small vocabulary voice recognition device that recognizes only 10 numbers, the number of standard patterns is about 10 even if a standard pattern of word units (single digit numbers) is used. However, it recognizes arbitrary sentences in Japanese and recognizes words in a large vocabulary (for example, all names of Japanese people).
In the speech recognition apparatus for the above, it is practically impossible to prepare standard patterns for each word because the number is too large. Therefore, standard patterns are prepared in units of syllables and phonemes, and standard patterns of words are created by concatenating and combining these, or results of recognizing in syllable units are post-processed to obtain word recognition results and sentence recognition results. To get it. In Japanese, there are about 120 syllables and about 4 phonemes.
Since it is 0, all Japanese can be supported by preparing a relatively small number of standard patterns. That is, if a small voice unit is used, a small number of standard patterns can cover a wide range of recognition targets, and if a large voice unit is used, a large number of standard patterns are required.

Next, let us consider the speech pattern expression capability of the standard pattern. As described above, Japanese speech can be represented in principle by about 120 syllables, but even in the same syllable, the speech pattern greatly varies depending on the preceding syllable and the following syllable. In addition, even if the syllable is the same, the voice pattern is greatly changed depending on the speaker. In the standard pattern of syllable unit or phonological unit, since the voice pattern that changes depending on the environment before and after is represented by the only standard pattern, the expression ability is naturally limited. On the other hand, when a standard pattern is registered in a large unit such as a word, variations in syllable and phoneme patterns based on the surrounding environment are incorporated into the word standard pattern, and as a result, the standard pattern faithfully expresses the speech phenomenon. That is, it is desirable to use a large voice unit (for example, a word) as the standard pattern from the viewpoint of the ability to express a voice phenomenon (especially, a change in voice pattern based on the surrounding environment).

Consider next from the viewpoint of the amount of available learning voice data. As described above, when a small voice unit is used, the number of standard patterns can be kept small, and when a large unit is used, the total number of standard patterns becomes large. Therefore, assuming that the total amount of learning voice data that can be used for learning the standard pattern is constant, the use of small voice units increases the amount of learning voice data for each standard pattern, resulting in a highly reliable standard pattern. Can be created. If the reliability is considered to be the same, the required amount of learning voice data is larger when the standard pattern of a larger unit is used.

Under the above circumstances, it is common sense to use a standard pattern in units of words when the number of vocabularies to be recognized is small, and to use a standard pattern in units of syllables or phonemes or similar when the number of vocabularies to be recognized is large.

Among these conventional techniques, as an improvement of these techniques, a method of using different standard patterns depending on the front-back environment, and a clustering method for increasing reliability when creating different standard patterns depending on the front-back environment are used. There is a technique using the technique. The conventional technique will be briefly described below.

As a method of using different standard patterns depending on the surrounding environment, Kluwer Academic Publishers, Nor
wel, MA, 1989 “Automatic Speech Recognition”,
There are examples as described on pages 95-97. This example is for English, and the basic unit of recognition is the English phoneme (Phone).
Is adopted to have different phoneme standard patterns depending on the phonemes before and after. By preparing the standard pattern in this way, it is possible to reduce erroneous recognition due to changes in the voice pattern based on the surrounding environment. There are about 40 phonemes in English, and if different standard patterns are used depending on the surrounding environment, the total number will be large, exceeding several thousand. In order to reliably learn such an enormous number of standard patterns, an enormous amount of learning speech samples are required, which is not realistic. In this conventional example, in order to deal with this problem, similar front and rear environments are handled collectively, and the total number of standard patterns is suppressed.

As a similar conventional example in the case of Japanese, the Institute of Electronics, Information and Communication Engineers, D2, Vol.J76-D-2, No.10, PP.215.
5-2164, October 1993, "Automatic Generation of Hidden Markov Networks by Sequential State Division Method" is an example. In this example, Japanese phonemes (consonants and vowels) are the basic units of recognition,
Starting from an HMM that does not depend on the phonetic environment before and after, the state is divided sequentially to deal with changes due to the phonemic environment before and after. The state division corresponds to a plurality of models, and is similar to providing different models depending on the preceding and following phoneme environments. In this conventional example, the decision as to what phonological environment the model should be divided into is automatically determined from the distribution of learning samples.

[0010]

In the above-mentioned prior art, by providing different standard patterns depending on the preceding and following phonological environments, it is possible to prepare a standard pattern that appropriately expresses the variation of the speech pattern based on the preceding and following phonological environments. , The recognition accuracy is improved. In addition, the clustering is used to solve the problem that the number of standard patterns increases by considering the phonological environment before and after, and the number of training speech samples for each standard pattern decreases, and the reliability of each standard pattern decreases. By using the above technique, the number of standard patterns is reduced to prevent the deterioration of reliability.

However, in order to reduce the occurrence of erroneous recognition in speech recognition, not only the accuracy and reliability of individual standard patterns are improved, but also the standard patterns of opposite categories have the same accuracy and reliability. Therefore, it is important to ensure that the set of standard patterns has an even accuracy and reliability as a whole. Even if the accuracy and reliability of some standard patterns are high, if the standard patterns with low accuracy or reliability are included in the set, the standard patterns have an adverse effect and cause misrecognition.

It is an object of the present invention to provide a means for creating an overall balanced set of standard patterns as a set of standard patterns which has been insufficiently considered in the above prior art.

[0013]

The object of the present invention is to measure the spread of the distribution of each standard pattern at each stage of learning the standard pattern so that each standard pattern has the same spread of distribution. For standard patterns with a large distribution spread, multiple standard patterns should be used.For each of the standardized patterns, the distribution spread should not be large, and each standard pattern should have the same distribution spread as a whole. It is achieved by

[0014]

In the standard learning process, the standard pattern is learned by using the conventional standard pattern learning method. For each standard pattern created, measure the spread of the distribution. In the process of determining the balance as a standard pattern set, the balance as a set of standard patterns is determined according to the measurement result of the spread of the distribution. When the spread of the distribution is biased, it is determined that the balance is poor, and when the spread of the distribution of each standard pattern is approximately the same and even, the balance is determined to be good. If it is determined that the balance is unbalanced, a standard pattern with a large distribution spread is selected from the standard pattern set, and a plurality of standard patterns are selected so that the distribution spread of each standard pattern does not become large. By the above processing, a well-balanced standard pattern can be created as a set of standard patterns, and highly accurate speech recognition can be realized.

[0015]

Embodiments of the present invention will be described below with reference to the drawings. The present invention can be applied to both word voice recognition and continuous voice recognition, but here, for simplicity, the case of word voice recognition will be described as an example.

FIG. 1 is a block diagram showing the configuration of an embodiment of the word voice recognition apparatus of the present invention. The input voice is converted into an electric signal by the voice input means 1. The voice converted into the electric signal is further analyzed by the voice analysis means 2 and the time series of the feature vector is output. On the other hand, the standard pattern connecting means 7 connects the standard patterns of the recognition basic units stored in the standard pattern storing means 5 in advance according to the information stored in the word dictionary 6 to form a word standard pattern. The standard pattern created by the standard pattern connection unit 7 and the feature vector time series of the input voice are collated by the collation unit 3, and a score is obtained for each word to be recognized. The judging means 4 outputs the recognition result based on the score of each word.

Next, the standard pattern of the basic recognition unit used in the present invention will be described. In the present invention, a probabilistic model is adopted as the standard pattern. FIG. 2 is a probabilistic model (Hidden Ma) which is a standard pattern used in the present invention.
rkov Model, hereinafter abbreviated as HMM). In the figure, each circle represents a state, and arrows represent transitions between the states. The symbol aij attached to the arrow represents the probability that the transition from the state i to the state j will occur, and the symbol bij (k) indicates that the feature vector belonging to the kth classification when the transition from the state i to the state j occurs. It represents the probability of being output. Given the feature vector time series of the input voice, the probability that the feature vector time series of the input voice is output from this probability model (HMM) can be calculated using the state transition probability and the output probability. The matching means 3 in FIG. 1 performs this probability calculation process. For details on the probability calculation process, see Kluw
er Academic Publishers, Norwel, MA, 1989 “Auto
The known method described in “matic Speech Recognition”, pages 95 to 97 may be used.

Next, a method of connecting standard patterns used in the speech recognition apparatus of the present invention will be described with reference to FIG.
FIG. 3 is a diagram for explaining how standard patterns are connected according to the word dictionary 6. As described above, the speech recognition apparatus of the present invention uses the HMM, which is a state transition model, as the standard pattern, so that the standard patterns can be easily connected. The connection of the standard patterns may be such that the destination of state transition from the final state of the preceding model is the first state of the subsequent model. In FIG. 3, the Japanese syllable is adopted as the basic unit of recognition, and the word “Hitachi (/ hitachi)” in the dictionary is taken up. The standard pattern storage means 5 stores HMMs corresponding to Japanese syllables (voice units corresponding to Japanese kana characters “a” “i” ... “n”). To create an HMM for the word "Hitachi", first the word dictionary 6 is checked and it is read that the word "Hitachi" is composed of syllable strings / hi /, / ta /, / chi /. The standard pattern connection means 7 sequentially reads the / hi / HMM, the / ta / HMM, and the / chi / HMM from the standard pattern storage means 5 in accordance with the syllable string, and sets them as one large HMM.

Next, a normal learning method of the HMM which is a standard pattern used in the speech recognition apparatus of the present invention will be described. The HMM is implemented by parameter estimation using a large number of training speech samples. FIG. 4 is a flowchart showing an outline of the learning flow. First
Create an initial model of HMM by some method (10
1) and then the parameter re-estimation process (102) using the learning voice sample is repeated (103) until the convergence condition is satisfied. This learning method is originally an iterative estimation algorithm, and the accuracy of the model improves as the number of iterations increases. Therefore, it is not always necessary to create the initial model with high accuracy. There are several methods for creating the initial model, but a method such as giving a random number may be used. The method of parameter re-estimation will be described later. There are several possible methods for determining the convergence condition, but for example, the number of repetitions is fixed and a fixed number of times (for example, 5 times).
There is no problem in practice by the method of ending after repeating.

When the convergence condition is satisfied, the iteration is terminated and the parameters of each HMM obtained by the parameter estimation are stored (104).

Next, the parameter re-estimation processing of the HMM will be described. As shown in the flowchart of FIG. 4, the parameter re-estimation process of the HMM is repeatedly performed in the learning flow. Here, the one-time processing will be described with reference to the flowchart of FIG. The HMM parameter re-estimation process is performed using speech samples for learning. Assuming that the number of training voice samples is N, the parameter estimation calculation processing similar to N times is performed, and after this processing is completed, the parameters of each HMM are updated to new values. In the parameter estimation process using each voice sample, first, the HMMs of the basic recognition units are concatenated according to the utterance content of the voice sample (203), and a method called the Forward-Backward algorithm is used for this concatenated HMM. Parameter estimation is performed (204).
By decomposing the concatenated HMMs into the original recognition basic units, the parameter estimation value of the HMM of each recognition basic unit is obtained (205). However, at this point, the HM of each recognition basic unit
The parameters of M are not updated, and after the parameter estimates are obtained for all speech samples, the parameter estimates of all recognition parameters obtained so far are combined to update the parameters of the HMM of each recognition basic unit (207). . For the specific calculation procedure of parameter estimation (Forward-Backward algorithm), see Kluwer Academic Publishers, Norwel,
MA, 1989 “Automatic Speech Recognition”, page 95-9
The known method described on page 7 may be used.

Next, a method of learning a standard pattern considering balance as a set of standard patterns, which is the main point of the present invention, will be described. FIG. 6 is a flow chart for explaining the learning method. In this learning, first, HMM learning is performed by a standard method that has been used conventionally (301).
The spread of the distribution is measured for each HMM thus created (302). The spread of the distribution can be obtained by the following equation, for example.

[0023]

[Equation 1]

However,

[0025]

[Equation 2]

As the spread of the distribution, the sum of the variance VARk of each dimension of the feature vector was obtained as shown in (Equation 1). The dispersion formula is calculated by (Equation 2). Here, Vk is the value of the k-th dimension of the feature vector V, μ
k is the average value.

Next, it is judged whether there is a large variation in the extent of the distribution of each HMM and there is a balance as a set (303), and if it is judged that there is a balance, the process ends. The balance can be determined by a determination method such as whether the maximum value of the spread of the distribution of each HMM is within a threshold value. The threshold can be set by, for example, a method of multiplying the average value of the spread (dispersion) of the distribution of each HMM by a constant α (for example, 1.3 times).
This determination is specifically expressed by the following equation.

[0028]

(Equation 3)

If this condition is not satisfied, the HMM having the largest spread of distribution is obtained (304), and this is divided into a plurality (for example, two) (305). Next, multiple HM
A speech sample containing syllables or phonemes for M is selected from the learning-like speech samples, and recognition processing is performed using each of the HMMs that have been divided into a plurality of HMMs. Allocate (306). Using the voice samples divided into two by the above allocation, the parameters of each HMM that has been made multiple are estimated and the parameters of the divided HMMs are updated (3
07). After the above processing, the spread of the distribution is measured again for each HMM (302), and it is determined whether there is no large variation in the spread of the distribution of each HMM and it is balanced as a set (303). repeat. If it is determined that there is a balance, the process ends.

For a plurality of HMMMs, a plurality of appropriate numbers of HMMs are created for the original HMM. Here, 2
In order to divide the HMM into two, for example, to divide the output probability distribution of the HMM into a new first distribution with a minute change, add a minute distribution from the original distribution. A method is conceivable in which the distribution obtained by subtracting the large fluctuation is used as the new second distribution. As the first distribution, for example, within a certain range from the value of the feature vector that takes the maximum value of the output probability density, the value of the probability density is multiplied by a constant β (for example (1.01 times)). Then, the probability density is also multiplied by a constant γ so that the value of the probability density becomes smaller.The value of γ is smaller than 1. The increase of the probability density by β times and the decrease of the probability density by γ times cancel each other out. As a result, the first distribution is obtained, and the second distribution has a probability density value of a constant 2-β within a certain range from the value of the feature vector taking the maximum value of the output probability density. Times (0.9 when β is 1.01
9) Then, within other ranges, the probability density is multiplied by a constant 2-γ so that the value of the probability density becomes small. FIG. 7 shows how a new distribution is obtained by this method. 40 in FIG.
1 indicates the original distribution. The range from 404 to 405 indicates a certain range from the value of the feature vector having the maximum output probability density. Reference numeral 402 represents a new first distribution, and 403 represents a new second distribution.

As a method of measuring the spread of the HMM distribution, a method of detecting multimodality of the distribution shape of the output probability distribution of the HMM can be considered. In addition, as a method of HMM division, the distribution of each detected
A method of dividing the distribution of M is also conceivable.

[0032]

As described above, according to the present invention, since a well-balanced standard pattern can be created as a set of standard patterns, highly accurate speech recognition is possible.

[0033]

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of a voice recognition device of the present invention.

FIG. 2 is a diagram illustrating a hidden Markov model of a basic recognition unit used in the speech recognition apparatus of the present invention.

FIG. 3 is a diagram for explaining how to connect hidden Markov models of basic recognition units used in the speech recognition apparatus of the present invention according to a word dictionary.

FIG. 4 is a flowchart illustrating a standard pattern learning method of the present invention.

FIG. 5 is a flowchart illustrating a parameter estimation process in the standard pattern learning method of the present invention.

FIG. 6 is a flowchart illustrating another learning flow of the standard pattern learning method of the present invention.

FIG. 7 is a diagram illustrating a method of dividing a hidden Markov model performed in the standard pattern learning method of the present invention.

[Explanation of symbols]

1 ... voice input means, 2 ... voice analysis means, 3 ...
Collation means, 4 determination means 5 standard pattern storage means, 6 word dictionary, 7 standard pattern connection means 101 initial model creation processing, 102 parameter re-estimation Processing 204 ... Forward-Backward algorithm 302 ... Standard pattern distribution spread measurement processing 303 ... Balance determination processing as a set of standard patterns 305 ... Standard pattern multiple processing

Claims (9)

[Claims] 1. A voice input means for inputting a voice, a voice analysis means for analyzing the input voice and outputting a time series of feature vectors, and a standard for a syllable, a phoneme, or a voice basic unit smaller than a phoneme. A standard pattern storage means for storing patterns, a word dictionary that describes the words of the recognition target word as a syllable, or a phonological unit, or a sequence of phonetic basic units smaller than a phonological unit, and a syllable, a phonological unit, or a smaller than phonological unit. Standard pattern connecting means for connecting standard patterns for basic speech units to form a standard pattern for a recognition target word, and collating means for matching the time series of the feature vector of the input voice with the standard pattern formed by the connection. In the voice recognition device for recognizing based on the matching result output from the matching means, the syllable or the phoneme, Or, a standard pattern for a phonetic unit smaller than a phoneme, the syllable, or a phoneme, or to provide a plurality according to the degree of variation of the phonetic pattern of a phonetic unit smaller than a phoneme, a syllable having a plurality of standard patterns, Alternatively, a phonetic recognition unit or a phonetic basic unit smaller than a phoneme is matched with a plurality of standard patterns in the matching process, and a plurality of matching results are integrated to obtain a recognition result. 2. The speech recognition apparatus according to claim 1, wherein the standard pattern is composed of a probabilistic model, and the probabilistic model is learned by using a speech sample for learning. 3. The speech recognition apparatus according to claim 2, wherein the probabilistic model is a hidden Markov model. 4. A standard for representing the entire utterance content by concatenating standard patterns of syllables, phonemes, or basic units of speech smaller than the phoneme according to the utterance content to a learning voice sample whose utterance content is known. Create a pattern, learn the standard pattern as a standard pattern that represents the entire utterance content using the learning voice sample, and use the created standard pattern that represents the entire utterance content as a syllable, a phoneme, or a voice that is smaller than the phoneme. A learning method of a standard pattern such that a standard pattern of a syllable, a phoneme, or a voice basic unit smaller than a phoneme is decomposed into basic units, the number of which is equal to the number of learning voice samples given the learning process. In a method of learning a standard pattern that is repeated, a standard pattern of created syllables, phonemes, or basic units of speech smaller than phonemes. The degree of the spread of the distribution of the voice pattern of is measured, the standard pattern with a wide distribution is diversified, and the standardized pattern is trained again by using the speech sample for learning. A standard pattern learning method characterized by. 5. The standard pattern is constructed by a stochastic model, and the degree of spread of the distribution of speech patterns with respect to the standard pattern is measured based on the variance value in the distribution information of the probabilistic model. The standard pattern learning method according to claim 4, wherein 6. The standard pattern is constructed by a probabilistic model, and the degree of spread of the distribution of the speech pattern with respect to the standard pattern is measured based on the detection of the bimodality of the distribution possessed by the probabilistic model. The standard pattern learning method according to claim 4. 7. The standard pattern is constructed by a probabilistic model, and the plurality of standard patterns are made by adding a minute change to the distribution information of the original standard pattern. A learning method of the standard pattern according to claim 5 or 6. 8. The standard pattern is constructed by a probabilistic model, and the standard pattern is made plural by detecting bimodality with respect to the distribution information of the original standard pattern, and components from the detected plural distributions are detected. 7. A standard pattern learning method according to claim 4, 5 or 6, wherein a distribution having a large value of is selected and the distribution is performed based on the selected distribution. 9. The standard pattern learning method according to claim 5, wherein the probabilistic model is a hidden Markov model.