JPH04324499A - Speech recognition device - Google Patents

Abstract

PURPOSE:To offer the speech recognition device which can recognize the voice of an unknown user by selecting a proper standard speaker cluster. CONSTITUTION:The speech recognition device is equipped with an acoustic analysis part 13 which extracts the feature of the input voice by taking an acoustic analysis of the input voice, a speaker cluster selection part 14 which selects the standard speaker cluster corresponding to the acoustically analyzed voice according to similarity from a speaker model, a pattern matching part 18 which performs pattern matching according to the voice standard pattern of the selected standard speaker cluster, and a recognition result decision part 19 which decides the recognition result by the pattern matching and outputs the decision output. The speaker cluster selection part 14 determines the number of clusters according to the mutual information amount coefficient MIC of the selected standard speaker cluster.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition device that recognizes the speech of unspecified speakers using clustered standard patterns.

0002

2. Description of the Related Art Conventional speech recognition devices have used a method of creating a standard pattern using speech data uttered by a large number of speakers in order to recognize the speech of unspecified speakers. As standard patterns, attempts have been made to find an average pattern from patterns of unspecified speakers, and to use all patterns in a multi-template manner. In addition, methods for clustering standard patterns from a large number of speakers have also been attempted.

0003

[Problem to be Solved by the Invention] In the above-mentioned conventional speech recognition device that averages the standard patterns of all speakers, and in the above-mentioned conventional speech recognition device that uses multi-templates that use all the standard patterns, There is a problem that erroneous recognition caused by overlapping different phonemes cannot be avoided.

[0004]Furthermore, the above-mentioned conventional speech recognition device that uses the method of clustering standard patterns has the problem that even when performing clustering, it depends on the recognition vocabulary, and it is not possible to determine the optimal number of clusters. .

[0005] In view of the problems of the conventional speech recognition device described above, the present invention has been made to
To provide a speech recognition device capable of determining the optimum number of clusters of a standard pattern without depending on recognition vocabulary.

[0006]

[Means for Solving the Problems] The present invention provides an acoustic analysis means for acoustically analyzing input speech to extract speech features, and a standard speaker cluster corresponding to the acoustically analyzed speech from a speaker model. a sorting means for sorting based on the similarity of;
It is equipped with a pattern matching means for performing pattern matching based on the speech standard pattern of the selected standard speaker cluster, and a recognition result determination means for determining the recognition result by the pattern matching and outputting the determination result. This is achieved by a speech recognition device configured to determine the number of clusters based on mutual information coefficients of selected standard speaker clusters.

[0007]

[Operation] The acoustic analysis means acoustically analyzes the input speech to extract speech features, and the selection means selects standard speaker clusters corresponding to the acoustically analyzed speech based on similarity from the speaker model. The number of clusters is determined based on the mutual information coefficient of the selected standard speaker clusters, and the pattern matching means performs pattern matching based on the speech standard pattern of the selected standard speaker clusters, and the recognition result determining means determines the recognition result by pattern matching and outputs the determination result.

[0008]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of a speech recognition apparatus according to the present invention will be described with reference to the drawings.

FIG. 1 is a flowchart for explaining the operation of the speech recognition apparatus of the present invention. FIG. 2 shows the configuration of an embodiment of the speech recognition apparatus of the present invention which operates based on the flowchart shown in FIG.

First, the configuration of the speech recognition device shown in FIG. 2 will be explained.

The speech recognition device shown in FIG.
0, amplifier filter 11, analog/digital (A
/D) Consists of a conversion section 12, an acoustic analysis section 13, a speaker cluster selection section 14, speech standard patterns 15 to 17, a pattern matching section 18, and a recognition result determination section 19.

Next, the operation of each of the above-mentioned components will be explained.

The microphone 10 inputs the voice of an unknown user, and the amplifier filter 11 amplifies the voice input from the microphone 10. A/D converter 12
converts the analog audio signal amplified by the amplifier/filter 11 into a digital audio signal.

The audio analysis section 13 extracts features of the audio signal digitized by the A/D conversion section 12.

The speaker class selection unit 14 determines which standard speaker class the user's voice belongs to, using the degree of similarity from the speaker model, and selects the voice standard of one of the standard speaker clusters. Select patterns 15-17.

The pattern matching unit 18 performs pattern matching using the standard speech patterns selected by the speaker class selection unit 14, and the recognition result determination unit 19 uses the distance to the standard pattern found to perform pattern matching. It determines whether the standard speech pattern is appropriate and outputs the recognition result.

The feature extracted by the voice analysis section 13 may be either a feature obtained by Fourier spectrum analysis or a feature obtained by a band-pass filter, as long as it can effectively extract the features of the speech. Furthermore, as an independent speaker model that is used in the speaker cluster selection unit 14 and does not depend on the text, any model that can define the degree of similarity between speakers may be used, such as a method based on vector quantization. be. Furthermore, the pattern matching method used by the pattern matching section 18 is as follows:
Dynamic programming (DP matching), hidden Markov model,
Any method using a neural network may be used.

Standard patterns for word recognition as recognition vocabulary are stored in word units or in combinations of smaller units. Furthermore, standard patterns for continuous speech recognition are stored in units such as syllables and phonemes.

Next, the processing procedure for clustering standard speakers will be illustrated with reference to the flowchart of FIG.

First, prepare the voices of many standard speakers (
Step S1) A text-independent speaker model that does not depend on the text is created using a method based on vector quantization using part or all of the speech for standard pattern creation (Step S2), and the speech used for creating the speaker model is or other voices are input to each speaker model to obtain a similarity matrix between speakers (step S3), and two speakers or speaker classes are merged based on the similarity matrix between speakers. (Step S4), and perform clustering by merging similar speakers by repeating Step 4 until the mutual information coefficient MIC becomes maximum (Step S4).
5).

The above mutual information coefficient MIC will now be explained.

The mutual information MI is used in information theory and is defined by the following equation (1).

MI(x,y)=H(x)+H(y)−H
(x, y) ... (1) H(x) is the entropy for the probability p(x) of the input speaker x, H(y) is the entropy for the determined probability p(y) of the speaker y, H(x,y) is the entropy for the joint probability p(x,y) of the input speaker x and the determined speaker y.

The amount of mutual information MI is determined from the similarity matrix between speakers. In the similarity matrix between speakers, the smaller the similarity between different speakers, that is, the more independent the speakers, the larger the mutual information MI, and the smaller the speakers that occur are closer to each other.

[0025] Therefore, bottom-up clustering can be realized by merging any set of speaker clusters so that the mutual information of the similarity matrix between speakers becomes large.

However, since the mutual information MI tends to increase as the number of clusters increases, this value is normalized by the maximum mutual information obtained when the number of clusters is determined, and This is a measure to evaluate how well clustering is achieved.

The maximum mutual information MImax is defined by the number M of clusters as shown in the following equation (2).

MImax=logM...(
2) Here, the mutual information coefficient MIC is defined as the ratio from the maximum mutual information MImax as shown in the following equation (3).

MIC=MI(x,y)/MImax
(3) By using the mutual information coefficient MIC, clustering efficiency can be evaluated regardless of the number of clusters.

[0030]

[Effects of the Invention] The speech recognition device of the present invention includes an acoustic analysis means for acoustically analyzing input speech to extract speech features, and a standard speaker cluster corresponding to the acoustically analyzed speech from a speaker model. a sorting means for sorting based on the similarity of;
It is equipped with a pattern matching means for performing pattern matching based on the speech standard pattern of the selected standard speaker cluster, and a recognition result determination means for determining the recognition result by the pattern matching and outputting the determination result. Since the number of clusters is determined based on the mutual information coefficient of the selected standard speaker clusters, the number of clusters can be determined based on the mutual information coefficient of the selected standard speaker clusters. The optimal speaker cluster can be selected, and as a result, speech from any speaker can be recognized with high accuracy.

[Brief explanation of drawings]

FIG. 1 is a flowchart for explaining the operation of a speech recognition device according to the present invention.

FIG. 2 is a diagram showing the configuration of an embodiment of a speech recognition device that performs the operations shown in FIG. 2;

[Explanation of symbols]

10 Microphone 11 Amplifier filter 12 A/D conversion section 13 Speech analysis section 14 Speaker cluster selection section 15-17 Audio standard pattern 18 Pattern matching section 19 Recognition result determination section

Claims (1)

[Claims] 1. Acoustic analysis means for acoustically analyzing input speech to extract speech features; and selecting standard speaker clusters corresponding to the acoustically analyzed speech based on similarity from a speaker model. a pattern matching means for performing pattern matching based on the speech standard pattern of the selected standard speaker cluster; and a recognition result determining means for determining a recognition result by the pattern matching and outputting the determination result. A speech recognition device comprising: the selecting means configured to determine the number of clusters based on mutual information coefficients of the selected standard speaker clusters.