JPH0719158B2 - Speaker adaptation method of phoneme transformation rule - Google Patents

Description

DETAILED DESCRIPTION [Outline] In order to deal with phoneme transformation in continuous speech, a phoneme created by using a phoneme transformation rule given at least a phoneme transition network of sentences, words, and morphemes. In the recognition device, a speaker-specific phoneme transformation rule dictionary of a large number of speakers is connected to the phoneme network creation unit, only learning data is registered in the word dictionary, and an unknown speaker recognizes the voicing result for the learning data. When performing, the frequency of use of the phoneme transformation rules used at this time is counted for each phoneme transformation rule dictionary of the large number of speakers, and the phoneme transformation rule dictionary of the specific speaker with high frequency of use is integrated, This is a phoneme transformation rule dictionary peculiar to the learning speaker.

[Industrial application field]

The present invention relates to a speech recognition device, and more particularly to a speaker adaptation method of phoneme transformation rules for learning phoneme transformation rules peculiar to each speaker without registering enormous speech data for learning phoneme transformation rules.

The phoneme transformation rule in phonetics is a rule that shows how each speaker transforms when multiple phonemes that make up a voice are uttered. There is a phoneme transformation rule peculiar to each speaker, but conventionally, a method of automatically learning the phoneme transformation rule peculiar to each speaker has not been developed. The phoneme transformation rule is visually extracted from the learning (registered) voice data of.

It is also conceivable to construct a phoneme transformation rule dictionary that can deal with unspecified speakers, but this method is not realistic because the number of rules is large and the processing time becomes huge, and it is effective for a certain speaker. Some phoneme transformation rules are harmful to another speaker, and are generally inferior to individual phoneme transformation rules in terms of identification performance.

At present, it takes a long time before a high-performance, high-speed voice recognition device can be used, which is inconvenient for the user, and the labor of the designer is great.
As a major obstacle to the spread of voice recognition devices, there has been a demand for a method of effectively generating phoneme transformation rules for unknown speakers.

[Problems to be solved by conventional technology and invention]

FIG. 3 is a diagram showing an example of a conventional general voice recognition device.

First, the phoneme transformation rules stored in the common phoneme transformation rule dictionary and the specific phoneme transformation rule dictionary will be described.

As described above, the phoneme transformation rule is a rule (rule) showing how to transform a plurality of voices constituting each voice when a speaker utters the voices. )
There are two types of rules, which are common to each speaker and rules specific to each speaker.

The rules common to each speaker are widely known as phonetics for each language (Japanese, English, etc.), but the rules specific to each speaker need to be learned from the actually uttered voice data. is there.

* Examples of common rules for speakers: i The vowels a or u sandwiched between unvoiced consonants may be devoiced.

ii / G / in the word can be nasal.

iii Vowels adjacent to nasal consonants can become nasalized.

iv Consonants followed by the vowel a are palatized.

* Examples of speaker-specific rules: v The consonant whose articulatory point is on the lip, followed by the vowel c, is rounded.

vi / G / does not make nasal noise.

vii Nasal sounds do not palate.

Etc.

In the voice recognition device, the phoneme transformation rule peculiar to the speaker is used in preference to the rule common to the speakers.
For example, in the example rule above, vi has priority
ii will be ignored.

Next, the learning method of the phoneme transformation rule will be described.

The voice recognition device needs to learn phoneme transformation rules specific to a voice utterer prior to voice recognition. Conventionally, the speech data for learning uttered by an unknown speaker is read, the phoneme transformation rules peculiar to the speaker are extracted, and the peculiar phoneme transformation rule dictionary is created visually.

Next, a voice recognition method in such a voice recognition device will be described.

When the above phoneme transformation rule is completed, a voice recognition device,
For example, in a device for recognizing a prefecture name, each character string (that is, a prefecture name) is read from a word dictionary storing all the prefecture names, and the common phoneme transformation rule dictionary Phoneme Transformation Rule Based on the above, the phoneme transition network shown in FIG. 4 is created.

In the example of this figure, a general word is taken as an example,
It is shown that when the above phoneme transformation rule is applied to a word of a character string called'ASITAGA '(tomorrow ga), a phoneme transition network as shown in the figure is obtained.

The phoneme transition network creation unit of FIG. 3 creates a phoneme transition network by using the above phoneme transformation for all character strings included in the word dictionary.

When the phoneme transition network created in this way is sent to the phoneme transition network verification unit, the voice of the specific speaker input from the microphone is decomposed into each phoneme by the voice analysis unit, and any path of the phoneme transition network is If the match is verified and accepted, the character string data of the word corresponding to the matched phoneme transition network is read from the word dictionary, and the character string notation is output as the recognition result.

Therefore, in the conventional method, the speaker-specific phoneme transformation rule dictionary is read by reading the voice data for each speaker, the phoneme transformation rule unique to each speaker is visually extracted, and the phoneme transformation rule dictionary is manually extracted. Since it was created, the labor for creating the specific phoneme transformation rule dictionary becomes enormous,
This has been a major obstacle to the spread of voice recognition devices.

In view of the above conventional drawbacks, the present invention has an object to provide a method for realizing a practical voice recognition device by providing a method for automatically learning phoneme transformation rules peculiar to each speaker. It is a thing.

[Means for solving problems]

FIG. 1 is a block diagram of the principle of the speaker adaptation method of the phoneme transformation rule of the present invention.

In the present invention, all the phoneme transformation rule dictionaries' of a large number of speakers (N persons) are used, and a plurality of phoneme transformation rule learning speech data (stored in the word dictionary) whose utterance contents are known to an unknown speaker are used. The phoneme transition rule verification frequency used for recognition by the phoneme transition network verification unit is counted for each specific speaker dictionary ′.
It is configured such that a few to a dozen or more of the above-mentioned specific speaker dictionaries' are integrated to form the unknown speaker phoneme transformation rule dictionary.

[Action]

That is, according to the present invention, in order to deal with phoneme deformation in continuous speech, a phoneme recognition device that creates at least a phoneme transition network of sentences, words, and morphemes using a phoneme deformation rule given in advance. In, a speaker-specific phoneme transformation rule dictionary of a large number of speakers is connected to the phoneme network creation unit, only learning data is registered in the word dictionary, and when recognizing a voicing result for the learning data by an unknown speaker. , The frequency of use of the phoneme transformation rules used at this time is counted for each of the phoneme transformation rule dictionaries of the large number of speakers, and the phoneme transformation rule dictionaries of the specific speakers with high frequency of use are integrated to obtain the learning speaker. The phoneme transformation rule dictionary peculiar to the speaker is used, so that the phoneme transformation rules peculiar to the speaker can be automatically learned from a relatively small amount of registered speech data. There is an effect that it is possible to provide a location.

〔Example〕

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

FIG. 2 is a block diagram showing one embodiment of the present utterance, and the dictionary usage frequency counter, the dictionary selection switch control unit, and the dictionary selection SW are the functional blocks necessary for implementing the present invention. The same reference numerals indicate the same objects throughout the drawings.

The speaker adaptation method of the phoneme transformation rule of the present invention will be described below with reference to FIG. 1 and FIG.

First, when learning phoneme transformation rules peculiar to a speaker,
A speaker-specific phoneme transformation rule dictionary'of a large number of speakers is connected to the phoneme transition network creating unit, only word names of learning data are registered in the word dictionary, and the learning data is recognized.

When performing the learning with words, recognition is performed for each word, but since the phoneme transition network for the word has been generated in advance by the phoneme transition network creation unit, the microphone utters the word, When it is analyzed in phoneme units in the voice analysis unit and input to the phoneme transition network verification unit, the verification unit accepts it in any path of the network.

If it is not accepted, a new rule is required, and the phoneme transformation rule is manually registered by the conventional method, or the learning data is ignored, and the process of recognizing the next learning data is started.

In the present invention, since a dictionary usage frequency counter is provided for each speaker's peculiar phoneme transformation rule dictionary ', there is a phoneme transformation rule involved in the creation of the accepted phoneme transition network path. When adopted from the speaker's peculiar phoneme transformation rule dictionary ', the corresponding counter is counted up.

When the same phoneme transformation rule exists in the peculiar phoneme transformation rule dictionary 'of a plurality of speakers, it functions to count up the counters of the peculiar phoneme transformation rule dictionary' of all speakers in which the phoneme transformation rule exists.

When the processing of all the learning data is completed, the count value of the dictionary usage frequency counter provided corresponding to the phoneme transformation rule dictionary of all the speakers is referred to, and each of the M speakers having a large value is referred to. Dictionary selection of the peculiar phoneme transformation rule dictionary 'by the control signal of the dictionary selection switch control unit
It is integrated by operating the SW to form the learning speaker's peculiar phoneme transformation rule dictionary. The value of M may be given in advance by the device designer, or may be specified by the operator using a keyboard, a switch, or the like.

When the learning for creating the phoneme transformation rule for the unknown speaker is completed, the normal speech recognition operation is performed by the phoneme transformation rule dictionary integrated as described above.

As described above, in the present invention, when learning data whose content is known in advance is recognized by an unknown speaker, which of a large number of phoneme transformation rule dictionaries for specific speakers is used is counted. However, the feature is that the phoneme transformation rule dictionaries that are frequently used are selected and integrated to be used as the phoneme transformation rule dictionary of the unknown speaker.

〔The invention's effect〕

As described above in detail, the speaker adaptation method of the phoneme transformation rule of the present invention uses at least sentences, words, and phoneme transition networks of morphemes to deal with phoneme transformation in continuous speech. In a speech recognition device created using a given phoneme transformation rule, a speaker-specific phoneme transformation rule dictionary of many speakers is connected to a phoneme network creation unit,
Only the learning data is registered in the word dictionary, and the frequency of use of the phoneme transformation rule used at this time when recognizing the utterance for the learning data by the unknown speaker is determined by the phoneme transformation rule dictionary of the above-mentioned multiple speakers. The phoneme transformation rule dictionary of a speaker that is frequently used is integrated into a phoneme transformation rule dictionary that is unique to the learning speaker. The rules can be automatically learned from a relatively small amount of registered voice data, and a practical voice recognition device can be provided.

[Brief description of drawings]

FIG. 1 is a block diagram showing the principle of a speaker adaptation system for phoneme transformation rules according to the present invention, FIG. 2 is a block diagram showing an embodiment of the present invention, and FIG. 3 is an example of a general speech recognition apparatus. , And Fig. 4 are examples of phoneme transition networks. In the drawing, is a microphone, is a voice analysis unit, is a phoneme transition network verification unit, is a phoneme transition network creation unit,
Indicates a peculiar phoneme transformation rule dictionary, a common phoneme transformation rule dictionary, a word dictionary, a dictionary selection switch, a dictionary selection switch control unit, and a dictionary usage frequency counter.

Claims (1)

[Claims] 1. A number of speech recognition devices, which deal with phoneme deformation in continuous speech, using at least a phoneme transition network of a sentence, a word, and a morpheme using a given phoneme deformation rule. Using all the phoneme transformation rules dictionary for specific speakers of the above, recognizing a plurality of rule-learning speech data in which the utterance content of an unknown speaker is known, the frequency of use of the phoneme transformation rules used at that time is calculated. For each speaker-specific phoneme transformation rule dictionary for a specific speaker, a dictionary frequency counter is used to count, and several to a dozen or more phoneme transformation rule dictionaries for a specific speaker are integrated, Speaker adaptation method of phoneme transformation rules, which is a dictionary for phoneme transformation rules for mobile phones.