JPS5857195A - 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 uses a speech recognition method, particularly a standard dictionary, to assign priorities to a plurality of single syllables and/or words (hereinafter collectively referred to as speech). , and extract it as a candidate voice.

The present invention relates to a speech recognition method in which a plurality of speech candidate sequences are stored in a candidate sequence dictionary with identifiers χ attached, and the validity of the extracted results is determined based on the identifiers.

Conventionally, in speech recognition devices for monosyllables and/or words, a standard dictionary is provided, feature quantities are extracted for unknown input speech, and the extracted feature quantities and the standard time read from the standard dictionary are used. The category to which the unknown input voice belongs is determined by comparing it with a trace amount. However, in the case of the above-mentioned conventional configuration, when trying to improve the recognition group, an extreme recognition result that should not be originally obtained may sometimes be obtained.

The purpose of the present invention is to solve such problems, and it is an object of the present invention to enable users to check the validity of recognition results themselves. And for that reason, the audio i! of the present invention! The system includes a feature extraction circuit section that extracts the feature amount of the input audio signal based on the frequency analysis result of the input audio signal, a standard dictionary storing the standard time trace amount corresponding to the standard speech, R, and the above. In the voice twill weaving device which recognizes the voice corresponding to the input voice signal, the voice twill weaving device includes a collation circuit unit that collates the feature amount obtained by the feature extraction circuit unit and the standard percentage trace amount read from the standard dictionary. The circuit unit Kg is configured to be able to extract a plurality of candidate voices by assigning priorities to them and output the extracted voice candidate sequences, and also stores a plurality of types of voice candidate sequences and supports each of the candidate voices. Candidate series dictionary with job name 't'll #4,
and a comparison circuit section that compares the speech candidate series outputted from the matching circuit with the speech candidate series read from the candidate series dictionary, and in the comparison circuit section, the speech candidate series that is matched is added to the speech candidate series. The present invention is characterized in that it is determined whether or not to adopt the speech Z recognition result having the highest priority among the speech candidate series outputted from the collation circuit unit based on the identifier that is present. This will be explained below with reference to the drawings.

The figure shows the configuration of an embodiment of the present invention. In the figure, numeral 1 is a feature extraction circuit, and 2 is a standard dictionary, for example, the number roJ.
, rIJ, r2J *... The standard time trace amount for the case of pronouncing "9" is stored together with the category name, 3 is a matching circuit section, 4 is a dictionary of candidate series that can be changed in the present invention, 5 represents the comparator circuit section χ provided in the present invention.

Candidate' series dictionary 4 contains I! For example, for the word rOJ t rlJ @ r2J t... "9" in the recognition category, a candidate series "0-4-5J" consisting of m combinations is created.
, “o-1-2J.

. . . are stored, and an identifier iD obtained based on, for example, statistical processing is assigned to each candidate sequence. The identifier rFJ indicating the armpit IDK means that if the corresponding candidate series is obtained, the word with the highest priority in the current scissors candidate series may be considered as the correct answer,
Further, the identifier rRJ means that if the candidate sequence is obtained, the matching result by the matching circuit section 3 at the previous stage should be discarded and "reject" should be answered.

In the figure, similar to the conventional speech recognition device, the matching circuit section 3 determines, for example, the distance between the feature amount extracted by the feature extraction section 1 and the standard time minute amount read out from the standard dictionary 2. In the case of the present invention, K, for example, three categories are determined as candidates in order from the one with the smallest determined matching distance, and the candidate series is arranged in order from the one with the smallest matching distance, that is, in the priority layer, for example, as shown in the figure. (0-1-2)
v output.

The candidate series (0-1-2) is led to the comparison circuit section 5,
On the other hand, from the candidate series word w4, for example (0-4-5°F)s(
012, , R) . The comparison circuit unit 5 selects both of the above candidate series (0-1-2).
and (045)t(012), . . . are compared. Then, if a match is found, the corresponding identifier iDV is checked, and if the identifier is rFJ, the word t with the highest priority in the candidate series is considered the correct answer, and if the identifier is rRJ, the correct answer is K). Reject the matching results from . As shown in the figure, when the candidate series (0-1-2) is obtained from the matching circuit section 3, the comparison circuit section 5 issues a reject. This is the candidate system Da 11 mentioned above.
(0-1-2)K extracted words “0”, “1”
.

This is because, from a phonetic perspective, "2" is Nokurano (Unamo J), which means that the collation result for the word rOJ, which has the highest priority, is extremely reliable.

In addition, in the illustrated configuration, if the above-mentioned redirect error occurs, the utterance is redone, or the feature extraction is performed from a different perspective than the feature extraction by the illustrated feature extraction 1 section 1, and the matching is redone. It is possible to add a configuration to do so. Also, in the above explanation, words such as numbers ti! However, it goes without saying that it can also be applied to the case of recognizing monosyllables.

As described above, according to the present invention, it is possible to have the verification circuit section check the validity of the verification result by itself, and it is possible to improve the reliability of the am result.

[Brief explanation of drawings]

Is the figure an example configuration of the present invention? show. - In 1, 1 represents the % total extraction circuit, 2 represents the keyaki quasi-dictionary, 3 represents the collation circuit, 4 represents the candidate series dictionary, and 5 represents the comparison 10 circuit 87. Patent applicant Fujitsu Limited

Claims (1)

[Claims] A feature extraction circuit section that extracts feature quantities of an input speech signal based on the frequency analysis results of the input speech signal, a standard dictionary that stores standard feature quantities corresponding to standard speech, and a feature extraction circuit section that extracts feature quantities of the input speech signal based on the frequency analysis results of the input speech signal. It has a matching circuit unit that compares the feature quantity read out from the standard dictionary with the standard feature quantity read from the standard dictionary, and generates a voice vil! corresponding to the input voice signal. ! The speech recognition device is configured such that the collation circuit unit is capable of extracting a plurality of candidate speeches by assigning priorities to them and outputting the extracted speech candidate sequences. a candidate sequence dictionary in which the candidate sequence dictionary is stored and an identifier is assigned to each corresponding sequence, and a comparison circuit section t that compares the voice candidate sequence outputted from the matching circuit and the voice candidate sequence read from the candidate sequence dictionary. Make money,
The highest priority among the speech candidate sequences outputted from the collation circuit section based on the identifier assigned to the speech candidate series that matched in the comparison circuit section? Motsu audio tw
A speech recognition method characterized in that it is determined whether or not to be adopted as a recognition result.