JPH09258769A - Speaker adaptation method and speaker adaptation device - Google Patents

Abstract

(57) [Abstract] [Problem] In speaker adaptation by an unspecified speaker codebook created based on the voices of many unspecified speakers, there is a problem that an accurate speaker adaptation cannot be performed for an input speaker. there were. SOLUTION: A plurality of speaker classes are set based on voice characteristic data obtained from a large number of unspecified speakers, and an unspecified speaker code is obtained based on the speech characteristic data of speakers belonging to each speaker class. The books 23a and 23b are created, and the centroid vector sequence for each speaker adaptation word is obtained from the speech feature vector sequences of the speakers belonging to each speaker class and stored in the centroid vector string storage units 24a and 24b. . Then, the speaker feature determination unit 22 determines which speaker class the voice of the input speaker belongs to by associating the voice feature vector sequence of the input speaker with the center-of-gravity vector sequence, and the corresponding unspecified talk A speaker codebook is selected, and at the time of voice recognition, the voice of the input speaker is coded by the selected codebook and voice recognition is performed using the corresponding voice model.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speaker adaptation method and a speaker adaptation device applied to speech recognition using vector quantization.

[0002]

2. Description of the Related Art In speech recognition using vector quantization, speech recognition with high recognition performance is possible by performing speaker adaptation. As one of the conventional speaker adaptation methods in speech recognition using this vector quantization, when the speech of the input speaker is input, the input speech is based on a codebook created by a certain standard speaker. There is a method of converting a feature vector sequence of a speaker to a feature vector sequence of a standard speaker and outputting the sequence.

This will be described with reference to FIG. FIG.
(A) is a speech feature vector sequence of an input speaker, (b) is an input speaker codebook created in advance based on an unspecified speaker codebook described later, and (c) is a large number. It is an unspecified speaker codebook created based on the voice feature data of the unspecified speaker.

The speech of the input speaker is usually converted into a digital signal by an A / D converter and then frequency-analyzed to obtain a feature vector (LPC-LPC) representing the frequency characteristics of the voice waveform signal. The CEPSTRUM coefficient is generally output), but here, in order to simplify the explanation, it is shown as a five-dimensional feature vector sequence. The size of the unspecified speaker codebook is usually 256 or 512, but the size is 3 here. Then, the input speaker codebook and the unspecified speaker codebook are associated with respective data in advance,
For example, the data of A of the input speaker codebook is stored in A ′ of the unspecified speaker codebook and B of the input speaker codebook.
Is associated with B ′ of the unspecified speaker codebook, and C data of the input speaker codebook is associated with C ′ of the unspecified speaker codebook.

Now, the feature vector sequences (a), (b), (c), ... Of the input speaker's input voice as shown in FIG. 16 (a).
.. is input, the feature vector sequence (a), (b), (c), ... Which data in the input speaker codebook shown in FIG. Calculated by distance calculation. For example, the input voice (a) data (3,2,0,0) is closest to the input speaker codebook data A (3,2,0,0), and the input voice (b) The data (2.1.1.1.1) is the closest to the input speaker codebook data B (1.1.1.1.1.1), and the input voice (c) data (1.2.1.1) 1.1) is the input speaker codebook data B (1.1.1.1.1
1) is the closest to the input voice (d) data (0.0.2.
・ 2 ・ 2) is the data C (0.0) of the input speaker codebook
・ It is the closest to 0 ・ 2 ・ 2), and the input voice (e) data (0 ・ 0 ・ 0 ・ 2 ・ 3) is the same as the input speaker codebook data C (0 ・ 0 ・ 0 ・ 2 ・ 2). The closest thing is required.

In this way, the closest feature vector is selected by referring to the input speaker codebook for each data of the input voice. Therefore, in this case, considering only the input voices (a) to (e), A, B, B, C, C
The feature code vector is calculated.

The input speaker codebook is associated with the unspecified speaker codebook such that A is A ', B is B', and C is C '. The voice is A '・ B' ・ of the unspecified speaker codebook.
It will be converted into data of B ', C', and C '. In this case, the data A'of the unspecified speaker codebook is (3, 3, 0, 0, 0), the data B'is (2, 2, 2, 1, 1), and the data C is 'Is (0.0
・ It is 0 ・ 3 ・ 3).

As described above, the feature vector sequence of the input voice can be converted into the feature vector sequence of the unspecified speaker codebook, and the converted feature vector sequence is transmitted to the voice recognition processing section.

[0009]

In the speaker adaptation method shown in FIG. 16, an input speaker codebook is created based on an unspecified speaker codebook created from a large number of unspecified speakers,
Speaker adaptation was performed using this input speaker codebook. However, in speaker adaptation using an input speaker codebook created from a large number of unspecified speakers, accurate speaker adaptation cannot be expected so much, and as a result, speech using a speaker-adapted code vector can be expected. The recognition has a problem that sufficient recognition performance cannot be obtained. As a means for solving this, a voice model and its codebook are prepared for many speakers (for example, 500 people), and a voice model and a codebook whose voice features are close to the input speaker are selected. A speaker adaptation method may be considered, but this method requires preparing several voice models and codebooks corresponding to the voice characteristics of the input speaker, which is not practical to apply to a small and inexpensive device. However, it was not practical.

Therefore, according to the present invention, in the speaker adaptation process as a pre-stage of the voice recognition, a large number of unspecified speakers are classified based on the voice feature data, and the voice feature data belonging to each class is classified. Create a codebook using
When the speaker is adapted, it is determined which class the speech of the input speaker belongs to, and the codebook is selected according to the result of the determination. When speech is recognized, the input code is selected using the selected codebook. For the purpose of realizing a speaker adaptation method and a speaker adaptation device that enable speech recognition at a high recognition rate by encoding the speech of a person and recognizing the speech using a corresponding speech model. There is.

[0011]

According to a speaker adaptation method of the present invention, a plurality of speakers are classified based on voice feature data obtained from an unspecified number of speakers, as set forth in claim 1. Based on the result obtained by comparing the data created based on the voice feature vector sequence of the speaker belonging to each speaker class of the class with the voice feature data of the input speaker, the voice of the input speaker is determined. It is determined whether the speaker belongs to a speaker class, and a voice model for voice recognition corresponding to the determined speaker class is selected based on the determination result, and the voice of the input speaker is selected at the time of voice recognition. It is characterized by performing voice recognition processing using a voice model.In this way, it is determined which speaker class the input speaker belongs to based on the voice feature data of the input speaker, and Speech recognition using the corresponding speech model So it can be expected highly accurate speech recognition with a simple process.

Further, according to the invention of claim 2, for each of a plurality of speaker classes classified based on the voice characteristic data obtained from an unspecified number of speakers, the voices of the speakers belonging to the respective speaker classes. An unspecified speaker codebook is created based on the feature data, and the data created based on the voice feature vector sequence of the speakers belonging to each speaker class and the voice feature data of the input speaker are compared. Based on the obtained result, it is determined which speaker class the voice of the input speaker belongs to, and based on the determination result, the unspecified speaker codebook and the speaker corresponding to the determined speaker class. A voice model for voice recognition corresponding to a class is selected, an input speaker codebook is created based on the selected unspecified speaker codebook, and the voice of the input speaker is created at the time of voice recognition. Input speaker code After vector quantized using a book and selected unspecified speaker code book, sent to the speech recognition unit, characterized in that the speech recognition using a speech model wherein the selected speech recognition unit.

According to the second aspect of the present invention, the input speaker codebook is created by using the unspecified speaker codebook created based on the voices of the speakers belonging to the selected speaker class.
At the time of voice recognition, vector quantization is performed using the created input speaker codebook and the selected unspecified speaker codebook, and the result is sent to the voice recognition unit, and the voice recognition unit performs voice recognition using the selected voice model. As a result, more accurate speaker adaptation becomes possible, and recognition with a high recognition rate becomes possible.

And in the above-mentioned claim 1 or 2,
The process of determining which speaker class the voice of the input speaker belongs to based on the feature vector sequence is performed by using the voice feature vector sequence of the speaker belonging to each speaker class, and the centroid vector for each word for speaker adaptation. After obtaining the sequence, the distance between the center of gravity vector sequence of each word and the feature vector sequence of the voice of the input speaker is obtained by DP matching for each speaker class, and which speaker the voice of the input speaker is based on the distance. Determine if it belongs to a class.

With this, it is possible to accurately determine which speaker class the voice of the input speaker belongs to by a simple process.

Further, in the above-mentioned claim 1 or 2, the processing for determining to which speaker class the voice of the input speaker belongs based on the feature vector sequence is the voice feature of the speaker belonging to each speaker class. A voice model corresponding to each speaker class by the DRNN method is created based on the vector sequence, and a numerical value indicating the certainty of the existence of a predetermined word is output from the voice feature vector sequence of the input speaker and the voice model. However, it is also possible to determine to which speaker class the voice of the input speaker belongs based on the numerical value indicating the certainty.

According to this, it is possible to cut out the speaker adaptation word included in the input voice by the keyword spotting process, and it is easy to determine which speaker class the voice of the input speaker belongs to. The processing can be performed with high accuracy.

Further, in the above-mentioned claim 1 or 2,
The process of determining which speaker class the voice of the input speaker belongs to based on the feature vector sequence is performed by using the voice feature vector sequence of the speaker belonging to each speaker class, and the centroid vector for each word for speaker adaptation. A sequence is obtained, and for each speaker class, the distance between the center of gravity vector sequence of each word and the feature vector sequence of the voice of the input speaker is obtained by DP matching, and the voice feature of the speaker belonging to each speaker class is obtained. A voice model corresponding to each speaker class by the DRNN method is created based on the vector sequence, and a numerical value indicating the certainty of the existence of a predetermined word is obtained from the voice feature vector sequence of the input speaker and the voice model. It is also possible to determine which speaker class the voice of the input speaker belongs to based on the calculated DP matching distance and the numerical value indicating the certainty of the existence of the predetermined word.

According to this, it is possible to judge to which speaker class the voice of the input speaker belongs by a simple process with higher accuracy.

According to the invention of claim 6, in any one of claims 1 to 5, the processing for classifying into a plurality of speaker classes based on the voice characteristic data obtained from the unspecified number of speakers is performed. , A DP matching distance is calculated for each word between the speakers for a feature vector sequence for each word obtained from an unspecified number of speakers, and the sum of the distances is calculated. It is characterized in that the inter-speaker distances are determined, the inter-speaker distances are obtained among the unspecified large number of speakers, and the classification is performed based on the inter-speaker distances.

According to this, it is possible to perform highly accurate speaker classification based on the voice feature data, and an unspecified speaker codebook created from speakers belonging to the speaker class thus classified is used. By adapting the speaker
Precise speaker adaptation is possible.

Further, as described in claim 7, the speaker adaptation apparatus of the present invention has a plurality of speaker classes classified based on voice feature data obtained from an unspecified number of speakers. Feature data storage means for storing data created based on the voice feature vector sequence of the speaker belonging to each speaker class, input data storage means for storing the feature vector sequence of the voice of the input speaker, and the input speaker Based on the result obtained by comparing the voice feature data for a certain word and the feature data for the word stored in the feature data storage means, it is determined which speaker class the voice of the input speaker belongs to, And a speaker class determination processing unit that selects a voice model for voice recognition corresponding to the determined speaker class based on the determination result, and the voice of the input speaker is selected as the voice of the input speaker during voice recognition. Mo After vector quantized using Le, characterized in that passed to the speech recognition unit.

As described above, the speaker class to which the input speaker belongs is determined based on the voice feature data of the input speaker, and the voice recognition is performed using the voice model corresponding to the speaker class. Highly accurate voice recognition can be expected with simple processing.

Further, according to the invention of claim 8, the voices of the speakers belonging to each of the plurality of speaker classes are classified for each of the plurality of speaker classes classified based on the voice feature data obtained from an unspecified number of speakers. An unspecified speaker codebook created based on the feature data, the data created based on the voice feature vector sequence of the speakers belonging to each speaker class, and the voice feature data of the input speaker are compared. It is determined which speaker class the input speaker's voice belongs to based on the result obtained from the above, and based on the determination result, the unspecified speaker codebook and the speaker corresponding to the determined speaker class. A speaker class determination processing unit that selects a voice model for voice recognition corresponding to a class, and an input speaker codebook created based on an unspecified speaker codebook selected by this speaker class determination processing unit. Have and sound At the time of recognition, the voice of the input speaker is vector-quantized by using the created input speaker codebook and the selected unspecified speaker codebook, and then sent to the voice recognition unit. The feature is that the voice recognition is performed by using the voice model.

According to the invention of claim 8, the input speaker codebook is created by using the unspecified speaker codebook created based on the voices of the speakers belonging to the selected speaker class.
At the time of voice recognition, vector quantization is performed using the created input speaker codebook and the selected unspecified speaker codebook, and the result is sent to the voice recognition unit, and the voice recognition unit performs voice recognition using the selected voice model. As a result, more accurate speaker adaptation becomes possible, and recognition with a high recognition rate becomes possible.

And in the above-mentioned claim 7 or 8,
The means for determining which speaker class the voice of the input speaker belongs to based on the feature vector sequence is a centroid vector for each word for speaker adaptation from the voice feature vector sequence of the speaker belonging to each speaker class. Obtaining a column, having a storage unit for storing the center of gravity vector for each word for each speaker class, for each speaker class by the speaker class determination processing unit, the center of gravity vector string of each word and The distance between the voice of the input speaker and the feature vector sequence is obtained by DP matching, and it is determined to which speaker class the voice of the input speaker belongs.

Thus, it is possible to accurately determine which speaker class the voice of the input speaker belongs to by a simple process.

The means for determining which speaker class the input speaker's voice belongs to on the basis of the feature vector sequence is a speech feature of a speaker belonging to each speaker class. A numerical value indicating the certainty of the existence of a predetermined word is output from a voice model created for each speaker class by the DRNN method based on a vector sequence, a voice feature vector sequence of an input speaker, and the voice model. The speaker class determination processing unit determines which speaker class the voice of the input speaker belongs to, based on the numerical value indicating the certainty output from the word detection unit. It is also possible to do so.

According to this, it is possible to cut out a speaker adaptation word included in the input voice by the keyword spotting process, and it is easy to determine which speaker class the voice of the input speaker belongs to. The processing can be performed with high accuracy.

Further, in the above-mentioned claim 7 or 8,
The process of determining which speaker class the voice of the input speaker belongs to based on the feature vector sequence is performed by using the voice feature vector sequence of the speaker belonging to each speaker class, and the centroid vector for each word for speaker adaptation. A column is obtained, and a storage unit for storing the centroid vector of each word for each speaker class is provided, and each of the DRNNda method is used based on the voice feature vector sequence of the speaker belonging to each speaker class. A voice model created corresponding to the speaker class, a voice detection unit that outputs a numerical value indicating the certainty of the existence of a predetermined word from the voice feature vector sequence of the input speaker and the voice model, A distance obtained by DP matching for the distance between the centroid vector sequence of each word and the feature vector sequence of the voice of the input speaker for each user class, and a number indicating the certainty output from the word detection unit. Based on, by the speaker class determination unit, it is possible to determine belongs to which speaker class speech input speaker.

According to this, it is possible to determine to which speaker class the voice of the input speaker belongs by a simple process with higher accuracy.

The invention of claim 12 is the same as that of claim 7
11 to 11, the process of classifying into a plurality of speaker classes based on the voice feature data obtained from the unspecified number of speakers is performed by each speaker obtained from the unspecified number of speakers. For a plurality of feature vector strings for each word, a DP matching distance is obtained for each word between speakers, and the sum of the distances is set as the distance between the speakers, and the unspecified number of speakers The feature is that each inter-speaker distance is obtained with, and classification is performed based on this inter-speaker distance.

According to this, it becomes possible to perform highly accurate speaker classification based on the voice feature data, and an unspecified speaker codebook created from speakers belonging to the speaker class thus classified is used. By adapting the speaker
Precise speaker adaptation is possible.

[0034]

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

(First Embodiment) FIG. 1 is a block diagram for explaining a schematic configuration of a voice recognition device to which the present invention is applied. The configuration is roughly divided into a voice input unit 1,
It is composed of a speaker adaptation unit 2 and a voice recognition unit 3.

The voice input unit 1 includes a microphone 11,
A / D converter 12 for A / D converting the voice input from the microphone 11, the A / D converted voice waveform signal,
Frequency analysis is performed for each short time using an arithmetic unit, and a number-dimensional feature vector (generally LPC-CEPSTRUM coefficient) that represents the feature of the frequency is extracted, and the time series of this feature vector (hereinafter, the voice feature vector sequence). Is output).

The speaker adaptation unit 2 is a part of the present invention, and determines which speaker class the voice of the user (input speaker) belongs to, and the result of the determination
A corresponding codebook (an unspecified speaker codebook created by using voice feature data of speakers belonging to the determined speaker class) is selected, and at the time of voice recognition, the selected unspecified speaker is selected. After the user's voice is vector-quantized by using the codebook, the code vector is output to the voice recognition unit 3. The speaker adaptation unit 2 will be described below.

The speaker adaptation section 2 is a vector quantization section 2
0, input data storage unit 21, speaker class determination processing unit 2
2, pre-classified unspecified speaker codebooks for each speaker class 23a, 23b, ..., Center-of-gravity vector string storage unit 24a for each pre-classified speaker class,
24b, ...

The input data storage unit 21 stores a feature vector sequence of words for speaker adaptation spoken by the user. In other words, the user from the system side, as a word for the speaker adaptation, for example, "Please talk to good morning", gives an instruction such as "Please talk to Hello", according to the instruction, when the user utterance, The voice is input through the microphone 11, A / D converted by the A / D converter 12, and then voice analyzed by the voice analyzer 13 to output a voice feature vector sequence. The voice feature vector sequence for each word is stored.

The unspecified speaker codebook 23
a, 23b, ... For each of a plurality of speaker classes classified based on the voice characteristic data obtained from a large number of unspecified speakers, the voice characteristic data of the speakers belonging to each speaker class are shown. For example, the unspecified speaker codebook 23a is a codebook created based on the voice feature data of unspecified large numbers of males, and the unspecified speaker codebook 23b is the unspecified large number of females. It is a codebook created based on the voice feature data of.

The speaker classes, which are classified into several types, have higher performance if they are finely divided based on the voice feature data of a large number of speakers, but the description will be simplified here. For this reason, an example of classification into two types, male and female, will be described. Further, here, the unspecified speaker codebook 23a is referred to as a first unspecified speaker codebook, and the unspecified speaker codebook 23b is referred to as a second unspecified speaker codebook.

Further, the centroid vector sequence storage units 24a, 24
Similarly to the above, b, ... For each of a plurality of speaker classes classified based on the voice feature data obtained from a large number of unspecified speakers, a speaker belonging to the speaker class is associated with each word. The center of gravity vector sequence obtained from the speech feature data obtained by uttering is stored, and this will be described below.

By the way, the center-of-gravity vector sequence is a feature vector (for example, 10-dimensional) obtained by uttering a certain word for each speaker class by a large number of speakers and analyzing the voice for each short time. Is a vector sequence obtained by averaging the feature vectors for each speaker at each time.
This will be described with reference to FIG. FIG. 2 is an example in which the feature vector sequence is obtained by causing the four speakers, males A, B, C, and D, to speak the word “good morning”.

The center of gravity vector sequence of the feature vector sequence for the word "good morning" from the four speakers A, B, C, D is obtained as follows.

First, in FIG. 2, for example, the 10-dimensional feature vector at time t1 in the speech uttered by the person A is Ca1, and the 10-dimensional feature vector at time t2 is Ca.
2, the 10-dimensional feature vector at time t3 is Ca3, ...
The 10-dimensional feature vector at time t1 in the speech uttered by the person B is Cb1, the 10-dimensional feature vector at time t2 is Cb2, the 10-dimensional feature vector at time t3 is Cb3, ... C,
Of the voice uttered by the person Cc1 is a 10-dimensional feature vector at time t1, Cc2 is a 10-dimensional feature vector at time t2, and Cc is a 10-dimensional feature vector at time t3.
3 ..., the 10-dimensional feature vector at time t1 in the voice uttered by the person D is Cd1, the 10-dimensional feature vector at time t2 is Cd2, and the 10-dimensional feature vector at time t3 is Cd3. , ..., and so on.

As described above, even when the same word "Ohayo" is uttered, the feature vector sequence for the words "Ohayo" by the persons A, B, C, and D is long in time depending on the personality of each person. Differences occur in the pod feature vector.

Next, the center of gravity vector of the person A, B, C, and D for the word "Ohayo" is obtained at each time. When obtaining the center of gravity vector, It is necessary to normalize the temporal length of the feature vector sequence of, that is, make the number of each feature vector the same. In order to do this, one of the feature vector sequences is selected as the reference vector sequence, and DP matching with the reference vector sequence is performed for normalization. This normalization process will be described with reference to FIG.

In FIG. 3, in order to make the explanation easy to understand, the sampling times of the feature vector sequence of A are t1, t2, and t3.
And the sampling time of the feature vector sequence of B is four of t1, t2, t3, t4, and the sampling time of the feature vector sequence of C is t1, t2, t3, t4, t5.
And the sampling time of the feature vector sequence of D is six times t1, t2, t3, t4, t5, and t6, and the feature vector sequence of B is the reference vector sequence here.

Then, with respect to the feature vectors Cb1, Cb2, Cb3, Cb4 of the reference B feature vector sequence at times t1, t2, t3, t4, the feature vector Ca of the A feature vector sequence at each time.
Feature vectors Cc1, Cc2, Cc3, C at respective times in the feature vector sequence of 1, Ca2, Ca3, C
Feature vectors Cd1, Cd2, Cd3, Cd4 at respective times in the feature vector sequence of c4, Cc5, D
Cd5 and Cd6 are associated with each other by DP matching.

In this way, the feature vectors at other times of the reference feature vector sequence are associated with the feature vectors of other feature vector sequences by DP matching, thereby normalizing the number of feature vectors. be able to.

That is, in the reference B feature vector sequence and the A feature vector sequence, Cb1 is associated with Ca1, Cb2 is associated with Ca2, and Cb2 is associated with C2.
Correspondence is made such that Ca3 is respectively associated with b3 and Cb4, and the reference vector sequence of B and the feature vector sequence of C are Cc with respect to Cb1.
1, Cc2 for Cb2, C for Cb3
Cc5 is assigned to c3 and Cb4, respectively. Further, the reference vector sequence of B and the feature vector sequence of D are Cd1 for Cb1 and Cd1 for Cb2.
2, Cd4 for Cb3, Cd for Cb4
6 are associated with each other.

As described above, the number of feature vectors is normalized by associating the reference feature vector sequence with the other feature vector sequences by DP matching. Then, the centroid vector is obtained for each of the associated feature vectors.

Any method may be used as the method for obtaining the center of gravity vector, but here, the center of gravity vector is obtained as follows.

Feature vectors Ca1 and C at time t1
10-dimensional L of each of b1, Cc1, and Cd1
The PC cepstrum coefficient is Ca1 = (Ca10, Ca11, ..., Ca19) Cb1 = (Cb10, Cb11, ..., Cb19) Cc1 = (Cc10, Cc11, ..., Cc19) Cd1 = (Cd10, Cd11) , ..., Cd19), which is composed of an average value for each dimension.
Let the zero-dimensional LPC cepstrum coefficient be the center of gravity vector at time t1. That is, the average value Cα10 of the first dimension
Is Cα10 = (Ca10 + Cb10 + Cc10 + Cd1
0) / 4 The average value Cα11 of the second dimension is Cα11 = (Ca11 + Cb11 + Cc11 + Cd1
1) / 4 The average value Cα19 of the 10th dimension is Cα19 = (Ca19 + Cb10 + Cc19 + Cd1
9) / 4. 1 at time t1 thus obtained
Average of 0-dimensional LPC cepstrum coefficients (Cα10, Cα
, ..., Cα19) is the center of gravity vector at time t1 and is represented by Cs1. Similarly, at time t
2, t3, ... Centroid vectors Cs2, Cs
Ask for 3, ... The center-of-gravity vector sequence composed of the center-of-gravity vectors Cs1, Cs2, Cs3, ... Determined in this way is represented by a chain line in FIG. Is represented by a white circle.

The above is the case of obtaining the center-of-gravity vector sequence of the men A, B, C, D for the word "Ohayo". The same applies to several speaker adaptation words other than "Ohayo". Then, the centroid vectors of a large number of men are obtained and the centroid vector sequences for these several words are stored in the centroid vector sequence storage unit 24a (hereinafter referred to as the first centroid vector sequence storage unit 24a) of FIG. Similarly, for some speaker adaptation words spoken by many women,
The center of gravity vector is obtained, and the center of gravity vector sequence for each of these several words is calculated as the center of gravity vector sequence 24b (hereinafter,
The second center-of-gravity vector sequence storage unit 24b). Here, the classification is made only for males and females, but if it is further subdivided, the centroid vector sequences for some words obtained for each class are stored in the corresponding centroid vector sequence storage unit. .

Further, the speaker class determination processing unit 22 is
The input voice feature vector stored in the input data storage unit 21 and the first and second center-of-gravity vector sequence storage units 24a, 24
The contents of b are associated with each other by DP matching, the speaker class to which the input voice belongs is determined based on the result, and an unspecified speaker codebook is selected accordingly. In this case, since there are two types of speaker classes, male and female, the unspecified speaker codebook is, as described above, the first unspecified speaker codebook 23a and the second unspecified speaker. Two types of codebooks 23b are prepared, and either the first or the second unspecified speaker codebook is selected. The speaker class determination processing unit 22 will be described below.
Will be described.

The input data storage unit 21 stores a feature vector sequence for the voice uttered by the user based on an instruction from the system side. For example, the system says, "Please say good morning," and the user says "Good morning," and then "Hello," when you say "Please say hello." , The user speaks based on an instruction from the system side.

Then, the feature vector sequence for "Ohayo" of the user and the center of gravity vector sequence for "Ohayo" stored in the first center-of-gravity vector sequence storage unit 24a are set to D.
Correspondence is made by P matching, and the distance between the two is obtained. Similarly, the feature vector sequence for "Ohayo" of the user and the center of gravity vector sequence for "Ohayo" stored in the second center-of-gravity vector sequence storage unit 24b are associated by DP matching, and the distance between the two is obtained.

Similarly, the feature vector sequence for the user's "Hello" and the first centroid vector sequence storage unit 2
The center of gravity vector sequence for "konnichiwa" stored in 4a is associated by DP matching, the distance between the two is obtained, and the feature vector sequence for "konnichiwa" of the user and the second center of gravity vector sequence storage unit 24b are stored. a centroid vector sequence was mapped by DP matching for the stored "hello", the distance between them (DP
Called distance). In this way, the user's feature vector sequence and
Of the word stored in the center-of-gravity vector sequence storage unit 24a and the second center-of-gravity vector sequence storage unit 24b
Find the P distance.

Here, the words for speaker adaptation are the words w1 and w.
There are three, 2 and w3. Then, the feature vector sequence of the user for the word w1 and the first centroid vector sequence storage unit 24
The DP distance of the centroid vector sequence for the word w1 stored in a is da1, the user feature vector sequence for the word w1 and the DP of the centroid vector sequence for the word w1 stored in the second centroid vector sequence storage unit 24b. The distance is db1, the user feature vector sequence for the word w2 and the DP distance of the center-of-gravity vector sequence for the word w2 stored in the first center-of-gravity vector sequence storage section 24a are the user feature vector sequence for the distance da2, word w2, and The DP distance of the centroid vector sequence for the word w2 stored in the second centroid vector sequence storage unit 24b is db2, the user feature vector sequence for the word w3, and the word w3 stored in the first centroid vector sequence storage unit 24a. The DP distance of the center-of-gravity vector sequence with respect to DP centroid vector sequence for the word w3 stored in the second centroid vector sequence storage unit 24b
If the distance is db3, the total DP distance of the respective barycentric vector sequences for the user's voice feature vector sequence and the words w1, w2, w3 stored in the first centroid vector sequence storage unit 24a is Da. Then, Da is obtained by Da = da1 + da2 + da3 (1), and similarly, for the user's voice feature vector sequence and the words w1, w2, and w3 stored in the second centroid vector sequence storage unit 24b, respectively. When the total DP distance of the center-of-gravity vector sequence of is Db, Db is obtained by Db = db1 + db2 + db3 (2)

Each D obtained as described above
The magnitudes of the sums Da and Db of the P distances are compared, and the smaller one is selected. For example, if Da> Db, it is determined that the user's voice is female in this case, and thus the second unspecified speaker codebook 23b created based on the voices of many females is used. select.

In this way, the speaker class determination processing unit 2
According to 2, it is determined which speaker class the voice of the input speaker belongs to, and when the speaker class is determined, an unspecified speaker codebook created based on the voice of the speaker class is selected. To be done.

As a result, when recognizing the voice of the input speaker, the voice of the input speaker is vector-quantized using the unspecified speaker codebook created based on the voice close to the voice of the input speaker. Since this is output as a feature code vector, the voice recognition unit 3 can perform highly accurate recognition processing.

By the way, the voice recognition unit 3 stores the voice feature data for the recognizable word (registered word) in this case, based on the voices of a large number of men. 31a (hereinafter referred to as a first voice model storage unit 31a), a voice feature data storage unit 31b (hereinafter referred to as a second voice model storage unit) that also stores voice feature data created based on a large number of female voices. Section 31b), a word detection section 32 that outputs word detection data from the contents of any one of these speech model storage sections and the quantized feature code vector, based on the detection data from this word detection section 32. The voice recognition processing unit 33, which outputs the recognition data for the input voice, is configured. It should be noted that one of the first voice model storage unit 31a and the second voice model storage unit 31b is selected based on the determination result of the speaker class determination unit 22, and during speaker adaptation, When it is determined that the voice of the input speaker belongs to the female speaker, the second voice model storage unit 3
When 1b is selected and it is determined that the voice of the input speaker is male, the first voice model storage unit 31a is selected.

As described above, in the speaker adaptation process, it is determined whether the voice of the input speaker belongs to the male speaker class or the female speaker class, and at the time of voice recognition, the result of the classification is obtained. Speech recognition is performed using a speech model based on. In this case, the voice of the input speaker is determined to belong to the female speaker class, is converted into a feature code vector using the second unspecified speaker codebook 23b, and is output. Voice model storage unit 31b
Since voice recognition is performed using the contents of, high-accuracy voice recognition can be performed.

That is, it is determined which speaker class the input voice belongs to, and an unspecified speaker codebook created by using the voices of the speakers belonging to the speaker class is selected and set as a code vector. A series of processes in which the voice recognition unit 3 performs the voice recognition process by using the voice model corresponding to the determined speaker class is a process equivalent to performing the speaker recognition and then performing the voice recognition. Will be done.

The above description is an example in which the speaker class is determined using the DP distance between the feature vector of the input speaker voice and the center of gravity vector for some words for speaker adaptation, but the present invention is not limited to this. No, DR that the applicant has already applied for a patent
Speech recognition technology based on the NN (Dynamic Recurrent Neural Network) method. (For this technology, the applicant of the present invention has disclosed Japanese Patent Application Laid-Open Nos. 6-4097 and 6-119.
Patent applications have already been filed based on 476, etc. ), The speaker class may be determined based on a value (this value will be described later).

The speech recognition technique based on the DRNN method will be briefly described with reference to FIG.

The speech recognition technology based on this DRNN system is
For example, a recognizable word that has been registered in advance from a continuous voice such as “Good morning. Today is a nice weather” (in this case, “Good morning”, “Today”, “Weather”, etc.) is used as a keyword. It is possible to obtain a value indicating in which part of the input speech these keywords words are present and with certainty, and to understand continuous speech as described above based on the value indicating the certainty. It is possible.

In this embodiment, when the speaker class is selected, the user speaks some words for speaker adaptation based on an instruction from the system side. Now, based on an instruction from the system side, for example, the user utters "Good morning", and then based on an instruction from the system side, for example, the user utters "Hello" and
It is assumed that an audio signal as shown in (a) is output. It is assumed that these "good morning" and "konchiwa" are words that are keywords registered in advance.

Then, assuming that the words which are the keywords registered in advance are, for example, 10 words, each word is associated with these 10 words (this is referred to as word 1, word 2, word 3, ...). A signal for detection is output, and the degree of certainty that the registered word exists in the input voice is detected from information such as the value of the detection signal. That is, when the word "Ohayo" (word 1) is present in the input voice, the detection signal waiting for the signal "Ohayo" is the input voice uttered by the user as shown in FIG. Get up in the "Good morning" part. Similarly, when the word "Hello" (word 2) is present in the input voice, the detection signal waiting for the signal "Hello" is as shown in FIG. Stand up at the part. FIG.
In (c), a numerical value such as 0.9 or 0.8 is a numerical value indicating a certainty (degree of approximation), and a high numerical value such as 0.9 or 0.8 indicates that the word exists in the input voice with high certainty. be able to. That is, the registered word "good morning" is, as shown in FIG.
The registered word "konnichiwa", which exists in the w1 portion on the time axis of the input speech signal with a certainty of 0.9, has a w2 on the time axis of the input speech signal as shown in FIG.
It can be seen that there is a certainty of 0.8 in the part.

It is possible to determine which speaker class the user's voice belongs to by using the numerical value (approximation degree) indicating the certainty. This will be described with reference to the drawings.

FIG. 5 is a diagram showing the overall structure, and the same parts as those in FIG. 1 are designated by the same reference numerals. 5 is different from FIG. 1 in that the determination of the speaker class is output from the DRNN, that is,
The point is to obtain a numerical value (approximation degree) indicating the above-mentioned certainty and to judge which speaker class the input voice belongs to based on this approximation degree. Therefore, the speaker adaptation unit 2
Is a DRN as means for outputting the degree of approximation.
The N-type word detection unit 25 and a voice model data storage unit 26a (hereinafter, referred to as a first DRN) in which DRNN voice model data created based on voices of an unspecified number of men are stored.
N voice model storage unit 26a), a voice model data storage unit 26b (hereinafter referred to as a second D) that stores DRNN voice model data created based on voices of an unspecified number of women.
The RNN voice model storage unit 26b) is provided.
In this case as well, in order to simplify the explanation,
An example in which the speaker class is divided into male and female will be described.

In such a configuration, the system side gives an instruction, for example, "Please speak good morning", the user speaks "Good morning", and then,
In response to the instruction "Please speak hello",
When the user speaks based on an instruction from the system side such as "Hello", the voice is voice-analyzed, output as a feature vector sequence, and then stored in the input data storage unit 21.

The word detecting section 25 detects with what certainty the corresponding word exists in the input voice. That is, when the word "good morning" is present in the input voice, the detection signal waiting for the signal "good morning" is, as described above, as shown in FIG. 4B, "good morning" of the input voice. Stand up at the part. Similarly, when the word "hello" is present in the input voice, the detection signal waiting for the signal "hello" rises at the "hello" portion of the input voice as shown in FIG. .

Here, the first DR is applied to the input voice.
A detection signal is obtained using the NN voice model storage unit 26a and the second DRNN voice model storage unit 26b. For example, when the first DRNN voice model storage unit 26a is used for the user's voice "Ohayo", the detection signal waiting for "Ohayo" rises as shown in FIG. When the voice model storage unit 26b of FIG. 6 is used, the detection signal waiting for the “good morning” is shown in FIG.
It is assumed that the person stands up as shown in (b). Also, when the first voice model storage unit 26a is used for the user's voice "Hello", the detection signal waiting for the "Hello" is raised as shown in FIG.
6D, it is assumed that the detection signal waiting for "Hello" has risen as shown in FIG. 6D.

The rising portion (D
RNN output), a threshold th is set, and this threshold t
The area of a portion larger than h (hatched portion in the figure) is calculated, and based on the size of the area, the speaker class determination unit 22
Determines which speaker class the user's voice belongs to.

Now, the words for speaker adaptation are the words w1 and w.
There are three, 2 and w3. And the first for word w1
The area of the rising portion (DRNN output) of the detection signal which is equal to or larger than the threshold value when using the voice model storage unit 26a of No.
The area of the DRNN output that is greater than or equal to the threshold value when a is used is Sa2, and the first voice model storage unit 26 for the word w3
If the area of the DRNN output that is equal to or more than the threshold value when a is used is Sa3, the first DRNN voice model storage unit 26
DRN in words w1, w2, w3 when a is used
The total area Sa of the N output or more is equal to or more than the threshold value is obtained by: Sa = Sa1 + Sa2 + Sa3 (3)

Further, when the second DRNN voice model storage unit 26b for the word w1 is used, an area equal to or larger than the threshold of the DRNN output is Sb1, and the second DNN for the word w2.
DRNN using the RNN voice model storage unit 26b
Let Sb2 be the area above the output threshold and Sb3 the area above the DRNN output threshold when the second DRNN voice model storage unit 26b for the word w3 is used.
The total area Sb of the words w1, w2, and w3 that is equal to or larger than the threshold value of the DRNN output when the second DRNN voice model storage unit 26b is used is obtained by Sb = Sb1 + Sb2 + Sb3 (4).

The sizes of the areas Sa and Sb thus obtained are compared to determine the speaker class. That is, Sa
If <Sb, it is determined that the input voice belongs to a female speaker class, and as a result, a codebook created based on a large number of female voices (second unspecified speaker codebook). ) Select 23b.

In this way, the speaker class determination processing unit 2
According to 2, it is determined which speaker class the voice of the input speaker belongs to, and when the speaker class is determined, an unspecified speaker codebook created based on the voice of the speaker class is selected. To be done. As a result, when recognizing the voice of the input speaker, the voice of the input speaker is vector-quantized using a codebook created based on the voice close to the voice of the input speaker, and output as a feature code vector. Therefore, the voice recognition unit 3 can perform highly accurate recognition processing.

The voice recognition unit 3 performs voice recognition using a voice model corresponding to the selected speaker class. That is, as described above, in this case, since the input speaker is determined to belong to the female speaker class, the second voice model storage unit 31b is used to obtain the degree of approximation from the word detection unit 32 and The voice recognition processing unit 33 performs voice recognition based on the degree of approximation. In this case, if the DRNN method is used as the recognition means of the voice recognition unit 3, the first and second voice model storage units 26a and 26b (first and second DRNN voice model storage unit 31a) shown in FIG. , 31b) and the word detection unit 25 (word detection unit 32) can be shared by the voice recognition unit 3 and the speaker adaptation unit 2.

As described above, in the speaker adaptation process, it is determined whether the voice of the input speaker belongs to the male speaker class or the female speaker class using the DRNN output, and at the time of voice recognition, Speech recognition is performed using a speech model based on the result of the classification. In this case, the voice of the input speaker is determined to belong to the female speaker class, and is converted to a feature code vector using an unspecified speaker codebook created using the sounds of unspecified many female speakers. It is output, and the voice recognition unit 3 performs voice recognition using a voice model created using the voices of many female speakers.
Highly accurate voice recognition can be performed.

In the above description, the speaker class is divided into two speaker classes, that is, a male speaker and a female speaker. However, if the speaker class is further subdivided, Highly accurate recognition is possible.

Further, when the speaker class is determined, both the DP distance described first and the DRNN output described next are used, so that more accurate speaker class selection can be performed. In this case, the sum Da of the DP distances obtained by the above equation (1) is added by the sum Sa of the areas of the DRNN output which is equal to or more than the threshold value obtained by the equation (3), and D obtained by the equation (2) is added.
The sum Db of the P distances is added to the sum Sb of the areas of the DRNN output obtained by the equation (4) or more, and the speaker class is determined based on the value. However, for DP sums Da and Db, the one with the smaller numerical value is adopted, and DRNN
The sum of the areas equal to or larger than the output threshold value Sa, Sb has a larger numerical value. Therefore, when the DP distance is mainly considered, Sa and Sb take the reciprocals thereof and give appropriate weights to Sa and Sb, respectively. And add to Da and Db.

By performing such processing, it becomes possible to determine the speaker class with higher accuracy for the voice of the input speaker, and it is possible to perform subsequent voice recognition with even higher accuracy.

In the example shown in FIG. 1, when performing voice recognition, the voice feature vector sequence of the input speaker is coded by the vector quantization unit 20 and then passed to the voice recognition unit 3. In the case of the first embodiment, it is not always necessary to quantize, and the voice-analyzed feature vector sequence may be sent to the voice recognition unit 3 as it is. In this case, the unspecified speaker codebooks 23a and 23b corresponding to the speaker class are unnecessary, and when the speaker class is determined, a voice model for voice recognition corresponding to the determined speaker class is selected, At the time of recognition, voice recognition is performed using the selected voice model. In other words, it is determined which speaker class the input speaker belongs to, based on the determination result, a voice model belonging to the speaker class is selected, and recognition processing is performed using the selected voice model. is there. Even in this case, the voice recognition rate can be greatly improved.

(Second Embodiment) In the first embodiment described above, the speaker class is determined for the voice of the input speaker, and the codebook corresponding to the determination result (selected Select an unspecified speaker codebook created from an unspecified speaker belonging to the speaker class, and code the input speaker's voice using the selected unspecified speaker codebook during voice recognition. However, in the second embodiment, the speaker class is transmitted to the voice recognition unit, and the voice recognition unit performs the voice recognition using the voice model corresponding to the selected speaker class. The input speaker codebook is created using the unspecified speaker codebook selected based on the judgment result of 1., and at the time of voice recognition, the voice is recognized after the speaker adaptation is performed using this input speaker codebook. To do . This will be described in detail below.

FIG. 7 illustrates the overall structure of the second embodiment, and the same parts as those in FIG. 1 are designated by the same reference numerals. In the second embodiment, the case of speaker class determination using the DP distance will be described.
The same can be applied when the RNN output is used.

In FIG. 7, reference numeral 27 is an input speaker conversion codebook. This input speaker codebook 27 is an unspecified speaker codebook (in this case, the first speaker codebook selected based on the speaker class determination result). , Second unspecified speaker codebooks 23a, 23b). Here, a case will be described in which the second unspecified speaker codebook 23b is selected by the speaker class determination processing described in the first embodiment.

When the speaker adaptation is performed, the system side uses, for example, "Ohayo" as the speaker adaptation word.
When the user utters “Good morning” according to the instruction, the voice is input through the microphone 11, A / D converted by the A / D conversion unit 12, and then the voice. The analysis unit 13 outputs a voice feature vector sequence representing the feature of the frequency. Then, the voice-analyzed feature vector is temporarily stored in the input data storage unit 21. Similarly, as the next word for speaker adaptation, for example, an instruction to speak "Hello" is given, and when the user says "Hello" according to the instruction, the voice is analyzed by voice. The obtained voice feature vector sequence is output, and the feature vector sequence is stored in the input data storage unit 21. In this way, some feature vector strings of words for speaker adaptation are stored.

The input speaker codebook 27 is created by selecting the selected first and second unspecified speaker codebooks 23a and 23a.
3b (in this case, the second unspecified speaker codebook 23b), the first and second centroid vector sequence storage units 24.
This is performed using the data of either a or 24b (in this case, the second center-of-gravity vector sequence storage unit 24b) and the input data storage unit 21. Hereinafter, this process will be described with reference to FIG.

FIG. 8 shows the selected second unspecified speaker codebook 23b, which has a size of 256 and is composed of 256 code vectors indicated by white circles. Then, these code vectors are set to Ck1 and Ck.
2, Ck3, ..., Ck256, and actually 2
Although it is composed of 56 code vectors, only the code vectors Ck1, Ck2, ..., Ck9 are shown in FIG. This code vector is, for example,
200 words for each word, 200 for each word
If there are about 25 feature vectors obtained when an unspecified woman such as a person speaks, that is, about 25 feature vectors per word, about 1 million feature vectors can be obtained. Vector quantize
It is a collection of six representative code vectors.

For such a second unspecified speaker codebook 23b, for example, the center of gravity vector sequence for "Ohayo" obtained as described above (here, indicated by a black circle in the figure, the center of gravity vector Cs1). , Cs2, ...
., Cs7) is vector-quantized. That is, when the center of gravity vector sequence for “Ohayo” is vector-quantized with the code vectors of Ck1, Ck2, ..., Ck256, the first and second center of gravity vector Cs1 and Cs2 of the center of gravity vector sequence are associated with the code vector Ck1. And the third center of gravity vector Cs3
Is associated with the code vector Ck3, the fourth center of gravity vector Cs4 is associated with the code vector Ck4, and the fifth, sixth, and seventh center of gravity vectors Cs5, C
s6 and Cs7 are associated with the code vector Ck5, respectively, so that the center of gravity vector sequence of "good morning" is Ck1, Ck1, Ck3, Ck4, Ck5, Ck.
It will be replaced with a code vector sequence of 5, Ck5.

Next, the feature vector sequence of "Ohayo" from the user stored in the input data storage unit 21 is compared with the quantized "Ohayo" centroid vector sequence (called centroid code vector sequence). Correspondence is performed by DP matching.

This is shown in FIG. In FIG. 9, the content of the second unspecified speaker codebook 23b shows only the center of gravity code vectors Ck1, Ck3, Ck4, Ck5 of “Ohayo” and other code vectors for the sake of easy understanding. Are not shown.

When "Ohayo" is input from the user, the feature vector sequence of "Ohayo" (input speaker feature vector sequence) and the center of gravity code vector sequence Ck are input.
1, Ck1, Ck3, Ck4, Ck5, Ck5, Ck5
And are matched by DP matching. Feature vectors Ci1 and Ci of the input speaker feature vector sequence
If the positions 2, Ci3, Ci4, Ci5, Ci5 and Ci6 are as shown in FIG. 9, the center of gravity code vector sequence Ck
1, Ck1, Ck3, Ck4, Ck5, Ck5, Ck5
In this case, the input speaker feature vectors Ci1 and Ci2 are respectively associated with the centroid code vector Ck1, and the input speaker feature vector Ci3 is associated with the centroid code vector Ck3. The vectors Ci4 and Ci5 are respectively associated with the centroid code vector Ck4, and the input speaker feature vector Ci6
Is associated with the centroid code vector Ck5.

When the input speaker feature vector sequence is associated with the centroid code vector sequence in this way, then the difference vector between the associated vectors (input speaker feature vector-centroid) Code vector). In this case, since the input speaker feature vectors Ci1 and Ci2 are associated with Ck1, respectively, the difference vector V1 is obtained by averaging the input speaker feature vectors Ci1 and Ci2, and V1 = (Ci1 + Ci2) / 2−Ck1 Similarly, the input speaker feature vector Ci3 is C
Since it is associated with k3, the difference vector V3
Is calculated by V3 = Ci3-Ck3, and similarly, input speaker feature vectors Ci4, C
Since i5 is respectively associated with Ck4, the difference vector V4 is the input speaker feature vector Ci4, Ci5.
V4 = (Ci4 + Ci5) / 2−Ck4, and the input speaker feature vector Ci6 is C
Since it is associated with k5, the difference vector V5 is obtained by V5 = Ci6-Ck5. That is, the respective barycentric code vectors Ck1, Ck3, Ck4, Ck5 of the barycentric code vector sequence are
That is, it has the difference vectors of V1, V3, V4, and V5 obtained as described above with respect to the input speaker feature vector sequence.

In this way, the difference vectors V1, V
When 3, V4 and V5 are obtained, the difference vector is then used to obtain a code vector for "good morning" of the input speaker, and the code vector is mapped to the input speaker codebook 27.

Here, when the code vector to be obtained is represented by Ctx (where x is the code vector number, and the numerical values are 1, 3, 4, and 5 here), Ct1 = Ck1 + V1 Ct3 = Ck3 + V3 Ct4 = Ck4 + V4 Ct5 = Ck5 + V5.

These Ct1, Ct3, Ct4 and Ct5
Is the center of gravity code vector Ck1, Ck3, Ck of “Ohayo” in the second unspecified speaker codebook 23b.
4, Ck5 and the feature vector sequence of the input speaker are associated with each other, and their difference vectors V1, V3, V4, V5
Is a code vector obtained by adding to the barycentric code vectors Ck1, Ck3, Ck4, Ck5 of the unspecified speaker codebook 23b, and the code of the second unspecified speaker codebook 23b as shown in FIG. The vector is converted into the code vector of the input speaker codebook 27 by the difference vector.

However, in this case, 1 is called "Good morning".
Since we are only thinking about one speaker adaptation word, 4
Although only two code vectors Ck1, Ck3, Ck4, and Ck5 have been obtained as converted code vectors, the same process is performed on the other words for speaker adaptation, so that the input speaker code vector corresponding thereto is obtained. Is created.

In this way, the code vector of the second unspecified speaker codebook 23b is converted into the code vector of the input speaker codebook 27 to create the input speaker codebook. Specific speaker codebook 23
For example, if there are 256 code vectors in b, not all are converted, and many code vectors that are not converted (called unlearned code vectors) exist. The processing for converting the unlearned code vector (this is referred to as interpolation processing) will be described below.

Here, in order to simplify the explanation, it is assumed that only one word "Ohayo" for speaker adaptation is considered, and four centroid code vectors Ck1, Ck3, Ck4 for the word "Ohayo". , Ck5
Is converted as a code vector to the input speaker codebook 27, and the other code vectors to be converted (unlearned code vectors) are Ck2 and Ck as shown in FIG.
6, Ck7, Ck8, and Ck9.

This unlearned code vector Ck2, Ck
Of 6, Ck7, Ck8 and Ck9, the interpolation process for converting Ck2 into the input speaker codebook 23 will now be described with reference to FIG.

In FIG. 11, the unlearned code vector C
From the learned code vectors existing around k2, three code vectors are selected. In this case, there are four learned code vectors Ck1, Ck3, Ck4, and Ck5 around the unlearned code vector Ck2. Of these, code vectors Ck1, Ck4,
Assuming that three Ck5 are learned code vectors existing at a distance close to Ck2, three learned code vectors at these close distances are selected, and the difference corresponding to these code vectors Ck1, Ck4, Ck5 is selected. The difference vector V2 for the unlearned code vector Ck2 is determined using the vectors V1, V4, and V5. This V2
Is calculated by V2 = μ21 · V1 + μ24 · V4 + μ25 · V5. In this equation, μ21, μ24, μ2
5 is a coefficient representing a weight, μ21 is a weight according to the distance between Ck2 and Ck1, μ24 is a weight according to the distance between Ck2 and Ck4, and μ25 is a weight according to the distance between Ck2 and Ck5. The magnitude of the weight is set according to each distance, and is set to be μ21 + μ24 + μ25 = 1. In this way, the difference vector for Ck2 is determined, and by using the difference vector V2, the unlearned code vector Ck2 is converted into the code vector of the input speaker codebook 27 by Ct2 = Ck2 + V2.

Similarly, the respective difference vectors of the unlearned code vectors Ck6, Ck7, Ck8 and Ck9 other than Ck2 are obtained and converted using the respective difference vectors.

As a method for obtaining the weighting coefficient μ, there is a method using a fuzzy series, a method using a distance proportional to the reciprocal of the squared distance, or the like. As an example of a method using a distance proportional to the reciprocal of the squared distance, for example, if μ21 is taken, μ21 = {1 / (d21 × d21)} / {1 / (d21
× d21) + 1 / (d24 × d24) + 1 / (d25 ×
d25)}. Here, d21 is the code vector Ck2
, Ck1 and d24 are code vectors Ck2 and Ck
distance from k4, d25 is code vector Ck2 and Ck5
It shows the distance to. In the above equation, the distance of 2
Multiply 1 {1 / (d21 × d21)} by {1 / (d2
1 × d21) + 1 / (d24 × d24) + 1 / (d25
Xd25)} is divided by μ21 + μ as described above.
This is for obtaining μ21 which is 24 + μ25 = 1. Similarly, μ24 and μ25 are also obtained. Such 2
Since the method of obtaining the weighting coefficient by the distance proportional to the reciprocal of the squared distance requires a small amount of calculation and is simple, it is often advantageous to use this.

The input speaker codebook 27 can be created by the above processing.

In this way, an input speaker codebook is created using the unspecified speaker codebook corresponding to the selected speaker class, and at the time of voice recognition, this input speaker's voice is input. By performing voice recognition after encoding using the speaker codebook, it is possible to perform more highly accurate voice recognition.

By the way, in the above explanation, when the speaker is adapted, the system side gives an instruction such as "Speak good morning" or "Hello" and follow the instructions. When the DP matching processing is performed, since the voice feature data obtained by the user uttering “Good morning”, “Hello”, etc. is used.
The beginning and end of the voice feature data (feature vector sequence) for each word were known, and DP matching processing could be performed based on the data at the beginning and end, but the speaker uttered from the contents uttered by the user. It is also possible to detect a word for adaptation and automatically adapt to the speaker. Hereinafter, this will be described.

As described above, according to the present invention, as a means for cutting out the voice section of the word for speaker adaptation from the voice uttered by the speaker in the so-called "unsupervised" mode without any instruction from the system side, according to the present invention, The start / end free DP matching and the above-mentioned DRNN output are used.

In this start / end free DP matching, standard data of some word is compared with input voice data, a voice section for some word is cut out from the input voice data, and further the standard word data and the cut out voice are cut out. It inputs the DP matching distance to the section. By combining this start / end free DP matching output and the DRNN output, the words for speaker adaptation are automatically cut out from the contents uttered by the user. To do. That is, for example, when the word "Ohayo" is present in the content uttered by the user, the voice section for that word can be cut out by the start-end free DP matching, and similarly, "Ohayo" can be extracted. If "Good morning" is present, a DRNN output having a high degree of approximation is output, and from both of these outputs, it is highly possible that the word "Good morning" appears to exist at a certain time in the utterance content of the user. You can know by probability.

By using such means, while the user speaks to the system, a speaker adaptation word is cut out from the utterance contents, and the cut out speaker adaptation word is extracted. The words are stored in the input data storage unit 21, and the speaker adaptation as described above is automatically performed. Therefore, there is an effect that the recognition performance naturally increases during use.

In the first and second embodiments described above, the characteristic data of several words for speaker adaptation spoken by the user are temporarily stored in the input data storage unit 21, and then various data are stored. I am trying to process it. A RAM is used as the input data storage unit 21, but naturally, the capacity of the RAM is also limited in order to make the entire apparatus as small as possible. To cope with this, the input data may be stored in the input data storage unit 21 as quantized and coded data. In particular, in the case without the above-mentioned teacher, in order to reduce the amount of data, it is necessary to store a large number of word data.
It is advantageous to code it.

(Third Embodiment) In the third embodiment, speakers are classified into classes, and an unspecified speaker codebook for each class (the first and second embodiments). , The first unspecified speaker codebook 23a, the second
(Corresponding to the unspecified speaker codebook 23b). In the first and second embodiments, in order to make the description easier to understand, the speaker classes are specifically divided into male and female for convenience, but the speaker classes are actually As described below, the contents of the codebook created by an unspecified number of speakers are classified into some groups based on the voice feature data, and not classified as male / female. Each speaker class is defined as a speaker class, and as a result of the classification, the speaker class is divided into a male speaker, a female speaker, and even a child class. Therefore, for example, when the class is divided into two speaker classes, the class is divided into a female speaker class and a male speaker class. Then, which speaker class the input voice belongs to is determined, and an unspecified speaker codebook corresponding to the determination result is selected. The classification will be described below.

FIG. 12 is a feature vector sequence obtained when three people A, B, and C utter the words "good morning", "hello", and "just now", for example. A2 is a feature vector sequence for "Hello" of a speaker of A, A2 is a feature vector sequence for "Hello" of a speaker of A, A3 is a sequence of feature vectors for "Imaima" of a speaker of A, B1 is "Good morning" of a speaker of B , B2 is a feature vector sequence for "Hello" of the B speaker, B3 is a feature vector sequence for "Ima" of the B speaker, C1 is a feature vector sequence for "Ohayo" of the C speaker, C2 represents a feature vector sequence for "Hello" by the speaker of C, and C3 represents a feature vector sequence for "just now" of the speaker of C.

FIG. 12 shows feature vectors obtained when three speakers utter three words, but in reality, a large number of unspecified speakers utter some words. There are a large number of feature vector sequences for each word obtained in this way.

With respect to a feature vector sequence for each word obtained by such an unspecified number of speakers uttering every several words, the features obtained by each speaker uttering a certain word The DP distance of the vector sequence is obtained and classification is performed.

Here, the DP distance of the feature vector sequence obtained by each speaker uttering a certain word will be described with reference to FIG.

FIG. 13A shows a case where DP matching is performed with the feature vector sequence B1 of the speaker B "Ohayo" with the feature vector sequence A1 when the speaker A utters "Ohayo" as a reference. 5B shows the correspondence relationship between the feature vector comrades at each time, and FIG. 7B shows the feature vector sequence B when the speaker B utters “Ohayo”.
1 shows the correspondence relationship between the feature vector comrades at each time when DP matching is performed with the feature vector sequence A1 of "Ohayo" of the speaker A on the basis of 1.

As described above, the DP distance varies depending on whether the speaker A or the speaker B is used as a reference. Therefore, the DP distance normalized by the length when the speaker A is used as a reference is used. , The DP distance normalized by the length when the speaker B is used as a reference is calculated, and the average value thereof is calculated and used as the inter-speaker distance between the speaker A and the speaker B for a certain word. . That is, FIG.
In the example of 3, the DP distance d1 (A1, B1) between the speaker A and the speaker B for the word "good morning" is d1 (A1, B1) = {(dA1 + dA2 +, ...,
+ DB8) / 8 + dB1 + dB2 +, ..., + dB
7) / 7} / 2. In this formula, dA1, dA2, ...
, DA8 represents a distance corresponding to each time (time t1 to t8 and time t1 to time t7) with the feature vector sequence of the speaker B based on the feature vector sequence of the speaker A, and dB1, dB2, ..., dB8 represent distances corresponding to respective times (time t1 to t7 and time t1 to time t8) of the feature vector sequence of the speaker A based on the feature vector sequence of the speaker B. ing.

The above is the DP distance between the speakers A and B for the word "good morning". Similarly, the speakers A and B are
Find the DP distance for the other words of.

Now, assuming that the number of words is n, the sum dAB of DP distances in the words of speaker A and speaker B is: dAB = d1 (A1, B1) + d1 (A2, B2) +,
..., + d1 (An, Bn), and the sum dAB of DP distances in each word is called the inter-speaker distance between speaker A and speaker B. Where d1 (A
1, B1) is the DP distance between the speakers A and B with respect to the word "good morning" as described above, and d1 (A2,
B2) and d1 (An, Bn) represent the DP distances of the speaker A and the speaker B for each word.

In this way, the inter-speaker distances of all speakers are obtained, and based on the obtained inter-speaker distances of the speakers,
Speaker classification is performed by the process shown in the flowchart of FIG. The process will be described below with reference to this flowchart.

Here, a case will be described in which the class is divided into four speaker classes (the number of speaker classes N = 4). First, with N = 1 (step s1), in the class of N = 1 (where N = 1 is the entire range before the classification), the speakers of the same kind obtained as described above. Among the inter-speaker distances of, the two speakers with the longest inter-speaker distance are selected (step s2). FIG. 15 shows each speaker as a point based on the inter-speaker distance for convenience. In this case, it is assumed that the speakers A and Z having the longest inter-speaker distance are selected. For convenience of explanation, FIG. 14 shows each speaker by a dot, but originally, for example, an unspecified speaker such as several hundreds of speakers was obtained by uttering several words. There are many feature vector sequences for each word.

When the two speakers A and Z having the longest inter-speaker distance are selected in this way, next, in step s3 of FIG. 14, the speakers A and Z are selected. Then, the DP distance of the representative vector sequence of the other speakers B, C, D, ... Is calculated, and based on the calculated distance, it is divided into two as shown by the broken line L1 in FIG. CL21 is formed (first trial classification). Next, in each class, a centroid vector sequence is obtained for each word (step s4). Two classes C now
L11 and CL21, so class CL11 and CL2
In step 1, a centroid vector sequence is obtained for each word. The method of obtaining the center-of-gravity vector sequence is as described above. For example, taking the word "good morning" as an example,
For each of the classes CL11 and CL21, the center of gravity vector sequence is obtained using the feature vector sequence “Ohayo” of the speaker belonging to each class.

Next, it is judged whether or not convergence has occurred (step s
5) is performed, but this is to see whether or not a certain fixed result is reached by repeating steps s3 and s4 several times, and there is no convergence in one processing, Steps s3 and s4 are repeated several times until the result of is reached. However, step s3 in the second processing
Uses the center of gravity vector sequence obtained in the previous step s4,
In each class, the DP distance between the center-of-gravity vector sequence for each word and the feature vector sequences of all speakers existing in the class is calculated, and based on the DP distance, it is divided into two as shown by the broken line L2, and two Classes CL12 and CL22 are formed (second trial classification).

As described above, when the second trial classification is performed, the newly formed classes CL12 and CL2 are formed.
The speakers belonging to 2 are slightly different from those in the first time. Then, in each class, a centroid vector sequence is obtained for each word (step s4). Here, since the two classes CL12 and CL22 have been formed by the second trial classification, the class CL
12, CL22 obtains the centroid vector sequence for each word for each class.

Then, again, it is judged whether or not it has converged. The judgment as to whether or not it has converged is made as follows.

That is, assuming that the present is the n-th trial classification stage, for each class formed in this n-th classification, the centroid vector sequence of each word and its class are assigned. The sum of the DP distances with the feature vector sequence of each word of all the existing speakers is obtained. Then, the centroid vector sequence of each word for each class obtained in the previous classification (n-1st trial classification) and the feature vector sequence of each word of all speakers existing in that class Of D
The difference between the sums of the P distances is calculated, and it is determined that the difference has almost disappeared.

In this way, for example, if it is converged by two processes, the speaker class at the time of convergence becomes the speaker class, and in this case, the speaker classes CL12 and CL22 formed by the broken line L2. Two speaker classes will be created (the number of speaker classes N = 2).

Then, in step s6 of FIG.
The number of speaker classes set by the number of speaker classes (in this case, N =
4)), and if the set number of speaker classes is reached, the process ends, but in this case, since the set number of speaker classes is not reached, the process returns to step s2. .

The process in step s2 is a process of selecting two speakers having the longest inter-speaker distance for each class. In this case, the speaker classes CL12 and CL are used.
The process of selecting the two speakers having the longest inter-speaker distance is performed for each of the 22. Then, similarly to the above, steps s3 and s4 are performed, and the processing is continued until the convergence. As a result, the speaker classes CL12 and CL22
Will be divided into two and four speaker classes will be created.

The speaker class is created as described above. For example, the initially created speaker class CL12,
If the CL 22 sees the divided result, for example, CL
Speaker classes such as 12 having voice feature data close to that of a male speaker and CL22 class having voice feature data close to that of a female speaker, each speaker class being further divided into two The class means that each class becomes a speaker class that is further subdivided based on the voice feature data.

In the first and second embodiments, it is used for the unspecified speaker codebook 23a and voice recognition based on the voice data of the speakers belonging to the speaker class CL12 thus created. Create a voice model and use speaker class C
An unspecified speaker codebook 23b and a voice model used for voice recognition are created based on the voice data belonging to L22.

In the above explanation, the number of speaker classes is
The number is 2 (N = 2), 2 2 (N = 4), 2 3 (n = 8), and so on, but belongs to a speaker class created by division. If there is a large difference in the number of speakers between the speaker classes, the process of dividing only the class with the overwhelmingly large number of speakers will reduce the number of speaker classes to three.
It is also possible to divide into any number such as 5, 5, 6 and 7.

As described above, in the third embodiment, the speech feature data of a large number of unspecified speakers are used to classify into any number of speaker classes. On the other hand, regardless of the characteristic data of the voice, the voicebook of male speakers is artificially collected by artificially collecting voices of only males, and the voicebook of female speakers is collected by collecting voices of only females. Create a codebook, or collect voices of adult males only and codebooks of adult male speakers without specific speakers, collect voices of adult females only, codebooks of adult female speakers without specifics There is also a method of creating each codebook such as a codebook of unspecified speakers of children by collecting voices of only children, but collecting speakers of each hierarchy regardless of voice feature data and each code. When creating a book, some adult women and some women have voice data that is close to that of children, and some children have voice data that is close to that of adult men. Classification can not be. However, in the speaker classification according to the present invention, since a large number of unspecified speakers are classified based on their voice feature data, it is possible to perform highly accurate speaker classification based on the voice feature data.

The program for performing the processing of the present invention described above can be stored in a storage medium such as a floppy disk, and the present invention also includes this storage medium.

[0140]

As described above, according to the present invention,
Determine which speaker class the voice of the input speaker belongs to,
Based on the determination result, a voice model corresponding to the determined speaker class is selected, and at the time of voice recognition, the voice of the input speaker is recognized by using the selected voice model. With a variety of processes and configurations, it is possible to perform voice recognition with a high recognition rate.

Further, according to the present invention, an input speaker codebook is created using the unspecified speaker codebook corresponding to the selected speaker class, and the speaker adaptation is performed using this input speaker codebook. As a result, more reliable speaker adaptation is possible, and in the voice recognition unit, the speaker-adapted code vector by the input speaker codebook is associated with the selected speaker class. By performing the voice recognition process using the voice model, the voice recognition with a higher recognition rate becomes possible.

Then, by determining the speaker class to which the voice of the input speaker belongs, using the DP distance or the DRNN output, and combining both of them, it becomes possible to make an accurate class determination. By using the DRNN output, it is possible to cut out a speaker adaptation word included in the input voice by a keyword spotting process, and a simple process for determining to which speaker class the voice of the input speaker belongs. Can be performed with high precision.

Further, the processing for classifying speakers according to the present invention is carried out for a feature vector string for each of a plurality of words in each speaker obtained from an unspecified number of speakers.
The DP matching distance is calculated for each word between the speakers, the sum of the distances is set as the distance between the speakers, and the inter-speaker distances are calculated among the unspecified number of speakers. Since classification is performed based on the inter-speaker distance, it is possible to perform highly accurate speaker classification based on voice feature data, and unspecified talk created from speakers belonging to such classified speaker classes. By adapting the speaker using the speaker codebook, accurate speaker adaptation is possible, and by recognizing the voice using the voice model created corresponding to the classified speaker class, it becomes high. A recognition rate is obtained.

[Brief description of drawings]

FIG. 1 is a configuration diagram illustrating a first embodiment of the present invention.

FIG. 2 is a diagram for explaining an example of obtaining a center of gravity vector sequence from a feature vector sequence for a certain word by some unspecified speakers.

FIG. 3 is a diagram illustrating a process of normalizing the temporal length of a feature vector sequence when obtaining a center-of-gravity vector.

FIG. 4 is a diagram illustrating a word detection process according to the DRNN method.

FIG. 5 is a configuration diagram when a DRNN output is used for speaker class determination in the first embodiment.

FIG. 6 is a diagram for explaining speaker class determination based on DRNN output.

FIG. 7 is a configuration diagram illustrating a second embodiment of the present invention.

FIG. 8 is a diagram illustrating a process of associating a code vector in a selected unspecified speaker codebook with a centroid vector and quantizing the centroid vector in the second embodiment.

FIG. 9 is a diagram for explaining correspondence between a quantized centroid vector (centroid code vector) and an input speaker feature vector.

FIG. 10 is a diagram illustrating a process of converting a learned code vector of an unspecified speaker codebook into an input speaker codebook using a difference vector.

FIG. 11 is a diagram for explaining interpolation processing of an unlearned code vector.

FIG. 12 is a diagram showing a feature vector sequence obtained when some speakers speak some words for explaining the third embodiment.

FIG. 13 is an explanatory diagram for obtaining a DP distance between speakers for a certain word in order to obtain a distance between speakers.

FIG. 14 is a flowchart illustrating processing according to the third embodiment.

FIG. 15 is a diagram illustrating classification processing.

FIG. 16 is a diagram illustrating a general vector quantization process that has been conventionally performed.

[Explanation of symbols]

1 voice input unit 2 speaker adaptation unit 3 voice recognition unit 11 microphone 12 A / D conversion unit 13 voice analysis unit 20 vector quantization unit 21 input data storage unit 22 speaker class determination processing unit 23a first unspecified talk Person codebook 23b second unspecified speaker codebook 24a first centroid vector storage section 24b second centroid vector storage section 25 word detection section according to the DRNN method 26a first DRNN speech model storage section 26b second DRNN Speech model storage unit 27 Input speaker codebook 31a First speech model storage unit 31b Second speech model storage unit 32 Word detection unit 33 Speech recognition processing unit Ca1, Ca2, ... Speech feature vector of speaker A Cb1, Cb2, ... Speech feature vector of speaker B Cc1, Cc2, ... Speech feature vector of speaker C Cd1, Cd2, ... D speech feature vector of speaker Cs1, Cs2, ... Centroid vector Ck1, Ck2, ... Code vector of unspecified speaker codebook Ci1, Ci2, ... Input speaker Voice feature vector V1, V2, ... Difference vector Ct1, Ct2, ... Code vector converted using difference vector CL12, CL22 Classified speaker class

Claims (12)

[Claims] 1. A voice feature vector sequence of a speaker belonging to each of the plurality of speaker classes classified based on the voice feature data obtained from an unspecified number of speakers is created. It is determined which speaker class the voice of the input speaker belongs to based on the result obtained by comparing the data with the voice feature data of the input speaker, and the determined speaker is determined based on the determination result. A speaker adaptation method, characterized in that a voice model for voice recognition corresponding to a class is selected, and at the time of voice recognition, the voice of an input speaker is subjected to voice recognition processing using the selected voice model. 2. A plurality of speaker classes classified based on voice feature data obtained from an unspecified number of speakers,
Create an unspecified speaker codebook based on the voice feature data of the speakers belonging to each speaker class, and input the data created based on the voice feature vector sequence of the speakers belonging to each speaker class. Based on the result obtained by comparing with the voice feature data of the speaker, it is determined which speaker class the voice of the input speaker belongs to, and based on the determination result, it corresponds to the determined speaker class. A voice model for speech recognition corresponding to the unspecified speaker codebook and its speaker class is selected, an input speaker codebook is created based on the selected unspecified speaker codebook, and speech recognition is performed. Sometimes the voice of the input speaker is vector-quantized using the created input speaker codebook and the selected unspecified speaker codebook, and then passed to the voice recognition unit, where the voice recognition unit selects the selected voice. Speaker adaptation method, characterized by recognition processing using the speech model. 3. The voice of the input speaker is obtained from the result obtained by comparing the data produced based on the voice feature vector sequence of the speaker belonging to each speaker class with the voice feature data of the input speaker. The process of determining which speaker class
Obtaining the centroid vector sequence for each word for speaker adaptation from the speech feature vector sequence of the speaker belonging to each speaker class, the centroid vector sequence of each word and the input speaker's speech for each speaker class 3. The speaker adaptation method according to claim 1, wherein the distance to the feature vector sequence is obtained by DP matching, and which speaker class the voice of the input speaker belongs to is determined based on the distance. 4. The voice of the input speaker is obtained from the result obtained by comparing the data produced based on the voice feature vector sequence of the speaker belonging to each speaker class with the voice feature data of the input speaker. The process of determining which speaker class
Dynamic recurrent neural network (D) based on the speech feature vector sequence of speakers belonging to each speaker class.
ynamic Recurrent Neural Networks: DRN
A voice model corresponding to each speaker class is created by a method called N), and a numerical value indicating the certainty of the existence of a predetermined word is output from the voice feature vector sequence of the input speaker and the voice model. The speaker adaptation method according to claim 1 or 2, wherein the speaker class to which the voice of the input speaker belongs is determined based on a numerical value indicating. 5. The voice of the input speaker is obtained based on the result obtained by comparing the data produced based on the voice feature vector sequence of the speaker belonging to each speaker class with the voice feature data of the input speaker. The process of determining which speaker class
Obtaining the centroid vector sequence for each word for speaker adaptation from the speech feature vector sequence of the speaker belonging to each speaker class, the centroid vector sequence of each word and the input speaker's speech for each speaker class The distance to the feature vector sequence is obtained by DP matching, and a voice model corresponding to each speaker class by the DRNN method is created based on the voice feature vector sequence of the speaker belonging to each speaker class, and the input speaker From the voice feature vector sequence and the voice model, a numerical value indicating the certainty of the existence of a predetermined word is obtained, and the calculated DP matching distance and the numerical value indicating the certainty of the existence of the predetermined word are used as the basis. 3. The speaker adaptation method according to claim 1, further comprising determining which speaker class the voice of the input speaker belongs to. 6. The process of classifying into a plurality of speaker classes based on the voice feature data obtained from the unspecified number of speakers is performed by a plurality of speakers of each speaker obtained from the unspecified number of speakers. For each feature vector sequence for each word, the DP matching distance is obtained for each word between each speaker, and the sum of the distances is set as the distance between the speakers, and the distances between the unspecified number of speakers are respectively calculated. 6. The speaker adaptation method according to claim 1, wherein the inter-speaker distance is determined, and the classification is performed based on the inter-speaker distance. 7. A voice feature vector sequence of speakers belonging to each speaker class of a plurality of speaker classes classified based on voice feature data obtained from an unspecified number of speakers is created. Data storing means for storing the data, an input data storing means for storing the feature vector sequence of the voice of the input speaker, voice characteristic data for a certain word of the input speaker, and the feature data storing means stored in the feature data storing means. Based on the result obtained by comparing with the feature data for the word, it is determined which speaker class the voice of the input speaker belongs to, and based on the determination result, the voice recognition corresponding to the determined speaker class. And a speaker class determination processing unit for selecting a voice model for use in speech recognition. When performing voice recognition, the voice of the input speaker is vector-quantized using the selected voice model and then passed to the voice recognition unit. Speaker adaptation device characterized by: 8. A plurality of speaker classes classified based on voice feature data obtained from an unspecified number of speakers,
An unspecified speaker codebook created based on the voice feature data of speakers belonging to each speaker class, and data created based on the voice feature vector sequence of speakers belonging to each speaker class, Based on the result obtained by comparing with the voice feature data of the input speaker, it is determined which speaker class the input speaker's voice belongs to, and based on the determination result, it corresponds to the determined speaker class. The unspecified speaker codebook and the speaker class determination processing unit that selects a voice model for speech recognition corresponding to the speaker class, and the unspecified speaker codebook selected by this speaker class determination processing unit Based on the input speaker codebook created based on the above, the voice of the input speaker is recognized as a vector quantity using the created input speaker codebook and the selected unspecified speaker codebook during speech recognition. After ized, passed to the speech recognition unit, a speech recognition unit, speaker adaptation device, characterized in that the recognition process using the selected speech model. 9. The voice of the input speaker is obtained from the result obtained by comparing the data produced based on the voice feature vector sequence of the speaker belonging to each of the speaker classes with the voice feature data of the input speaker. The process of determining which speaker class
Obtaining the centroid vector sequence for each word for speaker adaptation from the speech feature vector sequence of the speaker belonging to each speaker class, the centroid vector sequence of each word and the input speaker's speech for each speaker class 9. The speaker adaptation apparatus according to claim 7, wherein a distance to the feature vector sequence is obtained by DP matching, and which speaker class the voice of the input speaker belongs to is determined based on the distance. 10. The voice of the input speaker is obtained based on the result obtained by comparing the data produced based on the voice feature vector sequence of the speaker belonging to each of the speaker classes with the voice feature data of the input speaker. The process of determining which speaker class belongs to is to create a voice model corresponding to each speaker class by the DRNN method based on the voice feature vector sequence of speakers belonging to each speaker class, and From the voice feature vector sequence and the voice model, a numeric value indicating the certainty of the existence of a predetermined word is output, and based on the numeric value indicating the certainty, which speaker class the voice of the input speaker belongs to The speaker adaptation device according to claim 7 or 8, characterized in that 11. The voice of the input speaker is obtained based on the result obtained by comparing the data produced based on the voice feature vector sequence of the speaker belonging to each speaker class with the voice feature data of the input speaker. The process for determining which speaker class belongs to is to obtain a centroid vector sequence for each word for speaker adaptation from the speech feature vector sequence of speakers belonging to each speaker class, and for each speaker class, The distance between the center of gravity vector sequence of each word and the feature vector sequence of the voice of the input speaker is obtained by DP matching, and each talk by the DRNN method is performed based on the voice feature vector sequence of the speaker belonging to each speaker class. Create a voice model corresponding to the person class, and from the voice feature vector sequence of the input speaker and the voice model, obtain a numerical value indicating the certainty of the existence of a predetermined word, and the obtained DP matching distance, Based on the numerical value indicating the serial certainty of the presence of a given word,
9. The speaker adaptation device according to claim 7, wherein it is determined which speaker class the voice of the input speaker belongs to. 12. The process of classifying into a plurality of speaker classes based on the voice feature data obtained from the unspecified number of speakers is performed by a plurality of speakers of each speaker obtained from an unspecified number of speakers. For the feature vector sequence for each word, the DP matching distance is calculated for each word between each speaker,
The sum of the distances is used as the distance between the speakers, the inter-speaker distances are obtained among the unspecified large numbers of speakers, and the classification is performed based on the inter-speaker distances. The speaker adaptation device according to any one of claims 7 to 11.