JPS62245294A - Voice recognition system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition method suitable for use in technical fields such as speech recognition devices, pattern matching, and signal coding.

Regarding a speech recognition method based on a vector quantization method that converts both an unknown input speech pattern and a standard speech pattern into a code string of representative vectors, for example,
No. 99500 is mentioned, but in the invention described in the publication.

Although efforts have been made to create a tree structure of distance measures and representative vectors, there is no mention of application to speech recognition methods for unspecified speakers. In recognition, efficient system configurations such as dictionary construction methods, structuring of representative vectors, distance table creation methods, and vector quantization methods for input speech are required.

■--The present invention was made in view of the above-mentioned circumstances.
In particular, after expressing the audio signal as a series of feature vectors,
This was done with the aim of improving recognition processing speed and recognition accuracy in a speech recognition method that uses a vector quantization method that recognizes unknown input cover patterns by converting them into code sequences of representative vectors.

SUMMARY In order to achieve the above object, the present invention includes an audio input section that inputs an audio signal, a feature analysis section that analyzes the input audio and converts it into feature parameters, and a vector quantum A vector quantization section that converts the data into a code string, a codebook storage section that stores representative code vectors necessary for vector quantization, and a codebook storage section that stores standard patterns to be recognized as code strings. a recognition processing unit that matches input speech and standard patterns; and a distance table storage unit that uses distance calculations between patterns, which are necessary during recognition processing, as distances between codes. , and a terminal for outputting recognition results, and is characterized in that the code vectors are clustered, distances are calculated using Mahalanobis' general distance between the code vectors belonging to each cluster, and the result is stored in a table. be. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is an electrical block diagram for explaining one embodiment of the present invention. In the figure, 1 is an audio input terminal, 2 is a feature vector extraction section, and 3 is a vector quantization section.

4 is a codebook storage unit, 5 is a recognition processing unit, 6 is a standard pattern storage unit, 7 is a distance table, and 8 is a recognition result storage unit.The audio input from the audio input terminal 1 is processed by the feature vector extraction unit 2. It is converted into a feature vector representing the characteristics of the voice. This feature vector is further stored in the codebook storage unit 4.
is converted to one of the representative vectors stored in
The code of the corresponding representative vector is registered in the vector quantization unit 3. The recognition processing unit 5 performs pattern matching between the code sequence registered in the vector quantization unit 3 and the standard pattern code sequence registered and stored in the standard pattern storage unit 6 in advance. The distance can be obtained quickly by quoting the distance calculated in advance and stored in the distance table 7 from the combination of codes. The recognition processing unit 5 outputs the standard pattern name having the minimum distance as a recognition result to the recognition result storage unit 8 through processing such as DP matching (pattern matching using dynamic programming).

FIG. 2 is a diagram showing the structure of the codebook 4 according to the present invention. In FIG. vector), 41b,
42b, 43b, and 44b represent the average vector of each cluster.For the method of creating the code vector, for example,
Y., Linda et al., “An Algor.
ithm for Vector Quantizer
Design”, IEEE, Trans, C
omg*un, Vol Cow-28°No, 1
．． Jan, 1980. It is described in. To put it simply, we use feature vectors that include various phoneme and syllable patterns as training samples, perform repeated calculations to determine clusters, and obtain representative vectors so that the overall average distortion is minimized. It is something. In this way, the obtained representative vectors 41a, 42a, 43a,
FIG. 2 shows a result of clustering using the same algorithm using 44a and the like as learning samples.

FIG. 3 is a diagram showing a distance table between code vectors existing in each cluster 41 to 44 in FIG. 2, and 71 to 74 correspond to 41 to 44 in FIG. 2. Further, 71a to 74a are distance tables of portions that do not belong to the distance tables 71 to 74 within the cluster. In 71 to 74, when storing the calculated distances between code vectors within each cluster, instead of using the usual Euclidean distance or urban distance, Mahalanobis' general distance or distance based on Bayes' law is used. In particular, it is possible to construct a distance table suitable for a speaker-independent recognition system. Furthermore, if normal Euclidean distances and urban area distances are stored in 71a to 74a, the distance between code vectors between clusters in FIG. 2 can be easily expressed. Alternatively, by storing representative distances (for example, the average value of the code vector distances between the clusters) in 71a to 74a, it is possible to reduce the memory required for this portion.

However, i and j in FIG. 3 are code numbers, N is the quantization level (number of code vectors), dl(it j)t...

d'(i, J) represents a distance table within the clusters 41 to 44, respectively.

FIG. 4 shows that when the vector quantization unit 3 of FIG. 1 vector quantizes the input speech and converts it into a code sequence,
This shows an example of tree-structuring of code vectors in consideration of consistency with code-vector clustering in Figure 2. In the figure, the depth of the tree structure is two levels, and each node (node)
The number is set approximately equal to the square root F1 of the quantization level N.6For example, when N=512,
It is 3. Therefore, the most efficient encoding can be achieved by setting both the number of clusters and the average number of code vectors within a cluster to about 23 in FIG. 2.

As is clear from the above description, according to the present invention, by clustering code vectors representing representative vectors among voice feature vectors, the distance table between code vectors can also be divided. As a result, it is possible to reduce the size of the table and the amount of memory. Also, by making the distance measure between code vectors within a cluster a statistical distance measure (Mahalanobis general distance, distance based on Bayes' law).

A recognition system suitable for unspecified speakers can be configured, and recognition accuracy can be improved. By creating a tree structure of code vectors based on clustering of code vectors, it is also possible to code input speech at high speed.

[Brief explanation of drawings]

FIG. 1 is an electrical block diagram for explaining one embodiment of the present invention, FIG. 2 is a diagram showing the structure of a codebook,
FIG. 3 is a diagram showing a distance table between code vectors,
FIG. 4 is a diagram showing a tree structure of code vectors. 1...Audio input terminal, 2...Feature vector extraction unit. 3... Vector quantization section, 4... Codebook storage section, 5... Recognition processing section, 6... Standard pattern storage section, 7... Distance table, 8... Recognition result storage section . Patent applicant Ricoh Co., Ltd. 1 Figure 2 Figure 3 Figure 4 KIIIJtg

Claims (5)

[Claims] (1) an audio input unit that inputs an audio signal, a feature analysis unit that analyzes the input audio and converts it into feature parameters, and a vector quantization unit that vector quantizes the feature parameters and converts them into a code string; A codebook storage section that stores representative code vectors required for vector quantization, a standard pattern storage section that stores standard patterns to be recognized as code strings, and input audio and standard pattern storage sections. It consists of a recognition processing unit that performs matching of the codes, a distance table storage unit that uses the distance calculation between patterns required during recognition processing as the distance between codes, and a terminal that outputs the recognition results. A speech recognition method characterized in that vectors are clustered, distances are calculated using Mahalanobis' general distance between code vectors belonging to each cluster, and the results are stored in a table. (2) The speech recognition method according to claim (1), wherein the distance calculation is performed using a distance measure based on Bayesian judgment. (3) In the distance table, the distance measure between each cluster is Euclidean distance or urban area distance, and the distance measure within each cluster is the statistic according to claim (1) or (2). The speech recognition method according to claim 1 or 2, characterized in that a distance measure is used as a distance measure. (4) In the distance table, the distance table is reduced by setting the distance between code vectors between each cluster to an appropriate constant value.
The speech recognition method described in section 1). (5) The number of clusters and the number of code vectors belonging to each cluster are the same and are equal to the square root of the total number of code vectors. Voice recognition method.