JP2001147698A - Method of generating pseudo word for voice recognition and voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To easily provide an out-of-candidate speaking rejection function for a recognition system which doesn't have this function. SOLUTION: Plural pseudo-words are generated with respect to a keyword, and reception or rejection of speaking is decided in accordance with how the degrees of similarity of pseudo-words for speaking are ranked.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a system for rejecting non-candidate utterances when a word other than a word registered as a recognition target word is uttered in a speech recognition system.

[0002]

2. Description of the Related Art Conventionally, as a method for rejecting non-candidate utterances, as disclosed in Japanese Patent Application Laid-Open No. 4-255900, the similarity of a recognition candidate word is normalized by the similarity of a reference model, and the utterance environment and There are some which reject stable non-candidate utterances without being affected by differences in individual voice quality.

[0003]

In the method of normalizing the similarity using the reference model as described above, it is necessary to perform necessary normalization in a voice-model collation unit which is the center of a recognition system. Therefore, it has been difficult to add the non-candidate utterance rejection function to a recognition system that does not have the non-candidate utterance rejection function. An object of the present invention is to use a recognition candidate dictionary that can be freely used by a user of a recognition system,
An object of the present invention is to easily provide a function for rejecting non-candidate utterances to a recognition system having no function for rejecting non-candidate utterances.

[0004]

In order to achieve the above-mentioned object, according to the present invention, one or more words are inputted,
A pseudo word generation method for speech recognition, comprising calculating a similarity between an input word and another word, and generating a word different from the input word based on the calculated similarity between the words. To automatically generate words similar to dictionary words.

Further, in the present invention, the similarity between the generated pseudo word and the similar word is retained.
The speech recognition apparatus has a rejection unit that rejects non-candidate utterances by referring to the similarity during speech recognition.

With this configuration, it is possible to create a pseudo word similar to a word in the dictionary but not stored in the dictionary, and it is possible to accurately reject non-candidate utterances and increase the speech recognition rate. it can.

[0007]

Embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a diagram showing an embodiment of the present invention.

In this embodiment, in a speech recognition apparatus, when a designated keyword is uttered, the utterance is accepted as a recognition candidate, and a speech other than the designated keyword is rejected. It generates recognition candidate dictionaries containing pseudowords, and takes the form of pseudoword generation tools, pseudoword generation functions, pseudoword generation APIs, and pseudoword generation middleware.

This pseudo-word generation tool for speech recognition includes a keyword input unit 101, a generation number designation unit 102, a pseudo-word generation unit 103, and a similarity calculation unit 104.

The keyword input unit 101 is an interface for inputting a word 901 to be recognized.
The generation number specification unit 102 is an interface for specifying the number 902 of pseudo words to be generated. Pseudo word generator 103
Is to generate a plurality of non-identical pseudo-words for the word 901 to be recognized, combine the word 901 to be recognized with the generated pseudo-word, and output it as a recognition candidate dictionary 903. The similarity calculator 104 is for calculating the similarity between two given words.

Next, the operation of the embodiment of FIG. 1 will be described.

The input word 901 to be recognized is
For example, one or more kana character strings are input using an interface provided by the keyword input unit 101. The interface provided by the keyword input unit 101 can use, for example, a method of reading a file listing the words 901 to be recognized or a method of passing a list of words 901 to be recognized as an argument of a function.

The number 902 of pseudo words to be generated is input using an interface provided by the generation number specifying unit 102. The interface provided by the generation number specification unit 102 is as follows:
For example, a method of passing the number of pseudo words to be generated as an argument of a function can be used.

The pseudo-word generation unit 103 generates pseudo words that are not identical to each other with respect to the input word 901 to be recognized, for example, by the following algorithm.

1. 1. Let n be the number of vowels included in the keyword. 2. Generate a word composed of only vowels using random numbers and having a vowel number of n ± α. 3. Use the similarity calculator 104 to calculate the similarity between the generated word, the keyword, and the already generated word. 4. If the similarity between the generated word and the keyword is equal to or less than a certain reference, the generated word is discarded and the process proceeds to step 2. If the similarity between the generated word and the keyword or the already generated word is equal to or more than a certain reference, the generated word is discarded, and the process proceeds to step 2. 6. Register the generated words, and if the number specified by the generation number specifying unit 102 has been reached, go to step 7; otherwise, go to step 2. The similarity calculator 104 outputs a list of keywords and the generated words as the recognition candidate dictionary 903. The similarity calculator 104 calculates the similarity between the two words given from the pseudo-word generator 103 by, for example, a dynamic programming matching method. The phonemes to be performed are classified into correspondence, substitution, dropout, and boiling out, and the score is given to each of them, and the total is calculated.

As described above, according to the second aspect, the recognition candidate dictionary 903 for rejecting utterances other than keywords can be generated.

FIG. 2 is a diagram showing an embodiment of the present invention.

In this embodiment, in a speech recognition apparatus, when a specified keyword is uttered, the utterance is accepted as a recognition candidate, and a utterance other than the specified keyword is rejected. It generates a recognition candidate dictionary that contains information on the similarity ranking of pseudowords and pseudowords for keywords, and includes forms such as pseudoword generation tools, pseudoword generation functions, pseudoword generation APIs, and pseudoword generation middleware. Eggplant

The speech recognition pseudo word generation tool includes a keyword input unit 101, a generation number designation unit 102, a pseudo word generation unit 103, a similarity calculation unit 104, and a correlation calculation unit 20.
1 and a recognition model 202.

The keyword input unit 101 is an interface for inputting a word 901 to be recognized.
The generation number designation unit 102 is an interface for designating the number of pseudo words to be generated. Pseudo word generator 103
Is to generate a plurality of non-identical pseudo-words for the word 901 to be recognized, combine the word 901 to be recognized with the generated pseudo-word, and output it as a recognition candidate dictionary 902. The similarity calculator 104 is for calculating the similarity between two given words. The correlation calculation unit 201 calculates, for each keyword,
The ranking of the similarity of the generated pseudo-word to the keyword is performed. The recognition model 202 is information for the correlation degree calculation unit 201 to calculate the similarity.

Next, the operation of the embodiment of FIG. 2 will be described.

Until the pseudo-word generation unit 103 outputs a list including the keywords and the generated words, FIG.
This is the same as the embodiment shown in FIG. The correlation degree calculation unit 201 ranks pseudo words for each keyword by, for example, the following algorithm.

1. The keyword is converted into a sequence of the recognition model 202 corresponding to the phoneme sequence forming the keyword. For example, an HMM (Hidden Markov Model) can be used as the recognition model. Since the HMM also includes information on the duration of a phoneme, it is equivalent to expressing a pseudo utterance having a temporal element by arranging recognition models corresponding to each phoneme according to the duration information. (Example) aki-> aaaakkiiiiiii 2. The recognition model sequence and each pseudoword are recognized by the recognition model 202.
Is collated using the information of, and the similarity is calculated. This includes
For example, a Viterbi algorithm can be used. 3. Pseudo-word ranking is performed for each keyword according to the calculated similarity. Outputting a Recognition Candidate Dictionary Containing Pseudo Word Ranking Information for Each Keyword As described above, according to the second invention, it is possible to generate a recognition candidate dictionary 904 for rejecting utterances other than keywords. it can.

FIG. 3 is a diagram showing an embodiment of the present invention.

This embodiment relates to a recognition apparatus for rejecting utterances other than keywords using the recognition candidate dictionary generated by the invention shown in FIG.

This speech recognition apparatus comprises an analysis section 301, a recognition candidate dictionary 302, a collation section 303, a rejection section 304, and a recognition model 202.

The analysis section 301 analyzes the input speech and converts it into a time series 906 of feature quantity vectors. The recognition candidate dictionary 302 is a word list including keywords and pseudo words generated by the embodiment shown in FIG. The matching unit 303 compares words included in the recognition candidate dictionary 302 with a time series 906 of feature amount vectors, calculates similarity, and outputs recognition candidates 907. The rejection unit 304 is for determining whether to output a candidate having the highest similarity among the recognition candidates 907 as a recognition result.

Next, the operation of the embodiment shown in FIG. 3 will be described.

The input speech 905 is converted by the analysis unit 301 into a time series 906 of feature quantity vectors. As the analysis means, for example, an LPC cepstrum (Linear Predictive Coding Cepstrum) as described on page 139 of “Basic of speech recognition (above)” (hereinafter referred to as Reference 1) issued by NTT Advanced Technology Corporation is used. be able to. Time series 906 of the feature vector
Is compared with each of the words included in the recognition candidate dictionary 302 by the matching unit 303, and the similarity is calculated. In the comparison by the matching unit 303, the recognition model 202, which is a database of information on how the voice is uttered, is used. As an expression method of the recognition model 202, for example, an HMM can be used. In this case, a Viterbi algorithm can be used as a comparison method. The matching unit 303 outputs a recognition candidate list 907 ranked according to the similarity as a result of the comparison. Rejection unit 30
4, whether the word having the highest similarity in the recognition candidate list 907 is a keyword or a pseudo word is determined by the information 909 of the recognition candidate dictionary, and if the word matches the keyword, the recognition result 908. Is output, and if the word is a pseudo-word, the recognition result is rejected.

As described above, according to the fifth invention, the speech recognition apparatus having the function of rejecting the utterance outside the keyword without modifying the core parts of the recognition system such as the analysis unit, the collation unit, and the recognition model. Can be realized.

FIG. 4 is a diagram showing an embodiment of the present invention.

This embodiment uses the recognition candidate dictionary generated by the invention shown in FIG. 2 to generate utterances other than keywords using fewer pseudowords than the recognition candidate dictionary according to the invention shown in FIG. It relates to a recognizing device to be rejected.

This speech recognition apparatus comprises an analysis section 301, a recognition candidate dictionary 401, a collation section 303, a rejection section 402, and a recognition model 202.

The analysis unit 301 analyzes input speech and converts it into a time series 906 of feature quantity vectors. The recognition candidate dictionary 401 is a word list generated by the embodiment shown in FIG. 2 and including the ranked keywords and pseudo words. The matching unit 303 compares the words included in the recognition candidate dictionary 401 with the time series 906 of the feature amount vectors, calculates the similarity, and outputs the recognition candidates 907. Rejection section 402 recognizes candidate 907
This is for determining whether or not to output the candidate with the highest similarity among the recognition results.

Next, the operation of the embodiment shown in FIG. 4 will be described.

The input speech 905 is converted by the analyzer 301 into a time series 906 of feature quantity vectors. As the analysis means, for example, an LPC cepstrum (Linear Predictive Coding Cepstrum) as described on page 139 of “Basic of speech recognition (above)” (hereinafter referred to as Reference 1) issued by NTT Advanced Technology Corporation is used. be able to. Time series 906 of the feature vector
Is compared with each of the words included in the recognition candidate dictionary 402 by the matching unit 303, and the similarity is calculated. In the comparison by the matching unit 303, the recognition model 202, which is a database of information on how the voice is uttered, is used. As an expression method of the recognition model 202, for example, an HMM can be used. In this case, a Viterbi algorithm can be used as a comparison method. The matching unit 303 outputs a recognition candidate list 907 ranked according to the similarity as a result of the comparison. Rejection unit 40
In 2, it is determined whether the uttered voice is a keyword as follows.

1. If the highest similarity in the recognition candidate list 907 is a pseudo word, it is not a keyword. When the keyword having the highest similarity in the recognition candidate list 907 is a keyword, (i) ranking information 9 of the pseudo-word related to the keyword from the recognition candidate dictionary 401
(Ii) the order of appearance of the pseudo-words in the recognition candidate list 907;
The pseudo-word ranking information 910 is compared. If the similarity is equal to or more than a certain value, the keyword is determined to be a keyword. As a method of calculating the similarity, for example, the top ten ranks of the recognition candidate list 907 are the same rank in the pseudo-word ranking information 910, and 10 points.
Point, eight points,..., And weighted and summed values can be used.

The rejection unit 402 outputs the keyword as a recognition result 908 when the determination is a keyword, and rejects the recognition result when the determination is not a keyword.

In the embodiment shown in FIG. 3, if the keyword having the highest similarity in the recognition candidate list 907 is a keyword, it is determined that the uttered word is the keyword. By using the word ranking information 910, a stable rejection performance can be obtained using a smaller number of pseudo words as compared with the embodiment shown in FIG.

As described above, according to the sixth aspect, the speech recognition apparatus having the function of rejecting the utterance outside the keyword without changing the core parts of the recognition system such as the analysis unit, the matching unit, and the recognition model. Can be realized.

[0042]

As described above, according to the present invention, a pseudo word for rejection is automatically generated from a given recognition candidate, and further, by using the rank information of the recognition candidate, a stable candidate can be easily obtained. An external voice rejection function can be realized.

[Brief description of the drawings]

FIG. 1 shows an embodiment of a pseudo-word generation tool for speech recognition.

FIG. 2 shows an embodiment of a pseudo-word generation tool for speech recognition.

FIG. 3 shows an embodiment of a voice recognition device.

FIG. 4 shows an embodiment of a speech recognition device.

[Explanation of symbols]

101: keyword input unit, 102: generated number unit, 10
3: pseudo word generator, 104: similarity calculator, 201:
Correlation calculation unit, 202: recognition model, 301: analysis unit, 302: recognition candidate dictionary, 303: collation unit, 304:
Rejection unit, 401: recognition candidate dictionary, 402: rejection unit, 90
1: Word to be recognized, 902: Number of pseudo words to be generated, 903: Recognition candidate dictionary, 904 Recognition candidate dictionary, 90
5: input voice, 906: time series of feature amount vectors, 907: ranked recognition candidate list, 908:
Recognition result, 909: recognition candidate dictionary information, 910: pseudo word ranking information.

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Akio Amano 1-280 Higashi Koigakubo, Kokubunji-shi, Tokyo F-term in Central Research Laboratory, Hitachi, Ltd. 5D015 AA04 LL03 LL05

Claims (4)

[Claims] 1. A method according to claim 1, wherein one or more words are input, a similarity between the input word and another word is calculated, and a word different from the input word is determined based on the calculated similarity between the words. A method for generating a pseudo-word for speech recognition, characterized by generating the pseudo-word. 2. A method for calculating a similarity between the generated word and each of the input words, and in a case where there are a plurality of generated words, according to the similarity to the plurality of generated words. 2. The pseudo-word generation method for speech recognition according to claim 1, wherein ranking is performed. 3. A dictionary unit for holding a pseudo word and a keyword to which a similarity degree is attached to another word as a recognition candidate dictionary, and a uttered voice and a word included in the recognition candidate dictionary stored in the dictionary unit. A speech recognition apparatus, comprising: a matching unit that outputs a similarity to the word; and a rejection unit that rejects a recognition result when the word having the highest similarity output by the matching unit is the pseudo word. . 4. A dictionary unit for holding a pseudo word and a keyword to which a similarity degree is attached to another word as a recognition candidate dictionary, and a uttered voice and a word included in the recognition candidate dictionary stored in the dictionary unit. A matching unit that calculates the degree of similarity with the keyword, and if the word having the highest similarity output by the matching unit is one of the keywords, the word is output by the matching unit among the words included in the recognition candidate dictionary. Having a rejection unit that determines whether the uttered voice is a keyword by comparing the appearance pattern of the word having a high similarity with the ranking information of the pseudo words made by the correlation degree calculation unit. Characteristic speech recognition device.