JPH11212592A - Pattern recognition device and standard pattern generating method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a standard pattern which has a high recognition degree by measuring how much temporarily generated probability distributions by classes overlap with each other, and correcting distributions which overlap with each other and reducing their overlap. SOLUTION: As for the completed probability distributions of the respective classes, how much the classes overlap with each other is measured (302). Denoting the number of the classes to be recognized as N, (N2-N)/2 pairs are present, so overlaps of the distributions are measured as to all the pairs. When a measured value exceeds a previously set threshold value, it is considered that the distributions overlap large and the number of distributions exceeding the threshold value is counted and denoted as K (303). Then it is decided whether or not the count value K is larger than the previously set threshold value Kth (304) and when the K is larger than Kth, the process is finished, but when not, those probability distributions are selected as distributions to be corrected (305) and corrected (306).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a pattern recognition device for identifying patterns of characters, figures, voices and the like.

[0002]

2. Description of the Related Art In a pattern recognition apparatus according to a statistical decision theory, a probability distribution of a feature vector is obtained in advance for each class, a probability value is calculated for given input data according to the probability distribution, and a maximum value is calculated. The class that gives the probability is the recognition result. Such a pattern recognition device according to the statistical decision theory is disclosed in various documents, for example, published by Corona, Makoto Nagao, edited by the Institute of Electronics and Communication Engineers, University of IEICE series 1-4.
This is described in “Pattern Information Processing”, pp. 106-109. A normal distribution is often used as the probability distribution. The normal distribution determines the distribution shape by giving the mean and the variance (or standard deviation). Probability distribution is determined by collecting sample data for each class and calculating the average and variance of the data. In a pattern recognition device that uses a normal distribution as a probability distribution, a normal distribution is created for each class using sample data in advance, and the probability is calculated for each class using the feature vector of the input data during recognition. Then, the class that gives the maximum probability is the recognition result.

As a conventional example of a speech recognition apparatus, there is a recognition apparatus which collects sample data for each class in advance and uses a standard pattern of each class by expressing the probability distribution of a feature vector. In the case of speech recognition, the processing target is not a feature vector at a temporary point but a time series of feature vectors, and a technique called a hidden Markov model is widely used as a technique for handling this. The speech recognition using the hidden Markov model is described in, for example, Seiichi Nakagawa, edited by the Institute of Electronics, Information and Communication Engineers, “Speech Recognition by Stochastic Model”, pp. 29-89. The Hidden Markov Model is a state transition model, and handles time series in the framework of state transition. When the hidden Markov model is used, the standard pattern is not a probability distribution of a feature vector at a temporary point, but a state transition model in which a probability distribution is associated with each state (or state transition) of the state transition model. . In the case of speech recognition, processing for associating the time series of the feature vector of the input speech to be recognized with each state of the state transition model is newly added, but the basic processing concept is a pattern recognition apparatus based on the statistical decision theory. Is the same as
In speech recognition, the probability distribution of a feature vector of speech is basically used as a standard pattern, and a normal distribution is often used as the probability distribution. In the case of speech recognition for unspecified speakers, a mixed distribution that expresses the probability distribution as a linear sum of multiple normal distributions instead of a single normal distribution in order to respond to the variation of the feature vector for each speaker In many cases, this is supported.

[0004]

In the above prior art, the standard pattern of each class is prepared as a single normal distribution or as a linear sum of a predetermined number of normal distributions for each class. .

[0005] In order to reduce the occurrence of erroneous recognition in pattern recognition or speech recognition, it is necessary not only to improve the expression accuracy and reliability of the probability distribution of each class, but also to reduce the overlap of the distribution between opposing classes. Is important.

An object of the present invention is to provide a standard pattern with a high recognition accuracy by modifying a probability distribution based on "reduction of overlap between distributions of opposing classes", which was not sufficiently considered in the prior art. It is in.

[0007]

SUMMARY OF THE INVENTION The object of the present invention is to measure the degree of overlap between probability distributions once created for each class, and to correct the overlap for distributions with overlapping distributions. Achieved by reducing.

[0008]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. The present invention can be applied to pattern recognition of various patterns such as characters, figures, and voices. Here, the case of voice recognition will be described as an example.

FIG. 1 is a block diagram showing the configuration of an embodiment of the word speech recognition apparatus to which the present invention is applied. The input voice is converted into an electric signal by the voice input unit 1.
The voice converted into the electric signal is further analyzed by the voice analysis means 2, and a time series of feature vectors is output. On the other hand, the standard pattern connection means 7 includes the standard pattern storage means 5.
Are linked in accordance with 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 determining means 4 outputs a 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. The recognition basic unit refers to a syllable that is a component of a word voice, a phoneme (vowel, consonant) that is a smaller unit, and the like. If a standard pattern is prepared for each basic recognition unit, a standard pattern of an arbitrary word voice can be constructed by combining the standard patterns, and speech recognition of a large vocabulary can be performed with a limited number of standard patterns. . In the case of speech recognition, various units such as a syllable, a phoneme, and a phoneme chain can be considered as a basic unit of recognition. However, in this embodiment, a case where a syllable is used as a basic unit for recognition will be described for simplicity. FIG. 2 shows a stochastic model (HiddenMarkov Model, corresponding to the standard pattern used in the present invention).
FIG. 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 a transition from the state i to the state j occurs. The symbol bij (k) indicates that the feature vector belonging to the k-th class when the transition from the state i to the state j occurs. Indicates the output probability. Given the input speech feature vector time series, the probability that the input speech feature vector time series is output from this probability model (HMM) can be calculated using the state transition probability and the output probability. In the matching means 3 in FIG. 1, the processing of this probability calculation is performed. For details on the probability calculation process, see Kluwer Academic
Publishers, Norwel, MA, 1989 “Automatic Speech Re
cognition ", 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 linked according to the word dictionary 6. As described above, in the speech recognition apparatus of the present invention, the HMM which is the state transition model is used as the standard pattern, so that the standard pattern can be easily connected. The connection of the standard patterns may be such that the state transition destination from the final state of the preceding model is set to the first state of the following model. In FIG. 3, Japanese syllables are 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. To create an HMM for the word "Hitachi", first, the word dictionary 6 is checked to read out that the word "Hitachi" is composed of the syllable strings / hi /, / ta /, and / chi /. The standard pattern connection means 7 sequentially reads the / hi / HM from the standard pattern storage means 5 according to the syllable string.
M, / ta / HMM, and / chi / HMM are read out and concatenated 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. HMM is implemented by performing 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) 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, the initial model does not always need to be created with high accuracy. There are several methods for creating the initial model. For example, a method of giving random numbers may be used. The method of parameter re-estimation will be described later. There are several methods for determining the convergence condition. For example, the number of repetitions is fixed and the number of repetitions is fixed (for example, five).
There is no practical problem in the method of terminating after repeating the above.

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

Next, the parameter re-estimation process 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, one process 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 speech samples for learning is N, a similar parameter estimation calculation process is performed N times, and after this, the parameters of each HMM are updated to new values. In the parameter estimation process using each voice sample, first, the HMMs of the recognition basic units are connected in accordance with the utterance content of the voice sample (203), and the connected HMM is forward-backwarded (forward).
-Backward) Parameter estimation is performed using a technique called an algorithm (204). By decomposing the concatenated HMM into the original recognition basic units, the HM of each recognition basic unit
A parameter estimate of M is obtained (205). However, at this time, the HMM parameters of each recognition basic unit are not updated, and after the parameter estimation values are obtained for all voice samples, Update HMM parameters (2
07). The specific calculation procedure for parameter estimation (Forward-Backward algorithm) is described in Kluwer Acade
mic Publishers, Norwel, MA, 1989 “Automatic Speec
h Recognition ", pages 95-97, may be used.

Next, a method of correcting a probability distribution constituting a standard pattern, which is the main point of the present invention, will be described. FIG. 6 is a flowchart illustrating the overall flow of the probability distribution correction method.

In this learning, first, a probability distribution is created by a conventional standard method (301). With respect to the probability distribution of each class thus completed, the degree of overlap between the classes is measured (302). The method of measuring the degree of overlap will be described later. The measurement of the distribution overlap is performed for all pairs of classes to be recognized.
Assuming that the number of classes to be recognized is N, (N2-N) /
Since there are two pairs of combinations, the overlap of the distribution is measured for all the combinations. And
When the measured value exceeds a preset threshold value, it is assumed that the overlap between distributions is large, the number of distributions exceeding the threshold value is counted, and the value is set to K (303). Next, it is determined whether the count value K is greater than a preset threshold value Kth (304), and if it is smaller than Kth, the process is terminated. Kth
If it is larger, those probability distributions are selected as probability distributions to be modified (305), and their distributions are modified (306). After the correction of the selected probability distribution is completed, the degree of overlap between the probability distributions between the classes is measured again (30).
2) The process of counting the number of probability distributions having a large overlap between classes (303) and determining whether the number is greater than a set threshold Kth (304) is repeated. The number of probability distributions with large overlap between classes is the threshold K
When it becomes less than th, the process ends.

Next, a method of measuring the degree of overlap of the probability distributions performed in the processing step 302 will be described. The probability distribution handled here is a normal distribution. The shape of the normal distribution is determined by the mean and the variance (standard deviation). Therefore, the overlap between normal distributions can be calculated using their respective averages and variances. FIG. 7 is a diagram illustrating a method of calculating the degree of overlap of a one-dimensional normal distribution. FIG. 7 shows the distribution of the two classes A and B. μA and μB are the average values of classes A and B, respectively, and σA and σB are the classes A, respectively.
And the standard deviation of B. At this time, the degree of overlap between the distributions of class A and class B can be expressed by equation (1).

[0018]

(Equation 1)

That is, the closer the average value between the two distributions is, and the larger the variance (standard deviation) is, the greater the degree of overlap is. FIG. 7 and equation (1) show a one-dimensional case, but actual speech recognition deals with a multi-dimensional probability distribution. In the case of a multidimensional probability distribution, the variance is basically treated as a covariance matrix considering the correlation between dimensions, but in order to reduce the parameters of the distribution, it is possible to treat the components other than the diagonal components of the covariance matrix as zero. Many. Also in this embodiment, a covariance matrix having only diagonal components will be considered. Let μAi and μBi be the ith element of the mean value vector of classes A and B, and σAi and σBi be the ith row and ith column components of the covariance matrices of classes A and B, respectively. The degree of overlap of the distribution can be expressed by equation (2).

[0020]

(Equation 2)

When two probability distributions with the next largest overlap are given, the process of step 305 is to select which distribution is to be modified. In this embodiment, each of the distributions represented by the equation (3) is selected. A method is adopted in which the distribution of the modulus having a large sum of the variances of the dimensions is selected as a correction target.

[0022]

(Equation 3)

Next, the process of correcting the probability distribution in step 306 will be described. In the present embodiment, the probability distribution is given by a normal distribution. Correction of the probability distribution is performed by expressing a single normal distribution before correction by a linear sum of two normal distributions. FIG. 8 is a view for explaining the probability distribution correction processing. In FIG. 8, 401 and 402 are overlapping distributions. 403 is the center of the distribution 401, and 404 is the center of the distribution 402. In the case of FIG. 8, the distribution 401 is modified because the distribution 401 is larger. The sample data used to create the distribution 401 is indicated by an x in FIG. The correction is performed in the following procedure. First, the distribution 402
Of the sample data farthest from the center 404 of. In FIG. 8, 405 corresponds to this. Next, sample data far from the center 404 of the distribution 402 and located on the opposite side of the sample data 405 from the center point 403 of the distribution 401 is obtained. In FIG. 8, 406 corresponds to this. Next, for all of the sample data used to create the distribution 401, it is determined which of 405 and 406 is closer to the point, and all the sample data is classified into two. Next, an average value and a variance are calculated using the classified sample data, and a normal distribution is created for each. In FIG. 8, 407 and 408 are newly created normal distributions. FIG. 9 shows a flow of correcting the probability distribution.

Next, another embodiment of the method of correcting the probability distribution forming the standard pattern, which is the main point of the present invention, will be described. The method described with reference to FIGS. 6 to 9 is a method for reducing the overlap between the probability distributions constituting the standard pattern. What is explained here is to actually perform recognition, extract an overlap of distributions that may cause erroneous recognition, and reduce the overlap of distributions with respect to the overlap of distributions in this portion. FIG. 10 is a flowchart illustrating the overall flow of another example of realizing the probability distribution correction method.

First, a probability distribution is created by a conventional standard method (601). Recognition of the sample data for evaluation is performed using the probability distribution of each class thus completed (602). Next, regarding the data in which misrecognition has occurred, the number of distribution overlaps is counted assuming that there is an overlap between the probability distribution of the class of the input data and the probability distribution of the class of the misrecognition result (603). Next, it is determined whether or not the count value K is greater than a preset threshold value Kth (604), and if it is smaller than Kth, the process is terminated. If it is larger than Kth, those probability distributions are selected as probability distributions to be corrected (605), and those distributions are corrected (606). After the correction of the selected probability distribution is completed, the sample data for evaluation is recognized again (60).
2) For the data in which the misrecognition has occurred, the number of distribution overlaps is counted assuming that there is a distribution overlap between the probability distribution of the class of the input data and the probability distribution of the class of the misrecognition result (603). The process of determining whether the number is greater than the set threshold value Kth (604) is repeated. When the number of overlapping probability distributions becomes equal to or smaller than the threshold value Kth, the process ends. Processing steps 604, 60
The processing in steps 5 and 606 is the same as the processing step 3 in FIG.
04, 305, and 306, but slightly different. Although there is an overlap between the probability distribution of the class of the input data and the probability distribution of the class of the recognition result, processing is performed based on the idea that it is necessary to lower the probability value of the class of the recognition result to avoid erroneous recognition. In step 605, the probability distribution of the class of the recognition result is selected as a correction target. The processing of correcting the probability distribution in the processing step 606 is almost the same as the processing step 306. In the processing step 306, information on the center point p of the distribution of the opposing class is used as shown in the flowchart of FIG. In step 606, the center point of the erroneously recognized input data sample set is used instead of the center point p of the distribution of the opposing class.

[0026]

As described above, according to the present invention, it is possible to reduce the overlap between the recognition target classes of the probability distribution, which is a component of the standard pattern, so that highly accurate pattern recognition and speech recognition become possible.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of a word speech recognition apparatus to which the present invention is applied.

FIG. 2 is a diagram illustrating a hidden Markov model of a basic recognition unit used in the word speech recognition apparatus to which the present invention has been applied.

FIG. 3 is a view for explaining a state in which a hidden Markov model of a basic recognition unit used in the word speech recognition apparatus to which the present invention is applied is connected in accordance with a word dictionary.

FIG. 4 is a flowchart illustrating a method for learning a standard pattern according to the present invention.

FIG. 5 is a flowchart for explaining parameter estimation processing in the standard pattern learning method of the present invention.

FIG. 6 is a flowchart illustrating an overall flow of a probability distribution correction method according to the present invention.

FIG. 7 is a view for explaining a method of measuring the degree of overlap of the probability distributions of the basic recognition units used in the word speech recognition apparatus to which the present invention is applied.

FIG. 8 is a diagram illustrating a method for correcting a probability distribution according to the present invention.

FIG. 9 is a flowchart illustrating a probability distribution correction procedure according to the present invention.

FIG. 10 is a flowchart illustrating an overall flow of another probability distribution correction method according to the present invention.

[Explanation of symbols]

1 ... voice input means, 2 ... voice analysis means, 3 ...
-Matching means, 4-Judgment means 5-Standard pattern storage means, 6-Word dictionary, 7-Standard pattern concatenation means, 101-Initial model creation processing, 102-Parameter rewriting Estimation process, 204: Forward-Backward algorithm, 302: Overlap measurement process of distribution between each class, 306: Correction process of probability distribution, 401 ...
Probability distribution to be corrected, 402: Probability distribution of opposition class, 403: Center point of probability distribution to be corrected, 404
················································································ ············································································································································ Is the sample data 405 farthest from the center point of the probability distribution center point 403 of the probability distribution to be corrected.
407, 408 on the other side for
... New probability distribution created by correction.

Claims (10)

[Claims] A feature extraction means for analyzing input data of a recognition target to extract a feature vector; a standard pattern storage means for storing a probability distribution of a feature vector for each class of the recognition target as a standard pattern; Matching means for calculating a probability using a probability distribution which is a standard pattern for each class with respect to the feature vector of the input data, and performing recognition based on a probability value for each class output from the matching means. In the pattern recognition device that performs the above, the probability distribution of the feature vector that is the standard pattern of each class is corrected according to the degree of overlap between the classes in the probability distribution of the feature vector that is the standard pattern of each class. A pattern recognition device characterized in that: 2. The pattern recognition according to claim 1, wherein the correction of the probability distribution of the feature vector, which is a standard pattern for each class, is performed using a linear sum of a plurality of probability distributions. apparatus. 3. The method according to claim 1, wherein the probability distribution of the feature vector, which is a standard pattern of each class, is corrected by using input data of each class used to create the probability distribution of a standard pattern of each class. 3. The pattern recognition device according to claim 1, wherein the pattern recognition device corrects the pattern. 4. The modification of the probability distribution of the feature vector, which is a standard pattern for each class, differs from the input data for each class used to create the probability distribution, which is a standard pattern for each class. 3. The pattern recognition apparatus according to claim 1, wherein the pattern is corrected using input data. 5. The method according to claim 1, wherein the probability distribution of the feature vector, which is a standard pattern for each class, is corrected only when the result of pattern recognition is erroneous recognition and only for the probability distribution that caused the recognition. Claim 1 or Claim 2 or Claim 3 or Claim 4 wherein the processing is performed.
The described pattern recognition device. 6. A feature extracting means for analyzing an input voice at regular time intervals to extract a feature vector time series, and a standard for storing a time series of a probability distribution of a feature vector for each class to be recognized as a standard pattern. Pattern storing means, and matching means for calculating a cumulative probability using a probability distribution time series which is a standard pattern for each class with respect to the feature vector of the input voice, and each class output from the matching means. In the speech recognition apparatus that performs recognition based on the value of the cumulative probability for each class, the standard for each class is determined according to the degree of overlap between the classes in the probability distribution of the feature vector that is a component of the standard pattern for each class. A speech recognition apparatus, wherein a probability distribution of a feature vector which is a component of a pattern is corrected. 7. The method according to claim 6, wherein the probability distribution of the feature vector, which is a component of the standard pattern for each class, is modified using a linear sum of a plurality of probability distributions. Voice recognition device. 8. The method according to claim 1, wherein the correction of the probability distribution of the feature vector, which is a component of the standard pattern for each class, is performed for each class used to generate the probability distribution of the component of the standard pattern for each class. 8. The speech recognition device according to claim 6, wherein the correction is performed using the input speech. 9. The method of correcting the probability distribution of a feature vector, which is a component of a standard pattern of each class, comprises the steps of: modifying a probability distribution of a component of a standard pattern of each class; 8. The speech recognition apparatus according to claim 6, wherein the correction is performed by using an input speech different from the input speech of the speech recognition device. 10. The correction of the probability distribution of the feature vector, which is a component of the standard pattern for each class, is performed only when the result of speech recognition is incorrect recognition. 10. The speech recognition device according to claim 6, wherein the speech recognition is performed only on the speech recognition device.