JPH06308994A - Japanese language voice recognizing method - Google Patents

Abstract

PURPOSE:To provide a Japanese language voice recognizing method by which processing time is shortened, and the conversion performance can be improved by deleting statistically a grammatically erroneous candidate, and preventing many candidates by a combination of simple characters from being generated. CONSTITUTION:As for a statistical language model, a statistical language model related to occurrence sequence information 101 of a character, and occurrence sequence information 102 of a part of speech, generated from a learning text data base 100 is used, and as for a standard pattern, a syllable standard pattern 201 and a syllable standard pattern 202 for reading of a character of a learning voice data base 200 are used.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a Japanese speech recognition method, and in particular, as a method for recognizing phonemes, a method based on pattern matching and a method based on feature extraction are probabilistically performed for state transition. Hidden Markov model that outputs a symbol with a certain probability during state transition (for example, Seiichi Nakagawa "Speech Recognition by Established Model", The Institute of Electronics, Information and Communication Engineers (1
988)) and statistical language models (eg LR Bahl
Other “A Statistical Approach to Continuous Speech Rec
ognition "IEEE Trans. on PAMI (1983)" and Japanese speech recognition method.

[0002]

2. Description of the Related Art FIG. 4 is a diagram for explaining a conventional Japanese speech recognition system.

The system shown in FIG. 1 has a learning text database 10 having statistical language model information 11 regarding kana / kanji occurrence order, and statistical language model information 11 of the learning text database 10 stored in advance. A language model memory 50, a learning voice database 20 having phoneme standard pattern information 21 of a hidden Markov model, a phoneme sequence memory 60 corresponding to characters / readings in which the phoneme standard pattern information 21 of the learning voice database is stored in advance, input A phoneme candidate selection unit 3 that selects a phoneme candidate from the phonemes immediately before the recognition process for the voice has been completed.
0, a recognition unit that collates the standard model of the learning text database 10 with the input voice, and recognizes and outputs the candidate kana / kanji character having the highest overall likelihood based on the phoneme standard pattern information of the learning voice database 20. It consists of 40.

Next, the operation of the conventional system will be described.
As a method of converting an input speech into a kana / kanji sequence by speech recognition using a conventional hidden Markov model and a statistical language model, a kana /
The statistical language model 11 regarding the occurrence order of Chinese characters, the phoneme standard pattern 21 of the hidden Markov model from the learning speech database 20, and the statistical language model memory 5 are stored.
0, created in advance in the phoneme sequence memory 60 corresponding to character reading, and using the statistical language model 11 and the phoneme sequence information corresponding to character reading for the input speech in the speech candidate selection unit 30. , From the previous phonemes already recognized,
Next, a plurality of phoneme candidates with a high probability of occurrence are selected, and for each of these selected phoneme candidates, the recognition unit 40 compares the phoneme standard pattern with the input speech to determine the occurrence likelihood based on the statistical language model. , It has been proposed to output a candidate kana / kanji character having the highest total likelihood of similarity with a standard pattern by a hidden Markov model as a recognition result.

[0005]

However, in the above-mentioned conventional conversion method, the statistical language model handles only the character faces of kana / kanji, and the conversion information from the phoneme sequence to the converted characters of kana / kanji characters is used. As a result, since only the information of the phoneme sequence corresponding to the reading of each kana / kanji character is given, there is a problem that many candidates are generated by a simple combination of characters.

The present invention has been made in view of the above points, and solves the above-mentioned conventional problems and statistically selects grammatically correct candidates to statistically select grammatically incorrect candidates. An object of the present invention is to provide a Japanese speech recognition method that can be deleted to prevent a large number of candidates generated by simple character combinations from being generated, thereby reducing processing time and improving conversion performance.

[0007]

FIG. 1 is a diagram for explaining the principle of the present invention.

The present invention uses the input speech as a time series of characteristic parameters, and uses the statistical language model 101 relating to the occurrence order created from the learning text database 100, and uses a plurality of speeches for the characteristic parameter time series of the input speech. A recognition candidate is selected, and for each of these speech recognition candidates, the standard pattern of the hidden Markov model of the learning speech database 200 is compared with the characteristic parameter of the input speech and the time series of the characteristic parameter of the input speech, respectively, and the likelihood of occurrence occurs. In a Japanese speech recognition method in which a candidate having a high total likelihood of is as a recognition result, regarding the occurrence order information 101 of characters and the occurrence order information 102 of part-of-speech created from the learning text database 100 as a statistical language model. Use a statistical language model.

Further, according to the present invention, the input speech is set as a time series of the characteristic parameters, and the statistical language model regarding the occurrence order created from the learning text database 100 is used. A voice recognition candidate is selected, and for each of these voice recognition candidates, the standard pattern of the hidden Markov model of the learning voice database 200 is compared with the input voice feature parameter and the input voice feature parameter time series, respectively, and the likelihood of occurrence occurs. In the Japanese speech recognition method in which a candidate having a high total likelihood with respect to the degree is used as a recognition result, a syllable standard pattern 201 of the learning speech database 200 and a syllable standard pattern 202 for reading a character are used as standard patterns.

[0010]

The present invention considers the occurrence order of part-of-speech in addition to the occurrence order of kana / kanji as the information of the learning text database, and the phoneme standard pattern information (syllabic standard) of the hidden Markov model as the information of the learning speech database. In addition to (Pattern), by considering the reading standard pattern information of characters, grammatically correct candidates are statistically selected and candidates generated are narrowed down compared to candidates generated by simple character combinations. , It is possible to shorten the processing time required to convert the input voice into a kana / kanji sequence and improve the conversion performance.

[0011]

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

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

The system shown in FIG. 1 has a voice signal input terminal 1 for inputting a voice signal, a feature extraction unit 2 for extracting feature parameters, a recognition unit 3 for selecting character candidates and obtaining a likelihood of similarity with feature parameters, and a syllable. A standard pattern memory 4 for storing a chained standard pattern, a character / part-of-speech statistical language model memory 5 for storing a character part-of-speech statistical language model, and a recognition result output unit 6 for outputting a recognition result from the recognition unit 3. .

First, a standard pattern memory 4 for storing a statistical language model based on the occurrence order information 101 and the part-of-speech occurrence order information 102 in the learning text database 100 and the hidden Markov model speech in the learning speech database 200. The character / part-of-speech statistical language model is read out using the standard pattern information 201 and the character reading standard pattern information 202, and the character / part-of-speech statistical language model memory 5 is read.
To generate.

The voice input from the input terminal 1 is converted into a digital signal by the feature extraction unit 2, further LPC cepstrum analyzed, and then converted into a feature parameter every frame (for example, 10 milliseconds). This characteristic parameter is, for example, an LPC cepstrum coefficient.

From the learning speech database 200, standard patterns such as syllabic standard patterns of hidden Markov models and syllable chain standard patterns for reading Chinese characters are created in advance in the same format as the feature vectors and stored in the standard pattern memory 4. , Learning text database 10
A statistical language model concerning the occurrence sequence of characters (kana / kanji) and part of speech is created from 0 and stored in the character / part of speech statistical language model memory 5.

The recognizing unit 3 sets a standard pattern representing the reading of the character candidates for a plurality of character candidates selected by using the character / part of speech statistical language model stored in the character / part of speech statistical language model memory 5. It is read from the standard pattern memory 4 and the likelihood of similarity with the characteristic parameter of the input voice is obtained.

The recognition result output unit 6 outputs the result of recognition by the recognition unit 3 to a display or the like.

FIG. 3 is a diagram for explaining the recognition processing according to the embodiment of the present invention. In FIG. 3, the recognition unit 3 selects a candidate character for i-th recognition and the recognition result is selected from the candidate characters. Indicates the case of output. That is, for example, in order to recognize the i-th character with respect to the input speech, conditional probabilities regarding the occurrence order of kana / kanji and part-of-speech are used from the character / part-of-speech statistical language model to calculate the (i-2) th and (i The recognition result of each character of -1) th, the recognition result of the part of speech of each word of the (j-1) th and the (j-2) th, and the part of speech of the jth word assumed to appear next. Based on and, a plurality of characters with a high likelihood of being predicted to appear i-th are selected as candidate characters k1 to kn.

Here, the i-th predicted character is a part of the j-th word. As the j-th word, all existing part-of-speech may be assumed, or the part-of-speech that can be assumed by the character predicted to appear at the i-th may be limited. i-2, i-1, the character that appears i-th is c
_If the parts of speech of the j-2, j-1, j-th word are s _j-2 , s _j-1 , s _j , _i-2 , c _i-1 , c _i , then the above conditional probability is ,

[0021]

[Equation 1] Becomes If you don't have enough text and part-of-speech sequences to get this data,

[0022]

[Equation 2] It is also possible to obtain the same effect by using, for example. In addition, i-2, i-1, i-th appearing c _i-2 , c _i-1 ,
As c _i , not a simple character itself, but a character code to which reading information is added in advance may be used.

Next, for each of the selected candidate characters,
Create different candidates according to the assumed number of parts of speech.

For example, as shown in FIG. 2, the letter k₁Against
And part of speech h₁, H₂, H₃If we assume
Three candidates by combination are created. And these symptoms
For each supplement, read the character from the standard pattern memory 4.
Reads the standard pattern for the i-th input voice
And the likelihood of character and part-of-speech statistical language model memory 5
Characters and parts of speech read from
The sum of the i-th likelihood of the candidate character of
However, the candidate with the highest total likelihood, for example, k ₁, H₂
Is the highest overall likelihood, the letter k₁Is the i-th recognition result
The result is output to the recognition result output unit 6 as a character.

When a character is output, the total likelihood does not always become the highest even when the candidate having the highest total likelihood outputs the next character. Therefore, the above B candidates having a high total likelihood are left (the beam width is referred to as B), and the next process is carried over. The reason for setting the upper B is that it is difficult to hold all the candidates because the memory amount increases and the processing time increases accordingly. When the candidate having the highest total likelihood is changed, the candidate character output to the recognition result output unit 6 is also updated.

This kana / kanji candidate selection, the hypothesis of the part of speech of the character, the matching with the standard pattern for them, and the total likelihood thereof, the recognition result characters are output as a kana / kanji sequence in descending order of likelihood.

If there is a Kanji character in the input voice that is not in the learning text database, it cannot be recognized. In this case, the unrecognizable character (Kanji)
Is output as a blank in the kana / kanji sequence of the recognition result, or a statistical language model regarding the phoneme or kana occurrence sequence and the phoneme or syllable standard pattern of the hidden Markov model are obtained, and kanji that are not in the database Is
You may output as a phoneme sequence or a kana sequence.

Further, the feature extraction unit 2, the recognition unit 3, and the recognition result output unit 6 can be processed by a dedicated or combined microprocessor.

[0029]

As described above, according to the present invention, the pseudonym
Using a statistical language model for the order of occurrence of Chinese characters, a syllabic standard pattern of Hidden Markov Models, and a phonetic chain standard pattern for reading of Chinese characters, when outputting a series of mixed Kana and Kana characters directly from the input speech, only the characters are output. By creating a statistical language model in advance using information on the part-of-speech occurrence order, it is possible to statistically delete grammatically incorrect candidates and reduce processing time.
The conversion performance can be improved.

[Brief description of drawings]

FIG. 1 is a diagram for explaining the principle of the present invention.

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

FIG. 3 is a diagram illustrating a recognition process according to an embodiment of the present invention.

FIG. 4 is a diagram for explaining a conventional Japanese speech recognition system.

[Explanation of symbols]

1 voice signal input terminal 2 feature extraction unit 3 recognition unit 4 standard pattern memory 5 character / part-of-speech statistical language model 6 recognition result output unit 10 learning text database 11 statistical language model information about kana / kanji occurrence order 20 learning Speech database 21 Hidden Markov model phoneme / standard pattern information 30 Phoneme candidate selection unit 40 Recognition unit 50 Statistical language model memory 60 Phoneme sequence memory corresponding to reading characters 100 Text database for learning 101 Character occurrence sequence information 102 Part of speech occurrence order Information 200 Speech database for learning 201 Hidden Markov model syllable standard pattern information 202 Character reading standard pattern information 300 Statistical language model 400 Syllable standard pattern

Claims (2)

[Claims] 1. A plurality of speech recognition candidates for a characteristic parameter time series of the input speech are obtained by using an input speech as a time series of characteristic parameters and using a statistical language model regarding an occurrence order created from a learning text database. For each of these speech recognition candidates, a candidate with a high total likelihood of occurrence likelihood and similar likelihood is selected by comparing the standard pattern of the hidden Markov model with the time series of the characteristic parameters of the input speech. In the Japanese speech recognition method in which the recognition result is, a statistical language model relating to the order of occurrence of characters and the order of occurrence of parts of speech created from the learning text database is used as the statistical language model. Japanese speech recognition method. 2. A plurality of speech recognition candidates are selected for the characteristic parameter time series of the input speech by using the input speech as a time series of characteristic parameters and using a statistical language model regarding the occurrence order created from a learning text database. For each of these speech recognition candidates, a candidate with a high total likelihood of occurrence likelihood and similar likelihood is selected by comparing the standard pattern of the hidden Markov model with the time series of the characteristic parameters of the input speech. In the Japanese speech recognition method using as a recognition result, a syllabic standard pattern and a syllable standard pattern for reading a character are used as the standard pattern.