JPH02226200A - Voice recognition device - Google Patents

Abstract

PURPOSE:To improve the recognition rate by obtaining template information which is most similar to an input pattern according to the output of a matching part and selecting the template of a corresponding similar speaker cluster by referring to two tables. CONSTITUTION:A standard pattern part 4 is provided with a similar speaker cluster table stored with plural other speaker clusters whose voices are most similar to an actual speaker cluster and the actual speaker cluster as standard patterns which are predetermined correspondingly by speaker clusters. A template selection part 5 obtains the template information which is most similar to an input pattern according to the output template 17 of a matching part 2 and refers to a category and a similar speaker cluster table to select and supply the template 16 of the similar speaker clusters to a matching part 2. Consequently, the quality of a limited amount of templates can be increased as much as possible adaptively to the speaker, and the recognition rate can be improved.

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field] The present invention is applied to a speech recognition device for digital speech processing. In particular, the present invention relates to a speech recognition device for non-specific speakers and for multiple speakers using a multi-template method.

〔overview〕

The present invention provides a speech recognition device that estimates a cluster of similar speakers based on information about a template that is most similar to a human pattern obtained from the results of the previous matching process, and selects a template to be used when performing the next matching process. By making this selection, the utility value of the limited amount of templates can be maximized by adapting them to the speaker, thereby improving the recognition rate.

[Conventional technology]

Conventionally, speech recognition devices for non-specific speakers and for multiple speakers often use a multi-template method, but this method always uses multiple templates prepared in advance for each category for matching. .

[Problem that the invention seeks to solve]

However, in such a voice tRm device, recognition performance is greatly influenced by the quality and quantity of templates. Since the amount of templates is limited by hardware, there is a drawback in that recognition performance is limited when a limited amount of templates is applied to many speakers. That is, voices of different contents uttered by speakers with a high degree of dissimilarity may exhibit a high degree of similarity. As a result, voices with different content may have the same identification result, or conversely, voices with the same content may have different identification results.
As a result, there was a drawback that the recognition performance deteriorated.

The present invention solves the above-mentioned drawbacks, and aims to provide a speech delta intelligence device that can maximize the quality of a limited amount of templates by adapting them to the speaker and improve the recognition rate. purpose.

[Means for solving problems]

The present invention includes a standard pattern section including a category table storing templates divided into predetermined categories and corresponding to each speaker cluster that is clustered for speakers who have uttered sounds in a predetermined standard pattern. , an analysis section that analyzes input signals and converts them into human patterns;
In a speech recognition device that includes a matching unit that performs a matching process between the output of the analysis unit and the contents of the category table, the standard pattern unit generates the predetermined standard pattern for each speaker cluster. A similar speaker cluster table is provided in which multiple other speaker clusters and own speaker clusters whose uttered voice is most similar to the own speaker cluster are provided, and the above human pattern is determined based on the output of the matching section. The present invention is characterized by comprising a template selection unit that obtains template information most similar to the above, refers to the two tables, selects a template of a corresponding similar speaker cluster, and supplies the template to the matching processing unit.

[Effect]

The standard pattern section stores multiple other speaker clusters and self-speaker clusters whose voice uttered as a predetermined standard pattern corresponding to each speaker cluster is most similar to the self-speaker cluster. A Tarastar table will be provided. The template selection section obtains template information most similar to the input pattern based on the output of the matching section, refers to the category table and the similar speaker cluster table, selects the template of the corresponding similar speaker cluster, and sends the template to the matching processing section. give. Through the above operations, the quality of the limited amount of templates can be maximized by adapting them to the speaker, and the recognition rate can be improved.

〔Example〕

Embodiments of the present invention will be described with reference to the drawings. 1st
The figure is a block diagram of a speech recognition device according to an embodiment of the present invention. In FIG. 1, the speech recognition device stores templates divided into predetermined categories corresponding to each speaker cluster, which is clustered for speakers who have uttered speech according to a predetermined standard pattern. an analysis section 1 that analyzes the input signal 11 and converts it into a manual pattern;
The matching unit 2 includes a matching unit 2 that performs a matching process between the output human pattern 12 and the content of the category table, and an identification unit 3 that outputs an identification result 14 based on the matching result of the matching unit 2.

Here, the feature of the present invention is that the standard pattern section 4 is configured to select a plurality of other patterns in which the voice uttered as the predetermined standard pattern corresponds to each speaker cluster and is most similar to the own speaker cluster. A similar speaker cluster table in which speaker clusters and white speaker clusters are stored is provided, and template information most similar to the input pattern is obtained based on the output template H7 of the matching section 2, and the two tables are referred to. The present invention is characterized in that it includes a template selection section 5 that selects a template 16 of a corresponding similar speaker cluster and supplies it to the matching processing section 2.

The operation of the speech recognition device having such a configuration will be explained. Table 1 is a similar speaker cluster table of the speech recognition device of the present invention. Table 2 is a category table for the speech recognition device of the present invention. FIG. 2 is a flowchart showing the operation of the template selection section of the speech recognition device of the present invention.

(Hereinafter, this page margin) Table 1 Table 2 First, the analysis section 1 converts the input signal 11 into an input pattern 12. The matching section 2 performs a matching process between this input cover pattern 12 and a template 16 used for matching. The identification section 3 outputs the identification result 14 from the matching result 13.

Further, the template selection unit 5 receives the template 17 most similar to the input cover pattern 12 as a result of the previous matching process, and selects a template to be used for the next matching based on that information.

The selected template 15 is received from the standard pattern section 4 and then output as a template 16 used for matching.

Here, the processing in the template selection section 5 will be explained in more detail. Table 1 is a speaker cluster table showing similar speaker clusters corresponding to speaker clusters. First, the speaker cluster (St, S2, S,) is a clustering of speakers who have uttered voices used as predetermined standard patterns. The data for clustering and ringing may be, for example, using five vowels uttered by each speaker, using more phoneme data of isolated utterances, or using data for all words (or phonemes, etc.) that form a category. There are ways to use it.

The degree of similarity between each clustered speaker cluster is determined using the data used for clustering. For example, select several other speaker clusters that are most similar to each speaker cluster, and find several similar speaker clusters including the speaker cluster itself. This number does not need to be the same for all speaker clusters.

In this example table, the similar speaker clusters for S, are s, , s, , Sb, and 32 for S2,
Sc, Sa.

Next, Table 2 shows each speaker cluster (St, -1S15,
SL) for each category (W+, -1WJ,...
Wo) template (T10,..., Tl,...
・T□). In this example, there is one template for each category in each speaker cluster, but there may be more than one template.

In FIG. 2, first, the matching section 2 selects a template 1 that is most similar to the input cover pattern as a result of the matching process.
7, the information on the template T1 is received (Sl).

From now on, the speaker cluster Si is found by referring to the category table shown in Table 2 (S2). Next, referring to the speaker cluster table shown in Table 1, similar speaker cluster S1,
Find So and Sf (S3). Finally, referring to the table shown in Table 2 again, templates Ti0, TI□,
%Tin%Tel,T,2, ,Tan5 T'r
+, Tf21, and Tf21. Then, the matching section 2 uses this selected template as the template 16.
Output to. (S4).

〔Effect of the invention〕

As explained above, the present invention can maximize the quality of a limited amount of templates by adapting them to the speaker.
It has an excellent effect of increasing the recognition rate.

Diagram.

FIG. 2 is a flowchart showing the operation of the template selection section of the speech recognition device of the present invention.

DESCRIPTION OF SYMBOLS 1... Analysis section, 2... Matching section, 3... Identification section, 4... Standard pattern section, 5... Template selection section, 11... Input signal, 12... Input pattern , 13
...Matching result, 14...Identification result, 15...
- Selected template, 16...Template used for matching, 17...Template most similar to the cover pattern entered as a result of matching processing.

Agent: Patent attorney Naotaka Ide

[Brief explanation of the drawing]

FIG. 1 is an embodiment of a speech recognition device block according to one embodiment of the present invention. Figure 1 Example: Flowchart of template selection section Fig. 2

Claims (1)

[Claims] 1. Contains a category table storing templates divided into predetermined categories corresponding to each speaker cluster obtained by clustering speakers who have uttered a predetermined standard pattern of speech. A speech recognition device comprising a standard pattern section, an analysis section that analyzes an input signal and converts it into an input pattern, and a matching section that performs a matching process between the output of this analysis section and the contents of the category table described above. The pattern section stores a plurality of other speaker clusters and a self-speaker cluster whose voice uttered according to the predetermined standard pattern is most similar to the self-speaker cluster corresponding to each of the above-mentioned speaker clusters. A speaker cluster table is provided, template information most similar to the input pattern is obtained based on the output of the matching section, the template of the corresponding similar speaker cluster is selected by referring to the two tables, and the above is performed. A speech recognition device comprising a template selection section for providing a template to a matching processing section.