JPH0612091A - Japanese speech recognizing method - Google Patents

Abstract

PURPOSE:To convert an input speech to a character string mixed with KANA (Japanese syllabary) and KANJI (Chinese characters) at the high recognition ratio. CONSTITUTION:Concerning the first character of each word, similarity between the input speech and the standard pattern of phonemes is obtained by using Japanese clause syntax 6 and a word dictionary 7 showing the reading, tolerance to first generate that character is obtained, and the sum of the similarity and the tolerance is defined as total tolerance. While remaining characters having the high total tolerance, the second character having the high sum of the similarity and tolerance is similarily searched and concerning the obtained word having the high total tolerance, the similarity with the input speech and the tolerance of the generation are respectively obtained from the structure of noun and postpositional particle concerning the respective candidates of the postpositional particle coming next to the noun and afterwards, the obtained clause having the high total tolerance as a whole is obtained and defined as the recognized result.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention is applicable to, for example, a hidden Markov model (for example, Seiichi Nakagawa "Speech recognition by a probabilistic model" edited by The Institute of Electronics, Information and Communication Engineers (1988)), a statistical language model (for example, Bahl, L. et al. "A Statist
iCal Approach to Continuo
us Speech Recognition "IEE
E Trans. on PAMI (1983)) and a Japanese statistical model (for example, Yamada et al., "Statistical Language Model Using Kana-Kanji Character Chain Information," IEICE Technical Report, SP91-26 (19).
91)) and a Japanese speech recognition method.

[0002]

2. Description of the Related Art As a conventional speech recognition method using a hidden Markov model and a statistical language model, a statistical language model regarding a phoneme occurrence order and a phoneme standard pattern of the hidden Markov model are previously stored in a learning text database. For the input speech, a statistical language model is used to select a plurality of phoneme candidates that have a high probability of occurring next from the plurality of phonemes that have already been recognized, and each of the selected phoneme candidates is selected. It has been proposed that the phoneme standard pattern is collated with the input speech, and the phoneme with the highest total likelihood of the occurrence likelihood and the similarity with the standard pattern is output as the recognition result.

[0003]

However, in this recognition method, since the recognition result is output as a sequence of phoneme units,
To output the input speech as a Japanese sentence, it is necessary to convert the phoneme-based sequence of the recognition result into kana-kanji conversion. That is, since the input speech is converted into a sequence of phoneme units and the phoneme sequence is converted into a kana-kanji sequence, the conversion performance is relatively low in that a correct conversion result can be obtained as a whole. I have no choice.

[0004]

According to the present invention, a statistical language model relating to the occurrence order of kana and kanji whose readings are changed is created from a learning text database,
In addition, bunsetsu syntax, word dictionary with pronunciation, phoneme conversion rules for converting characters with pronunciation into phoneme symbols, and phoneme standard patterns are created in advance, and recognition candidates derived from this syntax and word dictionary are created. Using the statistical model of Kana-Kanji and the phoneme standard pattern, the input voice is converted into a mixed Kana-Kanji character sequence.

[0005]

FIG. 1 shows an embodiment of the present invention. Input terminal 1
The voice input from is converted into a digital signal in the feature extraction unit 2, further subjected to LPC cepstrum analysis, and then converted into a feature parameter every one frame (one time point, for example, 10 milliseconds). This characteristic parameter is, for example, an LPC cepstrum coefficient.

From the learning voice database, a phoneme standard pattern of the hidden Markov model is created in the same format as the above-mentioned characteristic parameters, stored in the standard pattern memory 4, and from the learning text database in which the reading of the characters is given, A statistical language model relating to the order of occurrence of read characters is created and stored in the statistical language model memory 5. Similarly, from the above learning text database, a bunsetsu syntax 6 in which the structure of a Japanese bunsetsu is described as a word transition rule, and a word dictionary 7 with reading (for example, FIG. 2).
A), a character-phoneme conversion rule 8 (for example, FIG. 2B) that converts kana and kanji with pronunciations into phoneme symbol strings is created and stored. It is to be noted that the pronunciation of the pronunciation with respect to the kana is, for example, “ha” is either “ha” or “wa”. In the recognition unit 3, a plurality of characters selected by using the clause syntax 6 and the word dictionary 7 with the pronunciation For a candidate, the standard pattern of phonemes obtained by the character-phoneme conversion rule 8 of the character candidate is read from the standard pattern memory 4, and the similarity (likelihood) of the characteristic parameter of the input speech is obtained. Also, the likelihood of occurrence of the above character candidates is obtained from a statistical language model.

That is, for example, in order to recognize the i-th character of the input speech, a trigram concerning the appearance order of the characters (kana and kanji) whose pronunciation is changed from the statistical language model is used to identify the (i-2) -th character. Based on the recognition result of each of the i-1) th character, the occurrence likelihood of the i-th character that appears is calculated for each character, and the likelihood indicating the similarity with the standard pattern of the phoneme sequence of that character and its The sum of the likelihood of occurrence of a character is the total likelihood.

The selection of kana / kanji candidates with this reading, the matching with the standard patterns for them, and the operation of obtaining the recognition result character from the total likelihood thereof are repeated until the end of the voice section, and finally, that The recognition result characters obtained up to this point are sent to the recognition result output unit 9, and are output in that order as a kana and kanji sequence. Next, a specific example of the recognition procedure will be described. For example, for the first character of each word in the word dictionary 7, the similarity between the phoneme sequence of the input speech and the standard pattern is obtained. That is, 1. For example, the similarity likelihood P1 is obtained by matching the standard pattern of / tou / with the input voice from the first character east of the word in the word dictionary 7. First, the total likelihood Q1 = P1 + L1 of the eastern part is obtained from the likelihood L1 of the character eastal part. 2. From the first character mountain of the word in the word dictionary 7, a standard pattern of / yama / is matched with the input voice to obtain the similarity likelihood P2. First, the mountain
Than the likelihood L2 that occurs in the mountain total likelihood Q2 =
Obtain P2 + L2. 3. In this way, the total likelihood of each word with the first character is obtained, and, for example, 10 pieces are left in order of high total likelihood, and the second character of that word, for example, the next character in the east, is set to K ( Today, the standard pattern of / kyou / and the input voice are matched to obtain the similarity likelihood P3. The total likelihood Q3 = Q1 + P3 + L3 of the eastern Kyo is obtained from the likelihood L3 that the Kyo occurs after the character east. 4. Take the shape as the next character of the mountain / gat
The similarity pattern P4 is obtained by matching the standard pattern of a / with the input voice. From the likelihood L4 that the shape (gata) occurs after the character mountain (yama), the overall likelihood Q4 = Q2 + P4 + L4 of the mountain shape (gata) is obtained. 5. In this way, one with a high overall likelihood is obtained, and from the relationship of the structure of the noun + particle in Japanese syntax 6, the candidate for that particle is obtained from the dictionary.
The similarity pattern P5 is obtained by matching the standard pattern of a / with the input voice. From the likelihood L5 that (wa) occurs after the letter east Kyoto, the total likelihood of the sentence east (to) Kyoto is (wa) Q5 = Q3 + P5
Get + L5. 6. Similarly, the similarity pattern P6 is obtained by matching the standard pattern of / no / of the candidate and the input voice.
From the likelihood L6 that occurs after the character Tokyo, we obtain the overall likelihood Q6 = Q3 + P6 + L6 of the phrase Tokyo. 7. Similarly, the candidate is better than / ga /
Get 7. From the likelihood L7 that a character appears after a mountain shape, the Bunshi mountain shape is
To obtain the total likelihood Q7 = Q4 + P7 + L7. 8. Moreover, the similarity likelihood P8 with / no / is obtained. From the likelihood L8 that occurs after the character mountain shape, the total likelihood Q8 = Q4 + P8 + L8 of the mountain shape is obtained. 9. Candidates for the bunsetsu obtained in this way are in the eastern Kyo (wa) eastern Kyo (mountain) shape. The overall likelihoods Q5, Q6, Q7, and Q8 of are compared. 10. For example, if Q6 is the highest, the recognition result will be East, and Tokyo will be output as the recognition result. The feature extraction unit 2, the recognition unit 3, and the recognition result output unit 9 can be processed by a dedicated or shared microprocessor.

[0009]

As described above, according to the present invention,
High recognition performance is expected by using a statistical language model for the appearance order of kana and kanji with pronunciation and phoneme standard patterns corresponding to the pronunciation. In addition, the input voice can be converted into a kana-kanji series at once, and higher conversion performance than that in the case of conversion in two steps is expected.

When the conversion rate is evaluated for 500 syllables uttered in syllable units, the phoneme recognition is performed using the syllable statistical information (syllable trigram) by the conventional method, and then the phoneme string recognition is performed. When the kana / kanji conversion was performed on the result, 70% (correct phrase / total phrase × 100) was correctly converted. On the other hand, when the method of the embodiment is used, the conversion performance is 94% due to the effect of kana / kanji statistical information with reading (trigram of kana / kanji group with reading) even with the same phoneme recognition rate. Improve to. Moreover, the processing speed is increased 10 times by using the clause syntax.

The present invention is not limited to the above embodiment. For example, the unit of the recognition standard pattern may be not only a phoneme but also a syllable or a continuous syllable. The recognition method is not limited to the hidden Markov model, and DP matching may be used. The statistical language model is not limited to the trigram, but may be a bigram statistic.

[Brief description of drawings]

FIG. 1 is a block diagram showing an example of an apparatus that implements a voice recognition method according to the present invention.

FIG. 2A is a diagram showing a description example of a word dictionary 7 in FIG.
FIG. 3 is a diagram showing an example of a character phoneme conversion rule 8 in FIG. 1.

 ─────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Kiyohiro Kano 1-1-6 Uchisaiwaicho, Chiyoda-ku, Tokyo Nihon Telegraph and Telephone Corporation

Claims (1)

[Claims] 1. The input speech is a time series of characteristic parameters, and a statistical language model relating to the occurrence order created from a Japanese syntax, a word dictionary, a graphoneme conversion rule, and a text database is used to input the input speech. For the characteristic parameter time series, a plurality of speech recognition candidates are selected, and for each of these speech recognition candidates, the standard pattern is compared with the characteristic parameter time series of the input speech, respectively, and the likelihood of occurrence and the likelihood of similarity are calculated. In the Japanese speech recognition method in which the candidate with a high total likelihood of is used as the recognition result, as the above-mentioned statistical language model, a kana with a read pronunciation, which is created from a text database with a read kanji,
, And a statistical language model for the occurrence order of kanji, a grammar that describes the structure of Japanese bunsetsu as word transition rules as the Japanese syntax, and a kana with kana and kanji with the reading as the word dictionary. A set of words in a sequence is used, the phoneme standard pattern is used as the standard pattern, using a kana with a reading and a rule for converting kanji into a phoneme symbol string as the character phoneme conversion rule. Japanese speech recognition method.