JP2001290494A - Method and device for generating registered word dictionary, and method and device for speech recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To drastically reduce a dictionary domain while maintaining a high recognition rate and to easily realize hardware by reducing the number of computations. SOLUTION: Parmeter time series computed from inputted speech are coded employing a degree of similarity vector code book 106 that is beforehand generated by using learning speech and a regression coefficient code book 108. The coded series are registered in a coded registration word dictionary 114 as word patterns. In order to recognize speech, the word pattern stored in the dictionary 114 are restored into parameter time series by using the books 106 and 108 and a recognition result is obtained by comparing the restored parameter time series and the parameter series computed from the speech to be recognized.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and an apparatus for creating a registered word dictionary for registering and recognizing a recognized vocabulary by voice, and a speech recognition method and an apparatus using the same.

[0002]

2. Description of the Related Art Conventionally, when a word dictionary for determining a vocabulary to be recognized is created using speech, a registrant utters the same word one or more times to use a dictionary word pattern from its parameters. They were created, and those created for each of a plurality of words were collected and used as a registered word dictionary. When using this to recognize speech, a method has been adopted in which input speech is recognized by comparing it with dictionary word patterns in this dictionary. Such a speech registration speech recognition method is described in, for example, "Study of a Registered Model English Speech Method" (Acoustic Society of Japan, 1-Q-11,
1995/3).

FIGS. 6 and 7 are block diagrams of this method. FIG. 6 creates a dictionary word pattern from the input voice,
FIG. 7 is a configuration diagram of a method of registering in a registered word dictionary, and FIG. 7 is a configuration diagram of a method of recognizing an input voice using the registered word dictionary. 6, 1 is an acoustic analysis unit, 2 is a similarity calculation unit, 3 is a standard pattern storage unit, 4 is a regression coefficient calculation unit, 9 is a word pattern storage unit, 10 is a dictionary word pattern creation unit, and 14 is a registered word. It is a dictionary. Also, in the configuration diagram of the recognition method in FIG. 7, steps 1 to 4 are the same as those in FIG. 6, 15 is a word pattern creation unit, 14 is a registered word dictionary, and 17 is a word recognition unit.

[0006] Here, the standard pattern storage unit 3 common to FIG. 6 and FIG. 7 optimizes the characteristics of phonemes for each of n phonemes with respect to training data uttered in advance by many speakers. Find the temporal position (feature frame) to represent,
A phoneme standard pattern created using a time pattern of feature parameters centered on the feature frame is stored. As the time pattern, LPC cepstrum coefficients (C0 to C
8) is calculated, a parameter series in which these are arranged one-dimensionally is obtained, and an average value vector of each element of the parameter series and a covariance matrix between the elements are obtained and stored as a standard pattern.

The configurations of the acoustic analysis unit 1, similarity calculation unit 2, standard pattern storage unit 3, and regression coefficient calculation unit 4 shown in FIGS. This is a configuration for obtaining a series of dimensional regression coefficient vectors, and these series are obtained as follows.

That is, the acoustic analysis unit 1 performs an acoustic analysis on the input speech for each analysis frame, and calculates a series of LPC cepstrum coefficients. The similarity calculation unit 2
The sequence of LPC cepstrum coefficients obtained by the acoustic analysis unit 1 is compared with each of the n patterns in the standard pattern storage unit 3 to obtain n similarities, and the n-dimensional similarity is defined as one vector. A series of vectors is calculated. The regression coefficient calculator 4 calculates the similarity vector from the time series
A regression coefficient is calculated for each vector component for each frame, and a series of n-dimensional regression coefficient vectors is obtained.

In the case of creating a dictionary word pattern for a registered word dictionary, in FIG. 6, the above-described similarity vector series and regression coefficient vector series are respectively obtained from speech data of the same word uttered twice. These are stored in the word pattern storage unit 9 as a set. The dictionary word pattern creation unit 10 performs time matching by DP matching on a set of a similarity vector sequence and a regression coefficient vector sequence for two utterances stored in the word pattern storage unit 9, and performs average matching according to the result. A similarity vector series and an average regression coefficient vector series are obtained and stored in the registered word dictionary 14 as a dictionary word pattern of the corresponding word.

When the input speech is recognized, the similarity is calculated from the input speech according to the configurations of the acoustic analysis unit 1, the similarity calculation unit 2, the standard pattern storage unit 3, and the regression coefficient calculation unit 4 in FIG. A set of a series of vectors and a series of regression coefficient vectors is obtained.

The two vector sequences are combined by a word pattern creating section 15 to form a word pattern of the input speech. In the word recognition unit 15, the word pattern created by the word pattern creation unit 15 is compared with each dictionary word pattern stored in the registered word dictionary 14, and the word with the highest score is determined as a recognition result. Here, a distance using time matching by DP matching is used for comparing the word pattern of the input voice with the dictionary word pattern in the registered word dictionary.

[0010]

As described above, in the conventional method described above, in order to store each dictionary word pattern in the registered word dictionary, an n-dimensional similarity vector and an n-dimensional similarity vector are stored for each frame of the word. Since it is necessary to store the regression coefficient vector, there is a problem that a very large memory area is required as an area for the registered word dictionary.
It is required to reduce this.

[0011] Further, in comparison between the input speech and the dictionary word pattern of the registered word dictionary, there is a problem that 2 × n multiplications are required per frame in one word, and the calculation takes time. It is required to reduce this.

The present invention solves the above-mentioned problem. By expressing a dictionary word pattern in a registered word dictionary using a compressed codebook, the dictionary area can be significantly increased while maintaining a high recognition rate. An object of the present invention is to reduce the number of operations and the number of operations, and to facilitate hardware implementation.

[0013]

In order to achieve this object, in the creation of a registered word dictionary according to the present invention, a time series of parameters calculated from an uttered voice is prepared by using a code prepared in advance using a learning voice. Encoding is performed using a book, and this code sequence is registered as a word pattern in a registered word dictionary. When recognizing speech, the word pattern stored in the registered word dictionary is restored to a time series of parameters using a codebook, and this is compared with the parameter series calculated from the speech to be recognized. To obtain the recognition result.

According to the present invention, the dictionary word pattern of the registered word dictionary, which has been conventionally expressed as an n-dimensional similarity vector series or an n-dimensional regression coefficient vector series, is:
Each of the similarity vector series and the regression coefficient vector series is represented as a simple one-dimensional code series, and the memory area for the registered word dictionary is reduced without lowering the recognition accuracy, thereby facilitating hardware implementation.

Further, in the present invention, a memory area for a codebook for encoding / decoding the similarity vector and the regression coefficient vector is newly required. However, the n-dimensional similarity constituting the codebook is required. By selecting only N or M valid vector components smaller than n for each of the vector and the n-dimensional regression coefficient vector, the input speech and the dictionary word pattern in the registered word dictionary are compared. The number of operations can be reduced, and the size of the codebook itself can be compressed.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS According to the first aspect of the present invention, a word pattern obtained from an utterance voice for vocabulary registration is converted into a word pattern.
It is characterized by encoding using a codebook and registering this code sequence as a dictionary word pattern in a registered word dictionary, and has the effect of greatly reducing the area for storing the registered word dictionary.

According to a second aspect of the present invention, a registrant utters dictionary registration words and obtains m (m is an integer) obtained for each frame.
Are matched with m feature parameters of each of n types (n is an integer) of standard patterns created in advance from learning data, and n similarities are obtained for each frame to obtain n-dimensional n-dimensional similarities. A similarity vector sequence is encoded for each frame using a codebook previously obtained from the similarity vector of the learning data,
This code sequence is registered as a dictionary word pattern in a registered word dictionary, and has the effect of significantly reducing the memory area for a registered word dictionary using a similarity vector sequence as a dictionary word pattern. .

According to a third aspect of the present invention, m words (m is an integer) obtained for each frame by a registrant uttering dictionary registration words.
Are matched with m feature parameters of each of n types (n is an integer) of standard patterns created in advance from learning data, and n similarities are obtained for each frame to obtain n-dimensional n-dimensional similarities. As a similarity vector series, a time change amount of the similarity is obtained for each frame for each of the n types of similarity time series, and a series of n-dimensional vectors of the time change amount of the similarity is learned in advance. Encoding is performed for each frame using a codebook obtained from an n-dimensional vector of a temporal change amount of data similarity,
This code sequence is registered as a dictionary word pattern of a registered word dictionary, and a memory area for a registered word dictionary in which a vector sequence of a time variation of similarity is used as a word pattern can be significantly reduced. It has the action of:

According to a fourth aspect of the present invention, in the method for creating a registered word dictionary according to the second or third aspect, a code sequence obtained by encoding an n-dimensional similarity vector sequence using a codebook is used. A method for creating a registered word dictionary, characterized in that an n-dimensional vector sequence of a degree of time change is encoded using a codebook and a code sequence obtained is combined and registered as a dictionary word pattern. This has the effect of significantly reducing the memory area for a registered word dictionary that is used as a word pattern by using both a sequence and a vector sequence of the time variation of similarity.

According to a fifth aspect of the present invention, in the method for creating a registered word dictionary according to the second aspect, when a codebook is created from the similarity vector of the learning data, each of the n-dimensional code vectors has The codebook is characterized by using a code vector created by selecting only small N components and reducing the size of the codebook necessary for encoding a word pattern by a similarity vector sequence. Has the effect of being able to.

According to a sixth aspect of the present invention, in the method for creating a registered word dictionary according to the third aspect, when creating a code book of a vector of the amount of time change of the similarity, each of the n-dimensional code vectors includes: It is characterized in that a codebook is created by using a code vector created by selecting only M components smaller than n, and is necessary for encoding a word pattern by a vector series of a time variation of similarity. This has the effect that the size of the codebook can be reduced.

According to a seventh aspect of the present invention, in the method for creating a registered word dictionary according to the fourth aspect, when creating a codebook of similarity vectors, only N components smaller than n are selected and created. When creating a codebook using a code vector and a codebook of a vector of the amount of time change of similarity, a codebook using a code vector created by selecting only M components smaller than n is used. And has the effect of reducing the size of a codebook required when encoding a word pattern that uses both a similarity vector sequence and a vector sequence of a temporal change in similarity.

According to an eighth aspect of the present invention, each dictionary word pattern encoded from a registered word dictionary created in advance is decoded using a codebook previously obtained from a similarity vector of learning data, and this decoding is performed. It is characterized by recognizing the uttered input voice by using the coded dictionary word pattern, and has the effect that the input voice can be recognized by using the coded and compressed word dictionary. .

According to a ninth aspect of the present invention, the uttered input voice is composed of m (m is an integer) acoustic feature parameters obtained for each frame and n types (n is an integer) prepared in advance from learning data. ) Are matched with m feature parameters of each of the standard patterns, and n similarities are obtained for each frame to obtain an n-dimensional similarity vector sequence.
This is restored as a dictionary word pattern from a registered word dictionary created in advance for each frame using a codebook previously obtained from a similarity vector of learning data, and the input speech is recognized using the dictionary word pattern. And has the effect that input speech can be recognized using an encoded and compressed word dictionary.

According to a tenth aspect of the present invention, the uttered input voice is composed of m (m is an integer) acoustic feature parameters obtained for each frame and n types (n is an integer) created in advance from learning data. M) of each of the standard patterns
Matching is performed with n feature parameters, n similarities are obtained for each frame, and an n-dimensional similarity vector series is obtained. For each of the n types of similarity time series,
The time change amount of the similarity is obtained for each frame, and a series of n-dimensional vectors of the time change amount of the similarity is calculated using a codebook previously obtained from the n-dimensional vector of the time change amount of the similarity of the learning data. A dictionary word pattern is restored from a registered word dictionary created in advance for each frame, and the input speech is recognized using the dictionary word pattern. This has the effect that the input speech can be recognized using the converted sequence.

According to an eleventh aspect of the present invention, the uttered input voice is composed of m (m is an integer) acoustic feature parameters obtained for each frame and n types (n is an integer) prepared in advance from learning data. M) of each of the standard patterns
Matching is performed with n feature parameters, n similarities are obtained for each frame, and an n-dimensional similarity vector series is obtained. For each of the n types of similarity time series,
The temporal change amount of the similarity is obtained for each frame, and the n-dimensional similarity vector sequence and n of the temporal change amount of the similarity are calculated.
A word pattern of an input voice obtained by combining a series of dimensional vectors into one is restored as a dictionary word pattern from a registered word dictionary created in advance for each frame using a codebook obtained in advance from the similarity of learning data, Using the dictionary word pattern to recognize the input voice, the input voice can be recognized by using a coded sequence of a similarity vector and a vector of a temporal change amount of the similarity. Has the effect of being able to.

According to a twelfth aspect of the present invention, in the speech recognition method according to any one of the eighth to eleventh aspects, the registered word dictionary is created by the method according to any one of the first to seventh aspects. It is characterized by using a dictionary, and has an effect that input speech can be recognized using an encoded and compressed word dictionary.

According to a thirteenth aspect of the present invention, the registrant utters the dictionary registration word and obtains m (m is an integer) acoustic feature parameters obtained for each frame,
N types (n is an integer) created in advance from learning data
Standard pattern storage means storing m feature parameters of each of the standard patterns, m feature parameters from the standard pattern storage means, and m (m is an integer) acoustic features from the acoustic analysis means A similarity calculating means for performing parameter matching, obtaining n similarities for each frame, and obtaining an n-dimensional similarity vector series, and using a codebook previously obtained from a similarity vector of learning data for each frame. And a coded registered word dictionary storage means for registering the coded code sequence as a dictionary word pattern in a registered word dictionary. This has the effect of significantly reducing the memory area for a registered word dictionary using a vector sequence as a dictionary word pattern.

According to a fourteenth aspect of the present invention, the registrant utters the dictionary registration word, and obtains m (m is an integer) acoustic feature parameters obtained for each frame,
N types (n is an integer) created in advance from learning data
Standard pattern storage means storing m feature parameters of each of the standard patterns, m feature parameters from the standard pattern storage means, and m (m is an integer) acoustic features from the acoustic analysis means Parameter matching, n similarities are calculated for each frame, and a similarity calculating means for obtaining an n-dimensional similarity vector series; and a time series of the similarity for each of the n types of similarity time series. Regression coefficient calculating means for obtaining the amount of change for each frame and obtaining a series of n-dimensional vectors of the amount of time change of the similarity, and a code which previously obtained this from an n-dimensional vector of the amount of time change of the similarity of the learning data Using a book,
A regression coefficient vector encoding unit that encodes each frame, and a coded registered word dictionary storage unit that registers the coded code sequence in a registered word dictionary as a dictionary word pattern. This has the effect that the memory area for a registered word dictionary in which a vector sequence of similarity changes over time is used as a word pattern can be significantly reduced.

According to a fifteenth aspect of the present invention, in the registered word dictionary creating device according to the thirteenth or fourteenth aspect, the coded registered word dictionary storage means stores the n-dimensional similarity vector series using a codebook. The code sequence obtained by encoding and the code sequence obtained by encoding an n-dimensional vector sequence of the time change amount of the similarity using a code book are combined and registered as a dictionary word pattern. This has an effect that a memory area for a registered word dictionary that is used as a word pattern by using the similarity vector sequence and the vector sequence of the time variation of the similarity together can be significantly reduced.

According to a sixteenth aspect of the present invention, the uttered input voice is converted into m (m is an integer) acoustic feature parameters obtained for each frame, and n is generated from learning data in advance. Standard pattern storage means for registering m feature parameters of each type (n is an integer) of standard patterns, and matching of the m (m is an integer) acoustic feature parameters with the m feature parameters A similarity calculating means for obtaining n similarities for each frame and obtaining an n-dimensional similarity vector series;
Word pattern restoring means for restoring the n-dimensional similarity vector series as a dictionary word pattern from a registered word dictionary created in advance for each frame using a codebook previously obtained from the similarity vector of the learning data; And a word recognizing means for recognizing the input voice using the dictionary word pattern, wherein the input voice can be recognized using the encoded and compressed word dictionary. Having.

According to a seventeenth aspect of the present invention, the uttered input voice is converted from an acoustic analysis unit for obtaining m (m is an integer) acoustic feature parameters obtained for each frame, and n generated from learning data in advance. Standard pattern storage means for registering m feature parameters of each type (n is an integer) of standard patterns, and matching of the m (m is an integer) acoustic feature parameters with the m feature parameters A similarity calculating means for obtaining n similarities for each frame and obtaining an n-dimensional similarity vector series;
Regression coefficient calculating means for obtaining, for each frame, a time change amount of the similarity for each of the n types of similarity time series, and obtaining an n-dimensional vector series of the time change amount of the similarity; Word pattern restoration that restores a dimensional similarity vector series as a dictionary word pattern from a registered word dictionary created in advance for each frame, using a codebook previously obtained from an n-dimensional vector of the time variation of the similarity of the learning data Means, and word recognition means for recognizing the input voice using the restored dictionary word pattern, wherein the input voice is recognized using an encoded and compressed word dictionary. Has the effect of being able to

The invention according to claim 18 is a sound analysis means for obtaining m (m is an integer) acoustic feature parameters obtained for each frame from the uttered input voice, and n which is previously created from learning data. Standard pattern storage means for registering m feature parameters of each type (n is an integer) of standard patterns, and matching of the m (m is an integer) acoustic feature parameters with the m feature parameters A similarity calculating means for obtaining n similarities for each frame and obtaining an n-dimensional similarity vector series;
For each of the n types of time series of similarities, a regression coefficient calculating means for obtaining a time change amount of the similarity for each frame, and obtaining an n-dimensional vector series of the time change amount of the similarity; Word pattern creating means which combines an n-dimensional similarity vector series from the degree calculating means and an n-dimensional vector series of similarity time change from the regression coefficient calculating means into a single word pattern of the input voice And a word for restoring a word pattern of the input voice as a dictionary word pattern from a registered word dictionary created in advance for each frame using a codebook previously obtained from an n-dimensional vector of a time change amount of similarity of learning data. Pattern restoring means, comprising using the restored dictionary word pattern, word recognition means for recognizing the input voice, An effect that can recognize an input speech with Goka and compressed word dictionary.

According to a nineteenth aspect of the present invention, in the speech recognition apparatus according to any one of the sixteenth to eighteenth aspects, the registered word dictionary is provided by the registered word dictionary creating apparatus according to any one of the thirteenth to fifteenth aspects. It is characterized by using the created registration dictionary, and has the effect that input speech can be recognized using an encoded and compressed word dictionary.

Hereinafter, an embodiment of the present invention will be described with reference to FIG.
This will be described with reference to FIG.

(Embodiment 1) FIG. 1 is a block diagram of a registered word dictionary creating apparatus according to Embodiment 1 of the present invention, which will be described.

In FIG. 1, reference numeral 1 denotes an acoustic analysis unit for obtaining an LPC cepstrum from an input voice, 2 denotes a similarity calculation unit for obtaining a similarity vector from a comparison between the input voice and a standard pattern, and 3 denotes a standard pattern storage used for similarity calculation. , 4 is a regression coefficient calculation unit for calculating a regression coefficient vector from a time series of similarity vectors, S1 is a switch for distributing the calculated similarity vector between codebook creation and word registration, and S2 is a calculated regression. This is a switch for distributing coefficient vectors between when creating a codebook and when registering words.

A similarity calculation unit 105 performs clustering on a large number of similarity vectors created from the learning speech, obtains one or more representative (average) vectors, and uses these as code vectors. A code vector creation unit 106 is a similarity vector codebook that collectively uses the similarity code vectors obtained by the similarity code vector creation unit 105 as a codebook, and 107 is a regression coefficient calculation unit 4 created from learning speech. Perform clustering on many regression coefficient vectors, one or more representatives
A regression coefficient code vector creation unit 108 that obtains (for example, average) vectors and uses these as code vectors, and a regression coefficient vector code book 108 that collects the regression coefficient code vectors obtained by the regression coefficient code vector creation unit 107 and creates a code book , 9 is a word pattern storage unit that stores the time series of the similarity vector and the time series of the regression coefficient vector obtained by the similarity calculation unit 2 and the regression coefficient calculation unit 4 as a set. 110 is stored in the word pattern storage unit 9. An average word pattern creation unit that creates an average pattern from the two word patterns obtained.

A similarity vector encoding unit 111 encodes the similarity vector sequence in the average word pattern using the similarity vector codebook stored in the similarity vector codebook 106 and converts it into a code sequence. A regression coefficient vector encoding unit that encodes a regression coefficient vector series using the regression coefficient vector codebook stored in the regression coefficient vector codebook 108 and converts the regression coefficient vector series into a code series. A coded dictionary word pattern creating unit that collects code sequences corresponding to the regression coefficient vector series into a dictionary word pattern, and 114 is a coded registered word dictionary that stores the dictionary word pattern created by the coded dictionary word pattern creating unit 113 It is.

The operation of the above configuration will be described below. First, an outline of the operation will be described below. The speech recognition device compares the word pattern obtained from the input speech with the word pattern stored in the word dictionary, and selects the most similar pattern. Here, since the word pattern in the dictionary is coded, it is necessary to restore the dictionary word pattern for comparison with the input voice pattern. In the present embodiment, this restoring operation is performed sequentially without performing collective operation for each word pattern. That is, each time a vector of one frame in the voice pattern is calculated from the input voice, the vector of the word pattern in the dictionary required for the distance comparison with the vector is restored each time. By sequentially repeating such processing, recognition calculation can be performed while the memory area for developing the dictionary word pattern is suppressed.

Next, a detailed description will be given. First, when creating a codebook, the similarity vector changeover switch S1 of FIG.
2 is connected to a side.

For speech for creating a code book, utterance data of many speakers prepared in advance is used. These sounds are analyzed by the acoustic analysis unit 1 (referred to as frames).
In this embodiment, linear prediction analysis is performed every 10 msec) to obtain LPC cepstrum coefficients (C0 to C11). The LPC cepstrum coefficient is used by the similarity calculation unit 2 to obtain a similarity vector using the phoneme standard pattern stored in the standard pattern storage unit 3.

The standard pattern storage 3 stores Japanese 24 characters.
The parameters indicating the statistical distribution of the LPC cepstrum of each type of phoneme are stored as standard patterns. These parameters can be accurately and visually inspected with respect to voice data of a large number of speakers separately prepared for creating phoneme standard patterns. Using labeled ones, it is created using LPC cepstrum coefficients (C0 to C11) in the range of 10 frames before and after (100 ms) centering on the part that best represents the characteristics of each phoneme.

The similarity calculator 2 calculates the statistical distance between the LPC cepstrum coefficient of the input speech obtained from the acoustic analyzer 1 and the phoneme standard pattern stored in the standard pattern storage 3 by using the 24 phoneme standard patterns. For each of them, a 24-dimensional vector using these as coefficients is created.

The regression coefficient calculation unit 4 calculates a regression coefficient in the time axis direction for each dimension as a parameter representing a time change of each dimension of the 24-dimensional similarity vector obtained by the similarity calculation unit 2. Is a regression coefficient vector.

A number of similarity vectors and a number of regression coefficient vectors obtained as described above are combined with a similarity code vector creation unit 105 and a regression coefficient code vector creation unit 10.
7

In the similarity code vector creation unit 105,
Vector quantization (VQ) by K-means clustering is performed on a large number of similarity vectors collected,
One or more representative vectors representing the spatial distribution of the similarity vector are calculated.

The representative vector is directly paired with a code for identifying each as a code vector as shown in FIG.
6 can be stored, but the top six are selected in descending order of the absolute value of the component of the vector, and this is used as a compressed code vector. When this is stored in the similarity vector code book 106, not only a compressed code vector and a code for identifying the compressed code vector but also an index indicating a selected component are paired, and a structure shown in FIG. Is stored.

Similarly, the regression coefficient code vector creation unit 10
Also in step 7, vector quantization is performed on a large number of regression coefficient vectors collected, one or more representative vectors are calculated, and for each vector, only the top six with a large absolute value of the vector component are determined. The code vector is selected and compressed. Regression coefficient vector codebook 10
At 8, the compressed code vector of the obtained regression coefficient vector is stored together with the identification code and the index indicating the selected component.

Next, a method of registering a word pattern for registration in the registered word dictionary will be described. When registering a word pattern in the dictionary, both the similarity vector changeover switch S1 and the regression coefficient vector changeover switch S2 are connected to the side b.

At the time of word registration, the registrant utters the same word one or more times. For each utterance, a sound analysis unit 1, a similarity calculation unit 2,
Through the regression coefficient calculation unit 4, a series of the similarity vector and the regression coefficient vector is calculated for each frame. These two vector sequences are stored in the word pattern storage unit 9 as word patterns for each word utterance.

The average word pattern creating section 110 obtains an average word pattern from a plurality of word patterns obtained from one or more utterances stored in the word pattern storage section 9. However, when the word for registration is uttered only once, the pattern is treated as it is as an average word pattern.

The similarity vector encoding section 111 performs similarity vector code book 106 on each vector in the similarity vector sequence forming the average word pattern.
Is selected and the one that is most similar is selected and converted to an identification code corresponding to that vector. As a result, the sequence of similarity vectors is converted into a one-dimensional code sequence.

Similarly, the regression coefficient vector encoding unit 112 also determines the most significant of the code vectors stored in the regression coefficient vector code book 108 for each vector in the regression coefficient vector sequence forming the average word pattern. One similar one is selected and converted to a corresponding identification code. Thus, the regression coefficient vector sequence is converted into a one-dimensional code sequence.

The code sequence of the similarity vector and the code sequence of the regression coefficient vector obtained as described above are combined by the coded dictionary word pattern creating unit 113, and
The parameters are grouped as parameters representing one word pattern and sent to the coded registered word dictionary 114. In this way, the coded registered word dictionary 114 stores word patterns represented by code sequences as many as the number of recognized vocabularies.

In the above-described embodiment, the method of using vector quantization by clustering as a method of creating a codebook of a similarity vector and a regression coefficient vector is described. A method using a method optimized by the method is also effective.

In this embodiment, the code vector is selected in the order of the absolute value of each component, and only six code vectors are selected. However, a component having a statistically significant meaning is left as a selection criterion. The method is effective and the number of choices is six
Although fixed individually, it is also possible to adopt a method that can be arbitrarily changed for each vector.

In the embodiment, the method of coding both the similarity vector and the regression coefficient vector by the code book has been described. However, a method using only one of them is also effective.

Further, in the embodiment, the similarity vector code book 106 and the regression coefficient vector code book 108
Have been described above, but it is also effective to create them in common to form a single codebook and encode the similarity vector and the regression coefficient vector.

(Embodiment 2) FIG. 3 shows a configuration diagram of a speech recognition apparatus according to Embodiment 2 of the present invention, which will be described below.

In FIG. 3, an acoustic analysis unit 1, a similarity calculation unit 2, a standard pattern storage unit 3, and a regression coefficient calculation unit 4
Are the same as those of the first embodiment shown in FIG. 1, and a detailed description thereof will be omitted. Reference numeral 15 denotes a word pattern creating unit that combines the similarity vector series and the regression coefficient vector series obtained for the input voice to form a word pattern, 11
4 is a coded registered word dictionary storing each word pattern created in the first embodiment, 106 is a similarity vector codebook created in the first embodiment, and 108 is a regression coefficient vector code created in the first embodiment Book, 116 is a word pattern restoring unit that restores a code sequence registered as a dictionary word pattern in the coded registered word dictionary into a similarity vector sequence and a regression coefficient vector sequence, and 17 is an input speech and restored dictionary word pattern. This is a word recognition unit for comparing and obtaining a recognition result.

Next, the operation of the speech recognition apparatus configured as described above will be described. First, in the same manner as in the first embodiment, the input voice is subjected to the similarity analysis using the acoustic analysis unit 1, the similarity calculation unit 2, the standard pattern storage unit 3, and the regression coefficient calculation unit 4 at every analysis period of 10 ms. A series of vectors and a regression coefficient vector are calculated.

In the word pattern creating section 15, the two vectors, the similarity vector and the regression coefficient vector, are put together into one, forming one frame in the word pattern of the input voice.

The word recognizing unit 17 performs time matching matching by DP matching with each frame of each word pattern stored in the coded registered word dictionary 114 every time one frame in the word pattern of the input voice is obtained. Are sequentially advanced to calculate a score. here,
Since the word patterns stored in the coded registered word dictionary 114 are coded, the similarity vector code book 106 and the regression coefficient vector code book 108
, The word pattern restoration unit 116 successively restores the similarity vector and the regression coefficient vector.

That is, in the DP matching of the word recognizing unit 17, when the i-th frame of the word w stored in the coded registered word dictionary 114 is requested, the code of this frame is stored in the similarity vector code book. 10
6 and the regression coefficient vector codebook 108, the word pattern is restored to a code vector corresponding to the code by the word pattern restoration unit 116, and is passed to the word recognition unit 17.

By repeating this process, it is possible to perform a recognition calculation that does not require a large area for storing a registered word pattern. Further, among the scores obtained by performing the above processing for all the words in the dictionary, the word giving the highest score is the recognition result.

In this embodiment, the similarity vector code book 106 and the regression coefficient vector code book 1 for restoring a word pattern from a registered word dictionary are used.
In 08, the same codebook as that used at the time of creating the registered word dictionary was used, but this is not necessarily the same, and a codebook optimized for word pattern restoration may be used. .

Although the DP matching is used for the matching for comparing the word pattern of the input voice with the word pattern restored from the registered word dictionary, other methods using time matching are also effective.

(Examples) Two recognition experiments were performed using the first and second embodiments described above. One is an experiment for examining the relationship between the size of the similarity vector codebook 106 and the regression coefficient vector codebook 108 and the word recognition rate, and the other is as a code vector compression method. An experiment was performed to determine the number of choices. The data used in the experiment was data for deciding the codebook size and the number of selected codevector components, using data in which 543 words in Japanese were uttered by five men and women for a codebook creation. , Using data of 100 Japanese place names uttered three times by a total of 50 people, 25 men and women, of which two utterances are used for word registration, and the remaining one utterance is used for recognition evaluation. Was.

In the present recognition experiment, Japanese standard 2 was used as a standard pattern for obtaining a similarity vector series from input speech.
A pattern of four phonemes is used, and the number of dimensions of the similarity vector and the regression coefficient vector is set to 24. When creating a codebook of a similarity vector and a regression coefficient vector, a code vector is created using VQ (vector quantization), and the distance between the vectors is defined by (Equation 1). The correlation cosine was used.

[0071]

(Equation 1)

Further, DP matching is used for obtaining the average pattern from a plurality of word patterns of the input voice for registration and for matching the input voice with the dictionary word pattern restored from the registered word dictionary.

FIG. 4 shows a similarity vector codebook 1
11 shows the result of examining the change in the recognition rate with respect to the size of the codebook 06 and the regression coefficient vector codebook 108. Here, it is assumed that the size of the similarity vector codebook 106 and the size of the regression coefficient vector codebook 108 are the same, and the horizontal axis represents the number of code vectors included in each codebook, and the vertical axis represents the recognition rate at that time. Shown in In this experiment, the selection of the components of the code vector in each code book was not performed.

According to the experimental results, when the size of each codebook is 40, the recognition rate is 97.4%, which is 1.0% lower than that of the conventional system of 98.4%. Can be suppressed. At this time, the size of the registered word dictionary is about 5.5 per registered word on average by the conventional method.
Where about a kilobyte is required, it can be compressed to about 120 bytes on average, resulting in 1/48 compression.

FIG. 5 shows a similarity vector codebook 1
6 shows a result of examining a change in recognition rate when a component of each code vector included in the codebook 06 and the regression coefficient vector code book 108 is selected. Here, the number of selected code vector components is the same in the similarity vector and the regression coefficient vector, and the selected number is shown on the horizontal axis, and the recognition rate at that time is shown on the vertical axis. In this experiment, the size of each codebook was the same, and the number of code vectors was 40.
According to this result, even when only six components are selected from the components of each 24-dimensional code vector, the recognition rate is 9
This is 7.3%, which is a reduction of 0.1% in recognition rate, compared to 97.4% when no code vector component is selected.

At this time, the size of the code book is
Each code vector is compressed from 48 bytes to 18 bytes.

[0077]

As described above, the present invention obtains a number of similarity vectors and a number of code vectors representing the spatial distribution of regression coefficient vectors obtained from learning speech data, thereby obtaining a continuous code. A one-dimensional parameter representing a parameter that needs to be stored as a vector value as a vector having a numerical value with a small error can be replaced with a small error, whereby the registered word dictionary can be reduced without substantially reducing the word recognition rate. Can be greatly reduced to 1/48 of that of the conventional method.

At this time, the codebook, which is separately required, is reduced to half or less the size of the normally created codebook by selecting the components of the code vector, and is used for distance calculation. The number of times of multiplication itself can be reduced. As described above, the present invention greatly contributes to the improvement of the practical application technology of the speech registration type speech recognition apparatus.

[Brief description of the drawings]

FIG. 1 is a block diagram of a registered word dictionary creation device according to a first embodiment of the present invention.

FIG. 2 is a structural diagram of a normal codebook and a compressed codebook according to the embodiment;

FIG. 3 is a block diagram of a speech recognition apparatus according to a second embodiment of the present invention.

FIG. 4 is a diagram showing an experimental result obtained by examining a change in a recognition rate with respect to a codebook size in the embodiment.

FIG. 5 is a diagram showing an experimental result of examining a change in recognition rate when a code vector component in a code book is selected in the embodiment.

FIG. 6 is a block diagram of a registered word dictionary creating apparatus according to a conventional utterance registration method.

FIG. 7 is a block diagram showing the configuration of a speech recognition apparatus using a conventional utterance registration method.

[Explanation of symbols]

 Reference Signs List 1 acoustic analysis unit 2 similarity calculation unit 3 standard pattern storage unit 4 regression coefficient calculation unit 9 word pattern storage unit 10 dictionary word pattern creation unit 14 registered word dictionary 15 word pattern creation unit 17 word recognition unit 18 code storage unit 19 code vector Storage unit 20 Compressed code vector storage unit 105 Similarity code vector creation unit 106 Similarity vector codebook creation unit 107 Regression coefficient code vector creation unit 108 Regression coefficient vector codebook 110 Average word pattern creation unit 111 Similarity vector encoding unit 112 Regression coefficient vector encoding unit 113 Coded dictionary word pattern creation unit 114 Coded registered word dictionary 116 Word pattern restoration unit

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Taisuke Watanabe 3-10-1 Higashi-Mita, Tama-ku, Kawasaki-shi, Kanagawa F-term in Matsushita Giken Co., Ltd. 5D015 AA02 BB01 FF05 GG04 5D045 CB01 DA11 9A001 EE04 HH17

Claims (19)

[Claims] 1. A method for creating a registered word dictionary, comprising encoding a word pattern obtained from a vocalized speech for vocabulary registration using a codebook, and registering the code sequence as a dictionary word pattern in a registered word dictionary. 2. A dictionary registration word is uttered by a registrant, and m (m is an integer) acoustic feature parameters obtained for each frame, and n (n is an integer) standard patterns created in advance from learning data By performing matching with m feature parameters of each, n similarities are obtained for each frame to obtain an n-dimensional similarity vector sequence, and this is determined using a codebook previously obtained from the similarity vector of the learning data. A method for creating a registered word dictionary, characterized by encoding each frame and registering this code sequence in a registered word dictionary as a dictionary word pattern. 3. A dictionary registration word is uttered by a registrant, and m (m is an integer) acoustic feature parameters obtained for each frame, and n (n is an integer) standard patterns created in advance from learning data Matching is performed with the m feature parameters of each of them, n similarities are obtained for each frame, and an n-dimensional similarity vector sequence is obtained. Is calculated for each frame, and a series of n-dimensional vectors of the time variation of the similarity is calculated for each frame using a codebook previously obtained from the n-dimensional vector of the time variation of the similarity of the learning data. And registering the code series as a dictionary word pattern of a registered word dictionary. 4. A dictionary word pattern is obtained by encoding an n-dimensional similarity vector sequence using a codebook and encoding an n-dimensional vector sequence of similarity change over time using a codebook. 4. The registered word dictionary creation method according to claim 2, wherein a code sequence obtained by encoding is registered as a dictionary word pattern. 5. When a codebook is created from a similarity vector of learning data, a codebook is created using a codevector created by selecting only N components smaller than n for each n-dimensional codevector. 3. The method for creating a registered word dictionary according to claim 2, wherein: 6. When creating a codebook of a vector of the amount of time change of the similarity, a codebook is created using a codevector created by selecting only M components smaller than n for each n-dimensional codevector. 4. The method for creating a registered word dictionary according to claim 3, wherein 7. When creating a codebook of a similarity vector, a codebook using a codevector created by selecting only N components smaller than n and a codebook of a time variation vector of the similarity are included. 5. The method according to claim 4, wherein a codebook using a code vector created by selecting only M components smaller than n is used when creating the registered word dictionary. 8. A coded dictionary word pattern from a previously created registered word dictionary is decoded using a codebook previously obtained from a similarity vector of learning data, and the decoded dictionary word pattern is decoded. And a speech recognition method for recognizing the uttered input speech using the speech recognition method. 9. An uttered input voice has m (m is an integer) acoustic feature parameters obtained for each frame and n types (n is an integer) of standard patterns created in advance from learning data. Matching with m feature parameters is performed, n similarities are obtained for each frame, and an n-dimensional similarity vector sequence is obtained. This is used for each frame using a codebook previously obtained from the similarity vector of the learning data. A speech recognition method characterized by reconstructing a dictionary word pattern from a registered word dictionary created in advance, and recognizing the input speech using the dictionary word pattern. 10. An m-type (m is an integer) acoustic feature parameter obtained for each frame based on the uttered input voice,
N types (n is an integer) created in advance from learning data
Is performed with m feature parameters of each standard pattern, and n similarities are obtained for each frame.
As a dimensional similarity vector series, a time variation of similarity is obtained for each frame for each of the n types of similarity time series, and an n-dimensional vector sequence of this similarity time variation is calculated as Using a codebook obtained in advance from an n-dimensional vector of the time variation of the similarity of the learning data, a dictionary word pattern is restored from a registered word dictionary created in advance for each frame, and using this dictionary word pattern, A speech recognition method characterized by recognizing input speech. 11. An m-type (m is an integer) acoustic feature parameter obtained for each frame based on an uttered input voice;
N types (n is an integer) created in advance from learning data
Is performed with m feature parameters of each standard pattern, and n similarities are obtained for each frame.
For each of the n types of time series of similarities, a time change amount of the similarity is obtained for each frame, and the n-dimensional similarity vector series and the time of the similarity are calculated. The sequence of the n-dimensional vector of the change amount is 1
Using the codebook obtained in advance from the similarity of the learning data, the combined word pattern of the input voice is restored as a dictionary word pattern from a registered word dictionary created in advance for each frame, and this dictionary word pattern is used. And recognizing the input voice. 12. The speech recognition method according to claim 8, wherein the registered word dictionary uses a dictionary created by the method according to any one of claims 1 to 7. 13. A registrant utters dictionary registration words, and acoustic analysis means for obtaining m (m is an integer) acoustic feature parameters obtained for each frame, and n types (n is an integer) created from learning data in advance. A standard pattern storage unit that stores m feature parameters of each of the (integer) standard patterns, m feature parameters from the standard pattern storage unit, and m (m is an integer) sounds from the acoustic analysis unit Using similarity calculating means for matching characteristic feature parameters, obtaining n similarities for each frame, and obtaining an n-dimensional similarity vector sequence, and using a codebook previously obtained from the similarity vector of the learning data. A similar vector encoding means for encoding each frame, and a code for registering the encoded code sequence as a dictionary word pattern in a registered word dictionary. A registered word dictionary creating apparatus, comprising: a registered word dictionary storage unit. 14. A registrant utters dictionary registration words, and acoustic analysis means for obtaining m (m is an integer) acoustic feature parameters obtained for each frame, and n types (n is an integer) created in advance from learning data. A standard pattern storage unit that stores m feature parameters of each of the (integer) standard patterns, m feature parameters from the standard pattern storage unit, and m (m is an integer) sounds from the acoustic analysis unit Similarity calculating means for performing matching of characteristic feature parameters, obtaining n similarities for each frame, and obtaining an n-dimensional similarity vector series, and a similarity measure for each of the n types of time series of similarities Regression coefficient calculating means for calculating the time change amount of each frame, and obtaining a series of n-dimensional vectors of the time change amount of the similarity, Regression coefficient vector encoding means for encoding each frame using a codebook obtained from an n-dimensional vector of the amount of change, and a code for registering the encoded code sequence in a registered word dictionary as a dictionary word pattern A registered word dictionary creation device, comprising: a registered word dictionary storage unit. 15. A coded registered word dictionary storage means, comprising: a code sequence obtained by encoding the n-dimensional similarity vector sequence using a codebook; and an n-dimensional vector sequence of the time change amount of the similarity. 15. A code sequence obtained by encoding a code sequence using a code book is registered and registered as a dictionary word pattern.
Registered word dictionary creation device described in. 16. An acoustic analysis means for obtaining m (m is an integer) acoustic feature parameters obtained for each frame from an uttered input speech, and n kinds (n is an integer) of pre-created learning data. Standard pattern storage means for registering m feature parameters of each standard pattern;
A similarity calculation means for matching the m (m is an integer) acoustic feature parameters with the m feature parameters, obtaining n similarities for each frame, and obtaining an n-dimensional similarity vector sequence Word pattern restoring means for restoring the n-dimensional similarity vector series as a dictionary word pattern from a previously created registered word dictionary for each frame using a codebook previously obtained from the similarity vector of the learning data; A speech recognition device comprising: a word recognition unit that recognizes the input speech using the restored dictionary word pattern. 17. An acoustic analysis means for obtaining m (m is an integer) acoustic feature parameters obtained for each frame from an uttered input speech, and n kinds (n is an integer) of pre-created learning data. Standard pattern storage means for registering m feature parameters of each standard pattern;
A similarity calculation means for matching the m (m is an integer) acoustic feature parameters with the m feature parameters, obtaining n similarities for each frame, and obtaining an n-dimensional similarity vector sequence Regression coefficient calculating means for obtaining a time change amount of the similarity for each frame for each of the n types of similarity time series, and obtaining an n-dimensional vector series of the time change amount of the similarity, A word for which the n-dimensional similarity vector series is restored as a dictionary word pattern from a registered word dictionary created in advance for each frame, using a codebook previously obtained from an n-dimensional vector of the time variation of the similarity of the learning data. Pattern restoration means;
A speech recognition device comprising: a word recognition unit that recognizes the input speech using the restored dictionary word pattern. 18. An acoustic analysis means for obtaining m (m is an integer) acoustic feature parameters obtained for each frame from an uttered input speech, and n kinds (n is an integer) of pre-created learning data. Standard pattern storage means for registering m feature parameters of each standard pattern;
A similarity calculation means for matching the m (m is an integer) acoustic feature parameters with the m feature parameters, obtaining n similarities for each frame, and obtaining an n-dimensional similarity vector sequence Regression coefficient calculating means for obtaining a time change amount of the similarity for each frame for each of the n types of similarity time series, and obtaining an n-dimensional vector series of the time change amount of the similarity, A word pattern as a word pattern of an input voice, by combining the n-dimensional similarity vector series from the similarity calculating means and the n-dimensional vector series of the time variation of similarity from the regression coefficient calculating means into one Using a creating means and a codebook which previously obtained the word pattern of the input voice from an n-dimensional vector of the time change amount of the similarity of the learning data, for each frame. A word pattern restoring means for restoring the word dictionary as the dictionary word pattern, using the restored dictionary word pattern, the speech recognition apparatus characterized by comprising a word recognition means for recognizing the input speech. 19. The registered word dictionary according to claim 13, wherein
19. A registered dictionary created by the registered word dictionary creating apparatus according to any one of claims 16 to 18.
The speech recognition device according to any one of the above.