JP2003177779A - Speaker learning method for speech recognition - Google Patents

Abstract

(57) [Summary] [Problem] In conventional speaker adaptation and speaker registration learning to improve the performance of speakers with low recognition performance, does the speaker become burdensome because the amount of utterance for learning increases? If the amount of utterance is limited to reduce the burden, the recognition performance may not be improved in all utterances, and there is a possibility that a word with a lower recognition rate may appear. Yes. By estimating whether the utterance content depends on the recognition result with less utterance, performing speaker adaptive learning if not dependent, and speaker registration learning if dependent, With a learning utterance that is not burdened by the speaker,
It is possible to provide a speaker learning method that can surely improve the recognition rate.

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD The present invention relates to a speaker learning method in speech recognition.

[0002]

2. Description of the Related Art A conventional speaker learning method will be described below. In the conventional unspecified speaker speech recognition system, a standard acoustic model that can handle as many unspecified speakers as possible is constructed and used, but in practice, the utterance characteristics of the speaker are diverse, It is difficult to learn an acoustic model that guarantees high performance for all speakers. Therefore, conventionally, speaker adaptation that guarantees performance for all speakers by re-learning the acoustic model parameters of the unrecognized speaker using the speaker's own utterance and reconstructing the acoustic model adapted to the speaker. I am taking means. This speaker adaptation requires a large amount of learning speech to capture the characteristics of the speaker,
Since it imposes a burden on the speaker, various measures have been taken to limit the minimum number of times of vocalization (for example, Japanese Patent No. 2037877).
On the other hand, as another learning method, it is possible to add an acoustic model sequence corresponding to the recognition result of an erroneously recognized word to the pronunciation dictionary as a correct answer sequence and recognize what is recognized as an incorrect sequence as a correct sequence. There is also a person registration method (Japanese Patent Laid-Open No. 8-171396).

[0003]

The conventional speaker adaptation method is as follows.
In principle, this method can certainly improve the recognition performance if there is sufficient learning data.However, as is the case with almost all practical systems, the number of utterances is narrowed in consideration of the learning burden on the speaker. In that case, there is a problem that the recognition rate may be reduced on the contrary for some utterances that do not exist in the learning data. On the other hand, the conventional speaker registration method surely improves the recognition rate of learned utterances, but in the case of a speaker that is difficult to recognize due to a large amount of utterance content, all utterances that are difficult to recognize during learning must be used. There is a problem that learning is burdensome.

An object of the present invention is to solve the problems of conventional speaker adaptation learning and speaker registration learning, and to improve the recognition rate after learning with a learning voicing amount that does not burden the speaker. It provides the law.

[0005]

In order to solve the above-mentioned problems, the speaker learning method according to any one of claims 1 to 5 re-learns acoustic model parameters using a speaker's learning voice,
A means to create an acoustic model adapted to the speaker, a means to add the acoustic model sequence corresponding to the recognition result of the misrecognized word to the pronunciation dictionary as a correct sequence, and whether the recognizability depends on the utterance content. And a means for determining.

[0006]

BEST MODE FOR CARRYING OUT THE INVENTION A speaker learning method according to claims 1 to 5 of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram of a speaker learning method according to claims 1 to 5 of the present invention.

In the speaker learning function set so that each speaker selects when he / she needs to improve the recognition performance for himself / herself, the system first prompts the user to speak a specific word to identify the speaker. The word utterance is input. This utterance content is the minimum content necessary for each speaker to judge how appropriate the standard speech prepared in advance is, and in the case of Japanese recognition, for example, the word " Contents such as "microphone test" are appropriate. When the system recognizes words, a plurality of words may be selected from the target words so that all five vowels are included.

A normal recognition process is performed on this utterance in the voice recognition process 1, and a recognition result and a recognition reliability score are calculated in the recognition score calculation process 2. As for the recognition result, the phoneme or syllable sequence of the recognition result is compared with the correct phoneme sequence, and the difference is regarded as an error and the matching part is regarded as the correct answer, and the correctness is recorded for each phoneme of the correct answer sequence. . The reliability score is, for example, an acoustic distance score for each phoneme or syllable between the correct phoneme or syllable sequence and the uttered result, and when the weighted cepstrum distance is used as the distance measure, the reliability of each phoneme is calculated. May be calculated by Equation 1.

[0010]

[Equation 1]

In the learning method determination process 3, a phoneme or syllable (referred to as an adaptive candidate phoneme or syllable) which is erroneously recognized even if the reliability score is equal to or lower than the threshold value or equal to or higher than the threshold value.
Calculate the ratio of phonemes or syllables included in all utterances of. If this ratio is large, it is estimated that the speaker's utterance features do not apply to the standard voice, regardless of the utterance content, and it is necessary to learn to apply all standard voices to the speaker. To be If this ratio is small, the misrecognition depends on the utterance content, and although the utterance feature of the speaker and the standard voice are applied, it is considered that learning is necessary only for a specific utterance. Therefore, if this ratio is a certain value or more, speaker adaptive learning is selected, and if it is less than a certain value, speaker registration learning is selected.

When the speaker adaptation learning is selected, the speaker adaptation processing 4 prompts the user to make a minimum necessary utterance for further adaptation. In the speaker adaptation method, for example, when the VFS method described in Japanese Patent Laid-Open No. 5-53599 is used, a standard acoustic model is matched with a learning input speech parameter, and a fuzzy class function is obtained from the relationship between the corresponding parameters. , The parameters of the standard acoustic model are updated so that the standard speech approaches the learning input speech with the obtained function as a weight.

If speaker registration learning is selected,
In the speaker registration process 5, the utterance of only the word including the adaptive candidate phoneme or syllable calculated in the learning determination process is urged, and the phoneme sequence of the utterance is added to the phoneme sequence of the word including the phoneme sequence corresponding to the adaptive candidate phoneme. The syllable recognition result sequence is added to the pronunciation dictionary 7. For example, if the word "menu" causes erroneous recognition, it is assumed that only this word is uttered and the recognition result is "denu". If you are using a phoneme model as an acoustic model, click "Menu".
The correct phoneme model sequence of is / me ny uu /, and the recognition result phoneme sequence is / d eny uu /. For this speaker, it can be seen that the phoneme / m /, which is the beginning of a word and is followed by / e /, tends to be mistaken for / d /. So, in the recognition target word,
/ M /, which is the beginning of a word and the next is / e /, is mistaken as / d /
Add a phoneme sequence to the pronunciation dictionary to recognize it as m /.
In the case of this example, the menu originally "menu / men
Add "/ de ny uu /" instead of "y uu /" and change the dictionary to "menu / me ny uu / or / de ny uu /". This will cause this speaker to drop the "menu"
Even if it is recognized as uu /, the result is that the "menu" can be recognized.

As described above, it is estimated whether or not the utterance of the speaker is erroneous without depending on the utterance content, and if it does not depend on the utterance content, speaker adaptive learning is performed, and if so, speaker registration learning is performed. Thus, the problem that the recognition rate decreases in the conventional speaker adaptation learning despite many learning utterances can be solved by performing speaker registration learning instead of speaker adaptation learning. it can. In addition, the problem that cannot be learned without uttering many words in the conventional speaker registration learning can be solved by performing speaker adaptation learning instead of speaker registration learning.

[0015]

As described above in detail, the speaker learning method according to the first aspect of the present invention enables speaker adaptive learning to be performed depending on the recognizability of each speaker and the degree of dependence of the utterance content. By deciding whether to perform learning or speaker registration learning, and encouraging the speaker to learn either, recognition rate in the conventional speaker adaptation learning despite many vocalizations for adaptation Can be solved by automatically selecting speaker registration learning instead of speaker adaptation learning. Further, in the conventional speaker registration learning, a problem that cannot be learned without uttering many words can be solved by automatically selecting speaker adaptation learning instead of speaker registration learning. Therefore, the present invention provides a speaker learning method that can surely improve the recognition rate with a learning amount that does not burden the speaker.

As described in detail above, the speaker learning method according to the second aspect of the present invention determines that the easiness of recognition depends on the utterance content. By calculating the recognition score for the minimum possible training utterance and determining whether it depends on the score's height, it is possible to use many learning utterances for adaptation in conventional speaker adaptation learning. The problem of a low recognition rate can be solved by automatically selecting speaker registration learning instead of speaker adaptation learning. Further, in the conventional speaker registration learning, a problem that cannot be learned without uttering many words can be solved by automatically selecting speaker adaptation learning instead of speaker registration learning.
Therefore, the present invention provides a speaker learning method that can surely improve the recognition rate with a learning amount that does not burden the speaker.

As described in detail above, the speaker learning method according to the third aspect of the present invention determines whether or not the recognizability depends on the utterance content, and when it is determined that the recognizability depends on the utterance content. Performs speaker registration learning, and if it is determined that it does not depend on the speaker adaptation learning, the recognition rate is increased in the conventional speaker adaptation learning, even though many learning utterances are made for adaptation. Can be solved by automatically selecting speaker registration learning instead of speaker adaptation learning. Further, in the conventional speaker registration learning, a problem that cannot be learned without uttering many words can be solved by automatically selecting speaker adaptation learning instead of speaker registration learning. Therefore, the present invention provides a speaker learning method that can surely improve the recognition rate with a learning amount that does not burden the speaker.

As described above in detail, in the speaker learning method according to the fourth aspect of the present invention, the recognition score is calculated based on the accuracy of the recognition result, the distance value from the standard voice, or the reliability of the distance value on the left. In the conventional speaker adaptive learning, the problem that the recognition rate decreases even though many learning utterances for adaptation are calculated by calculating them individually or in combination, instead of the speaker adaptive learning. The problem can be solved by automatically selecting person registration learning. Further, in the conventional speaker registration learning, a problem that cannot be learned without uttering many words can be solved by automatically selecting speaker adaptation learning instead of speaker registration learning. Therefore, the present invention provides a speaker learning method that can surely improve the recognition rate with a learning amount that does not burden the speaker.

As described above in detail, in the speaker learning method according to the fifth aspect of the present invention, the recognition score is calculated as the correctness / incorrectness result of the recognition result, the distance value from the standard voice, or the reliability of the distance value on the left. In the conventional speaker adaptive learning, the problem that the recognition rate decreases even though many learning utterances for adaptation are calculated by calculating them individually or in combination, instead of the speaker adaptive learning. The problem can be solved by automatically selecting person registration learning. Further, in the conventional speaker registration learning, a problem that cannot be learned without uttering many words can be solved by automatically selecting speaker adaptation learning instead of speaker registration learning. Therefore, the present invention provides a speaker learning method that can surely improve the recognition rate with a learning amount that does not burden the speaker.

[Brief description of drawings]

FIG. 1 is a block diagram of a speaker learning method according to an embodiment of the present invention.

[Explanation of symbols]

1 voice recognition 2 Recognition score calculation 3 Learning method decision 4 speaker adaptation 5 Speaker registration 6 acoustic models 7 Pronunciation dictionary 8 recognition score buffer

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Shinichi Yoshizawa             1006 Kadoma, Kadoma-shi, Osaka Matsushita Electric             Sangyo Co., Ltd. F-term (reference) 5D015 AA02 AA03 GG01 GG04 GG05                       GG06

Claims (5)

[Claims] 1. A means for re-learning acoustic model parameters using a speaker's learning voice (hereinafter referred to as speaker adaptive learning), and a means for erroneously recognizing a word. It has means for adding an acoustic model sequence corresponding to the recognition result as a correct answer sequence to the pronunciation dictionary (hereinafter referred to as speaker registration learning), and means for determining whether the recognizability depends on the utterance content. , It is characterized that the speaker adaptive learning or speaker registration learning is selected according to the recognizability of each speaker and the dependence of the utterance content, and the speaker is urged to learn either of them. Speaker learning method. 2. The speaker learning method according to claim 1, wherein
The means to judge whether the easiness of recognition depends on the utterance content calculates the recognition score for the minimum learning utterance that can be judged to be dependent, and determines whether it depends from the height of the score. A speaker learning method characterized by that. 3. The speaker learning method according to claim 1, wherein
As a result of judging whether the recognizability depends on the utterance content, if it is judged to be dependent, speaker registration learning is performed, and if it is judged not to be dependent, speaker adaptation learning is performed. Speaker learning method. 4. The recognition score in the speaker learning method according to claim 2, is calculated by using the accuracy of the recognition result, the distance value to the standard voice, or the reliability of the distance value on the left, respectively, alone or in combination. Speaker learning method characterized by. 5. The recognition score in the speaker learning method according to claim 2, is calculated by using the accuracy of the recognition result, the distance value from the standard voice, or the reliability of the distance value on the left, respectively, alone or in combination. Speaker learning method characterized by.