JPH07113838B2 - Speech recognition method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition method for voice recognition of an unspecified speaker.

[0002]

2. Description of the Related Art Conventionally, in order to create a standard pattern for speech recognition of an unspecified speaker, data in which a large number of speakers have uttered recognition target words is used in advance, and a human being visually recognizes a speech segment. Have been cut out and processed statistically. In this method, speech data actually uttered by hundreds of speakers is used to create a recognition word dictionary for unspecified speakers. Recently, it has become possible to create a standard pattern for recognizing a voice of an unspecified speaker by using data in which a recognition target word is uttered by one or several speakers. For example, a configuration described in "Unspecified speaker voice recognition method in which dynamic characteristics of word voice are modeled by utterance of minority speaker" (Electronic Communication Information Society SP91-20) is known.

FIG . 7 is a block diagram of this technique. In FIG. 7, 1 is an acoustic analysis unit, 2 is a characteristic parameter extraction unit,
3 is a similarity calculation unit, 4 is a standard pattern storage unit, 5 is a regression coefficient calculation unit, 6 is a parameter series creation unit, 7 is a recognition unit,
Reference numeral 8 is a dictionary storage unit.

The standard pattern storage unit 4 stores, for each of the n phonemes, the data uttered by many speakers in advance.
The temporal position (feature frame) that best expresses the feature of the phoneme is obtained, and the phoneme standard pattern created by using the time pattern of the feature parameter centered on this feature frame is stored. As the time pattern, the LPC cepstrum coefficients (C0 to C8) are calculated for several frames before and after the characteristic frame, a one-dimensionally arranged parameter series is obtained, and the average value vector and the element of each element of this parameter series are calculated. The covariance matrix between them is obtained and stored as a standard pattern.

The dictionary storage unit 8 analyzes the recognition target speech uttered by one or two or more speakers in advance and calculates the similarity for each of the above-mentioned n standard patterns and each frame. The time series of the degree vector and the time series of the n-dimensional regression coefficient vector obtained as the inclination of the similarity vector for each frame are all registered as a dictionary.

When a voice is actually input, the acoustic analysis unit 1 performs acoustic analysis, the characteristic parameter extraction unit 2 calculates an LPC cepstrum coefficient, and the similarity calculation unit 3 compares the standard pattern in the standard pattern storage unit 4. The similarity is calculated, the regression coefficient calculation unit 5 calculates a regression coefficient from the change in the similarity for each frame, and the obtained similarity and regression coefficient are calculated by the parameter series creation unit of the parameter series creation unit 6 in a one-dimensional manner. The time series is used, and the recognition unit 7 performs recognition using the dictionary of the dictionary storage unit 8.

In the recognition, a matching method called DP matching is used. An example of a recurrence formula for performing DP matching is shown in (Equation 1). Here, the dictionary length is J frames, the input length is I frames, the function of the distance between the i-th frame and the j-th frame is l (i, j), and the cumulative similarity is g (i, j).
And

[0008]

[Equation 1]

In the DP matching, the sum of products is calculated for all the frames of the dictionary and the input, so the number of calculations is very large.

[0010]

As described above, the conventional method described above has a problem that it takes a long time because the number of calculations for collation is large, and that the storage area for storing the dictionary is large. It was

The present invention is to solve the above-mentioned problems, and by reducing the number of parameter sequences registered in the recognition word dictionary while maintaining a high recognition rate, the storage area of the dictionary and the number of operations in recognition are reduced. Both of them are reduced to enable high-speed recognition processing and facilitate hardware implementation.

[0012]

In order to achieve this object, according to the present invention, the speech to be recognized is uttered by a small number of speakers, one to several, and is obtained every analysis time (frame). and characteristic parameters of the number (m is an integer), the standard pattern each n type created from pre large number of speakers (n is an integer)
Matching with m feature parameters of
Input speech to be recognized by calculating n similarity degrees for each frame, selecting N (N is an integer) smaller than n from the time-series pattern created by this n-dimensional similarity vector, and registering it as a word dictionary. Similarly, m characteristic parameters obtained by the same analysis and m characteristic parameters each of which has n standard patterns
, Which is registered in the dictionary by finding the time series of the n-dimensional similarity vector
It is configured to recognize the speaker who registered the recognition target voice and other input voices by collating with the time series of the dimensional similarity vector.

[0013]

According to the present invention, with the above configuration, n types of phonemes and syllables created by a large number of speakers are obtained for feature parameters obtained by analyzing voices uttered by a small number of speakers, one to a few. The similarity with the standard pattern is obtained for each frame. It takes the longest time for recognition by reducing the number of n similarities to a number N smaller than n, and reducing the regression coefficient obtained as a variation amount of the similarity from n to a number M smaller than n. The number of times the matching portion is calculated can be reduced by that much, and the capacity of the dictionary storage unit can be reduced by the amount that the similarity is reduced from n to N and the regression coefficient is similarly reduced from n to M. .

[0014]

EXAMPLE An example of the present invention will be described below, but before that, an outline of the present invention will be described.

[0015] 1 and characteristic parameters Several speakers obtained by analyzing the recognition target voice uttered from people, advance a number of matching for each and analysis time n types of standard patterns created (one frame) from the speaker Then, N effective feature quantities are further determined from the obtained time series of n-dimensional similarity vectors, and these are registered as a dictionary. The input speech to be recognized is also matched with n kinds of standard patterns, the time series of the obtained n-dimensional similarity vector is obtained, and the feature quantity of N similarity degrees of the dictionary is found from the obtained time series. Match only with. Also, as for the regression coefficient obtained as the time variation of the similarity, M effective feature amounts are selected from n and used. By performing the voice recognition of the unspecified speaker in this manner, the number of collation calculations, which was 2 × n times per word per frame, can be reduced to N + M times.

[0016] For example, in the case, such as to say "umbrella" (kasa), / in the a / parts, / a / and / o / similarity of vowel part, such as, but larger, on the contrary, / k / , / T
The similarity of the consonant parts such as / and / s / becomes extremely small, and takes a value that is negligibly smaller than the similarity of / a /. That is, when the similarity of the input speech to the standard patterns such as n kinds of phonemes and syllables created by many speakers for each frame is calculated, the similarity to the phonemes and syllables different from the input is extremely high. Is a small value, and it can be said that it does not contribute much to actual recognition. Therefore, it is not necessary to register an element having a low degree of similarity, that is, an element that is not involved in recognition as a constituent element of the dictionary.

By doing so, the number of times of calculation of the collation portion, which requires the longest time for recognition, is reduced, and the capacity of the dictionary storage unit is also reduced by the reduction of the similarity and the regression coefficient.

[0018] Hereinafter, an embodiment of the present invention in conjunction with FIG. In FIG. 1, 9 is an acoustic analysis unit, 10 is a characteristic parameter extraction unit, 11 is a similarity calculation unit, 12 is a standard pattern storage unit, 13 is a regression coefficient calculation unit, 14 is a parameter series creation unit, and 15 is a parameter selection unit. , 16 are recognition units, 1
Reference numeral 7 is a dictionary storage unit.

[0019] Figure 2 is a conceptual diagram illustrating a method of matching, 18 is the index portion, 19 similarity storage unit, 20 is a regression coefficient storage unit. Parameter selection unit 1
When creating a dictionary to be stored in the dictionary storage unit 5, 5 selects only valid ones. In this embodiment, 23 will be described later.
From the similarities corresponding to the types of phonemes, only N pieces having the largest value are selected and stored in the similarity section 19. Also, regarding the dictionary stored in the regression coefficient storage unit 20,
From the regression coefficients corresponding to the three types of phonemes, only M pieces are selected in descending order of absolute value and created.

In the dictionary storage unit 17, conventionally, only the similarity or the similarity and the regression coefficient are stored in the dictionary as shown in FIGS. 5 and 6, but in the present invention, FIGS.
As shown in, the index unit 18 is newly provided, and the similarity or regression coefficient corresponding to which phoneme is left and stored. Then, as shown in FIG. 2, the index unit 18 determines which phoneme to perform the matching calculation, and the matching calculation is performed only on the phonemes registered in the index unit 18. Then, the similarity is stored in the similarity storage unit 19 and the regression coefficient in the regression coefficient storage unit 20 is used for matching. The regression coefficient is the amount of change in similarity over time and is represented by the slope of the straight line shown in FIG.

The standard pattern storage unit 12 stores 23 types of phoneme standard patterns created in advance from data uttered by many speakers. In this embodiment, / a
/, / O /, / u /, / i /, / e /, / j /, / jv
/, / Ju /, / w /, / m /, / n /, / s /, / h
v /, / hu /, / p /, / t /, / k /, / c /, /
b /, / d /, / r /, /

[0022]

[Outer 1]

[0023] /, use the / z / of 23 types of phoneme standard pattern. However, it is assumed that the phoneme / hv / is voiced / h /, the phoneme / hu / is unvoiced / h /, the phoneme / jv / is a vocal sound followed by voiced consonants, and the phoneme / ju / is a vocal sound followed by unvoiced consonants. The phoneme standard pattern is created by visually detecting the feature part of each phoneme (the temporal position that best expresses the feature of the phoneme) and using the time pattern of the feature parameter centered on this feature frame. To do.

As described above, the recognition experiment was conducted on the data of 20 persons who uttered 212 words using the configuration of this embodiment.
One of each of the 20 people, male and female, is registered as a dictionary in which 212 words are uttered, and the remaining 18 uttered words are recognized. Of the 23 phonemes, only the five with the highest degree of similarity are retained, and the regression coefficient is also with the highest absolute value.
When recognition was performed using a dictionary created by leaving only the number of words, a recognition rate of 95.35% was obtained.

[0025] As a result of the recognition using the dictionary that was created using the 23 phonemes all that does not perform exactly the reduction is 95.64%
Therefore, the recognition rate is slightly reduced by about 0.3%. On the other hand, the storage capacity of the dictionary was 23 types corresponding to 23 phonemes for both similarity and regression coefficient per frame, but this was reduced to 5 and 5 indexes were added respectively. It will be October 23. Also, regarding the time required for recognition, the number of times of collation per frame decreases and becomes less than half. Further, even if a similar recognition experiment is performed with five to four similarity degrees and regression coefficients remaining, a high recognition rate of 95.30% is maintained.

[0026] However, in this embodiment, a C0~C8 the LPC cepstrum as the acoustic parameters, an example has been described when performing recognition using 23 phonemes as a standard pattern, the LPC cepstrum orders and phonemes There is no problem even if the number of is changed. Further, it is also possible to use syllables, voice pieces such as VC / CV, and semi-syllables such as VCV instead of phonemes in the standard pattern.

[0027]

As described above, according to the present invention, the feature parameters obtained by analyzing the recognized word voices uttered by a small number of speakers, one to a few, are created in advance by many speakers. The similarity is calculated by calculating the similarity with n types of standard patterns, and the N-dimensional similarity vector of the n-dimensional similarity vector and the M-dimensional regression coefficient vector of the n-dimensional dictionary are used as the feature parameters for speech recognition. Even if is registered and created, the voice of an unspecified speaker can be accurately recognized.

In addition, this allows the dictionary to have a storage capacity less than half that of the conventional dictionary, the amount of calculation to be extremely small, and the recognition rate to be hardly reduced. As described above, the present invention greatly contributes to the improvement of the practical technology of the voice recognition device for the unspecified speaker.

[Brief description of drawings]

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

FIG. 2 is a conceptual diagram illustrating a matching method in the embodiment.

FIG. 3 is a conceptual diagram illustrating a dictionary and input in the same embodiment.

FIG. 4 is a characteristic diagram illustrating a regression coefficient in the example.

FIG. 5 is a conceptual diagram illustrating a time series of similarity vectors in a conventional voice recognition method.

FIG. 6 is a conceptual diagram illustrating a dictionary and input in a conventional voice recognition method.

FIG. 7 is a block diagram illustrating a conventional voice recognition method.

[Explanation of symbols]

 9 Acoustic Analysis Section 10 Feature Parameter Extraction Section 11 Similarity Calculation Section 12 Standard Pattern Storage Section 13 Regression Coefficient Calculation Section 14 Parameter Series Creation Section 15 Parameter Selection Section 16 Recognition Section 17 Dictionary Storage Section 18 Index Section 19 Similarity Storage Section 20 Regression Coefficient storage

Claims (4)

[Claims] 1. The speech to be recognized is produced by a small number of speakers, one to several, and m (m is an integer) acoustic feature parameter obtained at each analysis time (frame). Matching is performed with m feature parameters of each of the n types (n is an integer) of standard patterns created by the speaker, the n similarity is calculated for each frame, and the n-dimensional similarity vector is created. M obtained by selecting N (N is an integer) smaller than n from the time series pattern, registering it as a word dictionary, and similarly analyzing the input speech to be recognized
Characteristic parameters and each of the n standard patterns
The matching target speech is obtained by matching the m number of feature parameters possessed by, and obtaining the time series of the n-dimensional similarity vector and matching it with the time series of the N-dimensional similarity vector registered in the dictionary. A voice recognition method characterized by recognizing a registered speaker and other input voices. 2. A temporal change amount of similarity for each time series of n types of similarity is calculated for each frame, and M (n) smaller than n are selected from n-dimensional vectors of the temporal change amount of similarity. 2. The speech recognition method according to claim 1, wherein a word dictionary in which M is an integer) is selected and registered as a dictionary is used in combination with an N-dimensional vector of similarity. 3. The speech recognition method according to claim 1, wherein when reducing from N-dimensional similarity vectors to N-dimensional, N pieces are selected from the one having the larger value. 4. The speech according to claim 1, wherein M vectors are selected from those having a larger absolute value when reducing the vector of n-dimensional similarity temporal change amounts to M dimensions. Recognition method.