JPH10274992A - Method and device for forming voice model learning data - Google Patents

Abstract

PROBLEM TO BE SOLVED: To inexpensively form a voice model in a short time when the voice model related to a new word absent in a data base is formed. SOLUTION: The utterance data of at least a person of speaker among the utterance data obtained from many speakers beforehand having as a data base are made the standard speaker data, and others are made the learning speaker data (step s1), and a conversional function from a standard speaker data space to a learning speaker data space is formed based on the beforehand provided word data (step s2). When the learning data related to the new word are formed, the data obtained by that the standard speaker utters related to the new word are data converted to the learning speaker data space by using the conversional function, and the learning data related to the new word are formed (steps s3, s4).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a method and an apparatus for creating data for speech recognition learning for creating training data for learning a speech model for unspecified speaker speech recognition.

44 Standard speaker code data storage unit 45 Learning speaker code data storage unit 46 Code data conversion processing unit

[0003]

2. Description of the Related Art DRNN (Dynamic Recurrent Neural Network) is one of the speech recognition techniques for unspecified speakers.
s) There is a speech recognition technology using a speech model (this DR
The applicant of the present invention has already filed an application for speech recognition technology by NN in Japanese Patent Application Laid-Open Nos. 6-4079 and 6-119476.

[0004] This DRNN speech model is designed so that when a feature vector sequence of a word is input as time-series data, an appropriate output for the word is obtained.
The weight and bias between each unit are determined in accordance with a predetermined learning rule, whereby an output similar to the teacher output for the word is obtained for the voice data of the word spoken by a certain unspecified speaker. ing.

For example, "good morning" of a certain unspecified speaker
When the time series data of the feature vector sequence of the word “good morning” is input, in order to obtain an output close to the ideal output (teacher output) for the word “good morning”, the feature vector of the word “good morning” at each time Is given to the corresponding input unit, and is converted by the weight and bias set according to the learning rule. This is input as a time series data.
The time series processing is performed for all the feature vector strings of one word for each time. In this way, for speech data of a word spoken by an unspecified speaker, an output close to a teacher output for the word is obtained.

[0006] As described above, the learning rule for changing the weight of the DRNN speech model prepared for all the words to be recognized so that an appropriate output is obtained for each word is described in IEICE technical report: IEICI sp92-125 (1993-0
It is described on pages 17 to 24 of 1).

When creating a speech model for speech recognition of unspecified speakers, not limited to the DRNN speech model, an unspecified large number of speakers of about several hundreds are used for each word (for example, 2 words).
It is common to have learning data created from utterance data for each word obtained by uttering about 100 words) as a database, and perform learning based on the training data in the database to create a speech model. is there.

[0008]

However, there are cases where the user issues a request to create a speech model of a word that is not in the database. Conventionally, when creating a speech model for a word that does not exist in the database, about hundreds of speakers speak the word about the word and create learning data for the word based on the utterance data. Therefore, it was necessary to create a speech model based on the learning data.

As described above, when a speech model for a new word is created, in order to create training data for learning the speech model, utterance data of several hundred people is collected, and then the training data is created. Creation must be done. Therefore, there is a problem that it takes a lot of time to create a voice model and the creation cost increases.

Accordingly, the present invention makes it possible to create learning data for hundreds of people by using one or several utterance data selected as a standard speaker when creating a speech model of a word that does not exist in the database. Accordingly, it is an object of the present invention to realize a speech model learning data creating method and an apparatus thereof that can provide a speech model for a new word in a short period of time and at low cost.

[0011]

According to a first aspect of the present invention, there is provided a voice model learning data generating method for generating learning data for learning a voice model for voice recognition. In the creation method, the utterance data of at least one of the utterance data obtained from a large number of speakers having a database in advance is used as the standard speaker data, and the others are used as the learning speaker data. When creating a conversion function from the standard speaker data space to the learning speaker data space based on the data and creating learning data for a new word, the standard speaker speaks for the new word. The method is characterized in that the obtained data is converted into a learning speaker data space using the conversion function to generate learning data for a new word.

Thus, when a speech model is created for a new word that is not in the database, learning data for that word can be created based on utterance data generated by a small number of standard speakers. As described above, in order to create a speech model for a new word, there is no need to collect hundreds of utterance data to create learning data, and a speech model can be created in a short period of time and at low cost.

According to a second aspect of the present invention, in the first aspect of the present invention, the data existing in the standard speaker data space and the learning speaker data space are each word obtained by frequency analysis of a speech signal. The feature vector of each
Further, the process of converting the data obtained by the standard speaker uttering the new word into the learning speaker data space using the conversion function constitutes each word in the standard speaker data space. The feature is that it is performed using a difference vector between a feature vector and a feature vector constituting each word in the learning speaker data space.

According to a second aspect of the present invention, a feature vector (for example, data represented by an LPC cepstrum coefficient such as ten-dimensional) obtained by frequency-analyzing an audio signal is provided.
Because it uses itself, highly accurate data can be obtained,
Furthermore, since the data is converted from the standard speaker data space to the learning speaker data space using the difference vector obtained in advance,
Simple and high-precision data conversion can be performed.

According to a third aspect of the present invention, in the first aspect of the present invention, the data existing in the standard speaker data space and the learning speaker data space are each a word obtained by frequency analysis of a speech signal. Code data obtained by vector quantization of each feature vector, and data obtained by the standard speaker uttering the new word,
The process of converting to the learning speaker data space by using the conversion function is to perform vector quantization on new word data in the standard speaker data space to obtain code data, and map the code data to the learning speaker data space. Thus, data conversion from the standard speaker data space to the learning speaker data space is performed.

That is, according to the third aspect of the present invention, a feature vector obtained by frequency analysis of an audio signal is subjected to vector quantization for processing. Therefore, although the data is somewhat coarse, the processing is simplified and the processing time can be reduced.

According to a fourth aspect of the present invention, there is provided a voice model learning data generating apparatus for generating learning data for learning a voice model for voice recognition. A standard speaker data storage unit that stores, as standard speaker data, utterance data of at least one of utterance data obtained from a number of speakers having a database in advance, and utterance data other than the standard speaker. A pseudo-learning unit having a learning speaker data storage unit for storing as learning speaker data and a data conversion unit for performing data conversion from the standard speaker data space to the learning speaker data space using a conversion function obtained in advance; A word data generating unit; and a learning data effective unit for storing data generated by the pseudo learning word data generating unit. When creating training data, the data obtained by the standard speaker uttering the new word is converted into a learning speaker data space using the conversion function, and a pseudo-learning word for the new word is generated. Create data,
It is characterized in that learning data is created using the pseudo learning word data.

Thus, when creating a speech model for a new word that is not in the database, learning data for that word can be created based on the utterance data generated by a small number of standard speakers. As described above, in order to create a speech model for a new word, it is not necessary to collect hundreds of utterance data to create learning data, and a speech model can be created in a short time.

According to a fifth aspect of the present invention, in the invention of the fourth aspect, the standard speaker data stored in the standard speaker data storage section and the learning speaker data stored in the learning speaker data storage section are different from each other. Learning the feature vector for each word obtained by frequency analysis of the voice signal, and the data obtained by the standard speaker uttering the new word using the conversion function. The process of converting to the speaker data space is performed using a difference vector between a feature vector forming each word in the standard speaker data space and a feature vector forming each word in the learning speaker data space. And

According to the fifth aspect of the present invention, since the feature vector itself obtained by frequency analysis of the audio signal is used, highly accurate data can be obtained, and furthermore, using the difference vector obtained in advance. Since data is converted from the standard speaker data space to the learning speaker data space, simple and highly accurate data conversion can be performed. You.

According to a sixth aspect of the present invention, in the fourth aspect of the present invention, the standard speaker data stored in the standard speaker data storage section and the learning speaker data stored in the learning speaker data storage section are different from each other. Code data obtained by vector quantization of a feature vector for each word obtained by frequency analysis of a voice signal, and data obtained by uttering the standard speaker for the new word, The process of converting to the learning speaker data space by using the conversion function is to perform vector quantization on new word data in the standard speaker data space to obtain code data, and map the code data to the learning speaker data space. 5. The speech model training data creation device according to claim 4, wherein the data is converted from the standard speaker data space to the learning speaker data space.

That is, according to the invention of claim 6, a feature vector obtained by frequency analysis of an audio signal is subjected to vector quantization for processing. Therefore, although the data is somewhat coarse, the processing is simplified and the processing time can be shortened, and the memory for storing the standard speaker data and the learning speaker data requires only a small capacity.

[0023]

Embodiments of the present invention will be described below.

(First Embodiment) FIG. 1 is a flowchart schematically illustrating an embodiment of the present invention. FIG.
In the utterance data of about several hundred people (here, 300 people) already stored as a database,
Several arbitrarily selected persons (for example, five persons) are selected as standard speakers, utterance data of each standard speaker is used as standard speaker data, and others are used as learning speakers, and utterance data of each learning speaker are learned. It is set as speaker data (step s1).

Then, a conversion function from each of the standard speaker data spaces to the learning speaker data space is created by using word (about 200 words) data which is stored in a database in advance (step s2).

As described above, a conversion function from a certain standard speaker data space to each learning speaker data space is created in advance using all the words based on the word data held as a database. deep. Then, a conversion function from the standard speaker data space to each learning speaker data space is obtained for each learning speaker for each standard speaker.

When creating learning data for a new word that is not in the database, the standard words are uttered by the standard speakers, and the utterance data is converted by the conversion function. Then, data is converted into the respective learning speaker data spaces to generate pseudo-learning word data for several hundred persons (step s3).
Using the pseudo-learning word data thus created, new word learning data is created (step s).
4).

Thus, when creating a speech model for a new word that is not in the database, it is possible to create learning data for the word based on utterance data generated by several standard speakers. The above processing will be described in more detail.

FIG. 2 shows one speaker selected as a standard speaker and one speaker other than the standard speaker among the utterance data of several hundred speakers held as a database. This represents speech data (for example, a feature vector represented by a 10-dimensional LPC cepstrum coefficient) for some words of a (learning speaker). Actually, data for about 200 words exists. 3 shows feature vectors for three words.

FIG. 2 shows only one piece of standard speaker data 10 and one piece of learning speaker data 20 to simplify the explanation. The feature vectors of certain three words A, B, and C in the standard speaker data space and the feature vectors of the words A, B, and C in the learning speaker data space already correspond to each other. . Based on the respective word data in the standard speaker data space and the word data in the learning speaker data space as such a database, a data conversion function from the standard speaker data space to the learning speaker data space is determined in advance. Create it. Here, a difference vector is used as the data conversion function.

For example, the feature vector value Vk1 of the word A in the standard speaker data space and the feature vector value Vt1 of the word A in the learning speaker data space are represented by Vt1 = Vk1 + V1 if the difference vector is V1. . Also, the word A in the standard speaker data space
The feature vector value Vk2 of the word A and the feature vector value Vt2 of the word A in the learning speaker data space are represented by Vt2 = Vk2 + V2, where the difference vector is V2. That is, the difference vector V1 is added to the feature vector Vk1 of the standard speaker data space to obtain the feature vector value Vt1 of the learning speaker data, and the difference vector Vk2 is added to the feature vector value Vk2 of the word A in the standard speaker data.
2 plus the feature vector value Vt2 of the learning speaker data
Is obtained, the data conversion from the standard speaker data space to the learning speaker data space is performed. Similarly, the difference vector V3 is added to the feature vector Vk3 in the standard speaker data space to obtain the feature vector value Vt3 of the learning speaker data, and the difference vector V4 is added to the feature vector value Vk4 in the standard speaker data. Thus, a feature vector value Vt4 of the learning speaker data is obtained, and a feature vector value Vt5 of the learning speaker data is obtained by adding the difference vector V5 to the feature vector Vk5 of the standard speaker data space.

Using such a conversion function, utterance data of a new word obtained by uttering a standard speaker is converted into a learning speaker data space, so that pseudo learning on the new word is performed. Obtain word data. This will be described with reference to FIG.

FIG. 3A shows a new word N to be newly created in a database in a certain standard speaker data space.
Represents a feature vector sequence obtained by uttering the standard speaker, and the feature vector sequence is Vkn
1, Vkn2,..., Vkn5 (shown by white circles in the figure).

The feature vectors Vnk1, Vkn2,.
.., Vkn5, consider Vkn1 now.

In FIG. 3A, the feature vector Vkn
Among the feature vectors existing around 1 (feature vectors of about 200 words already stored as a database), some feature vectors existing at the closest positions are selected. That is, in the standard speaker data space, for example, 20
A large number of feature vectors constituting a feature vector sequence corresponding to 0 word are scattered in the space.
Some feature vectors close to the feature vector Vkn1 are selected. Here, it is assumed that three feature vectors close to the feature vector Vkn1 are selected, and Vk10, Vk20, and Vk30 are selected as the three feature vectors (indicated by black circles in the figure).

Then, the three selected feature vectors V
k10, Vk20, and Vk30 each have a difference vector to the learning speaker data space, and
10, V20 and V30. Using these difference vectors, the difference vector V for the feature vector Vkn1 is calculated.
Determine n1. This Vn1 is obtained by the following equation: Vn1 = μ1 · V10 + μ2 · V20 + μ3 · V30. In this equation, μ1, μ2, and μ3 are coefficients representing weights, μ1 is a weight according to the distance between Vkn1 and Vk10, μ2 is a weight according to the distance between Vkn1 and Vk20, and μ3 is a coefficient according to the distance between Vkn1 and Vk30. The magnitude of the weight is set according to each distance, and the closer the distance, the greater the weight. Note that μ1
Each value is set so that + μ2 + μ3 = 1.

In this way, the difference vector Vn1 with respect to the feature vector Vkn1 is determined, and using the difference vector Vn1, one feature vector Vkn1 constituting a new word feature vector sequence is obtained by Vtn1 = Vkn1 + Vn1. It is converted as a data space vector Vtn1.

Similarly, for each of the feature vectors Vkn2, Vkn3, Vkn4, and Vkn5 other than Vkn1, the respective difference vectors Vn2, Vn3, Vn4, and Vn5 are obtained by the above-described processing, and these difference vectors are calculated. Using the vector Vtn1, Vtn2, Vtn3, Vtn4, Vtn of the learning speaker data space.
Converted as 5. FIG. 3B shows how the feature vector sequence of the new word is converted into the learning speaker data space in this manner. This is referred to here as pseudo-learning word data.

The above process is a process between certain one standard speaker data and one learning speaker data. In practice, however, several standard speaker data and several hundred learning source data are used. Since speaker data exists, similar processing is performed for each.

With the above processing, for words that were not originally found in the database, based on the data uttered by several standard speakers, the pseudo-learning word data were stored in the learning speaker data space for several hundred people. Is created. And
Learning data is created using the pseudo-learning word data, and using the learning data, for example, the above-described DRNN
Learn a speech model.

FIG. 4 is a diagram showing an example of a device configuration for realizing the first embodiment. In general, a voice input unit 1, an A / D conversion unit 2, a voice analysis unit 3, a pseudo learning It comprises a word data creating section 4, an input data storage section 5, a learning data storage section 6, and the like.

The pseudo-learning word data creating section 4 includes a standard speaker data storage section 4 for storing standard speaker data.
1. Learning speaker data storage unit 4 for storing learning speaker data
2. It has a data conversion processing unit 43, and performs the processing described with reference to FIGS.

That is, of the hundreds of utterance data (feature vectors) already stored in the database, several arbitrarily selected persons are set as standard speakers, and the utterance data of each standard speaker is used as standard speaker data. The utterance data (feature vector) of each of the learning speakers is stored in the learning speaker data storage unit 42 while the others are stored as the learning speakers. Then, the conversion function (difference vector) from each standard speaker data space to the learning speaker data space
Is created for each word using about 200 words,
The data conversion unit 43 performs the above-described processing based on the conversion function to perform data conversion.

A process for creating learning data for a new word not in the database in such a configuration will be described.

First, when a certain standard speaker speaks a new word, the voice is input to the A / D conversion unit 2 via the voice input unit 1 and is A / D converted. Is converted into a feature vector represented by, for example, a 10-dimensional LPC cepstrum coefficient, and stored in the input data storage unit 5.

The input data conversion unit 43 is located at the closest position among the standard speaker feature vectors existing around the feature vector of the input voice stored in the input data storage unit 5 in the standard speaker data space. Several feature vectors are selected, and converted into respective learning speaker data using a difference vector obtained in advance. Thereby, pseudo learning word data is created in each learning data space, and the pseudo learning word data is stored in the learning data storage unit 6.
Is stored.

As described above, according to the first embodiment,
When creating a speech model for a new word that is not in the database, learning data for that word can be created based on utterance data generated by several standard speakers. Therefore, it is not necessary to collect hundreds of utterance data and create learning data in order to create a speech model for a new word as in the related art, and it is possible to create a speech model in a short period of time.

(Second Embodiment) In the first embodiment, the processing is performed using the feature vector itself represented by, for example, a 10-dimensional LPC cepstrum coefficient.
In the second embodiment, processing is performed using a code vector obtained by vector-quantizing a feature vector represented by an LPC cepstrum coefficient.

That is, when a certain standard speaker utters about 200 words for each word, if there are about 30 feature vectors per word, about 6000 feature vectors are generated. However, the standard speaker codebook is compiled by, for example, 256 code vectors into a standard speaker codebook, and the codebook is used to convert each code vector of the codebook into the learning speaker data space. In this case, a learning speaker codebook mapped to is prepared, and processing is performed using these codebooks. This processing will be described with reference to FIG.

The standard speaker codebook 50 in FIG.
Has 256 code vectors forming a code vector sequence for each word (about 200 words), but here, only code vector sequences for words A, B, and C are shown. . The learning speaker codebook 60 has a code vector mapped for each word in the standard speaker codebook 50. For example, the code vector Ck1 of the standard speaker codebook 50 corresponds to the code vector Ct1 of the learning speaker codebook 60, and the code vector Ck2 of the standard speaker codebook 50 corresponds to the code vector Ct2 of the learning speaker codebook 60. And standard speaker codebook 5
The code vector Ck3 of 0 is the learning speaker codebook 60.
Each code vector is associated with the corresponding code vector Ct3.

Using the standard speaker codebook 50 and the learning speaker codebook 60, a new word is vector-quantized by the standard speaker codebook 50, and the quantized code vector sequence is converted to a learning speaker code. By mapping to the book 60, pseudo-learning word data is created. This will be described with reference to FIG.

As shown in FIG. 6, a feature vector sequence of utterance data obtained by uttering a standard speaker with respect to a word N to be newly made into a database is represented by Vk1 and Vk.
2, Vk3,..., Vk7. This is vector quantized as follows.

Here, the standard speaker codebook has a size of 256, and is composed of 256 unspecified speaker code vectors. Then, the code vectors of these standard speaker codebooks 50 are Ck1, Ck2, Ck.
, Ck256, which is actually composed of 256 code vectors. In FIG. 6, for simplification of the drawing, these code vectors are Ck11, Ck1.
Only 10 code vectors 2, 2,..., Ck20 are shown (these code vectors are indicated by black circles in the figure).

With respect to such a standard speaker codebook 50, a feature vector sequence of a new word N obtained by uttering a standard speaker (indicated by white circles in the figure, feature vectors Vk1,
Vk2, Vk3,..., Vk7).

That is, each feature vector of the feature vector sequence of the word N and Ck scattered in the standard speaker codebook 50
1, Ck2, Ck3,..., Ck256, and the distance to the code vector is calculated, and the code vector having the shortest distance to the feature vector is selected. For example, the first and second feature vectors Vk1 and Vk of the feature vector sequence of word N
2 is associated with the code vector Ck11, the third feature vector Vk3 is associated with the code vector Ck13, the fourth feature vector Vk4 is associated with the code vector Ck14, the fifth, sixth, and seventh features. It is assumed that the vectors Vk5, Vk6, and Vk7 are respectively associated with the code vector Ck15.

Thus, the feature vector sequence of the word N is
Ck11, Ck11, Ck13, Ck14, Ck15,
Ck15 is replaced by a code vector sequence of Ck15.

When the feature vector sequence of a new word is quantized in this way, the quantized word N is obtained.
Are mapped to the learning speaker codebook 60. That is, since the code vectors of the standard speaker codebook 50 and the learning speaker codebook 60 correspond to each other, the standard speaker codebook 5
The code vector Ck11 of the word N at 0 is the code vector Ct11 of the learning story codebook 60, the code vector Ck13 is the code vector Ct13, the code vector Ck14 is the code vector Ct14, the code vector Ck15 is the code vector Ct15, and so on. Is mapped. Thus, the code vector sequence of the word N in the learning speaker codebook 60 is Ct1.
1, Ct11, Ct13, Ct14, Ct15, Ct1
5, Ct15.

As described above, based on the learning speaker's voice data in the database, the standard speaker codebooks for several speakers and the learning speaker codebooks for hundreds of other speakers are stored in the respective speakers. Is created for each speaker, the feature vector sequence obtained by uttering a new word by the standard speaker is vector-quantized to obtain a code vector sequence, and the code vector sequence is mapped to the learning speaker codebook. In this way, pseudo learning word data is created in each learning speaker codebook. Then, learning data is created using the pseudo-learning word data, and for example, the above-mentioned DRNN speech model is learned using the learning data.

FIG. 7 is a diagram showing an example of a device configuration for realizing the second embodiment. In general, a voice input unit 1, an A / D conversion unit 2, a voice analysis unit 3, a pseudo learning It is composed of a word data creating unit 4, an input data storing unit 5, a learning data creating unit 6, and the like.

The pseudo-learning word data creating section 4 includes a standard speaker code data storage section 44 for storing standard speaker data (code vectors), and a learning speaker for storing learning speaker data (code vectors). Code data storage unit 4
5. It has a code data conversion processing unit 46, and performs the processing described so far.

That is, of the hundreds of utterance data already stored as a database, arbitrarily selected several persons are set as standard speakers, and the utterance data of each of the standard speakers is vector-quantized to obtain a standard speaker code. Create a book,
The standard speaker code book data is stored in the standard speaker code data storage unit 44. A speaker other than the standard speaker is set as a learning speaker, and utterance data of each learning speaker is set as learning speaker data. Are mapped to create a learning speaker codebook, and the learning speaker codebook data is stored in the learning speaker code data storage unit 45.

When creating learning data for a word that is not in the database, when each of the standard speakers utters the new word, the voice is subjected to A / D conversion by the A / D conversion unit 2. After being converted into a signal of a feature vector represented by, for example, a 10-dimensional LPC cepstrum coefficient by the voice analysis unit 3, the signal is stored in the input data storage unit 5.

Then, the code data conversion processing section 46
The feature vector is vector-quantized using a standard speaker codebook, and then mapped to each learning speaker codebook. The pseudo-learning word data of each learning speaker created in this manner is stored in the learning data storage unit 6.
Is stored.

As described above, when creating a speech model for a new word that is not in the database, as in the first embodiment, based on speech data generated by several standard speakers,
Since it is possible to create learning data for that word, there is no need to collect hundreds of utterance data to create learning data in order to create a speech model for a new word, as in the past. Can create a voice model.

Further, in the second embodiment, since the data is subjected to vector quantization for processing, the data is somewhat coarse, but there is an advantage that the processing amount can be greatly reduced. On the other hand, the first embodiment has an advantage that high-precision data can be obtained although the processing amount is large.

In the first and second embodiments described above, when the standard speaker utters a new word and obtains the utterance data, it is not one utterance for one word but a plurality of utterances. By uttering and obtaining each utterance data, it is possible to obtain data with many variations corresponding to fluctuations in the utterance length.

Further, after learning data for a new word is created, when learning a speech model using the learning data, not only learning speaker data but also data spoken by a standard speaker is used as learning data. It is better to use it for learning.

In each of the above embodiments, several speakers are selected as standard speakers, but this can be done by one person. However, when a plurality of standard speakers are used, learning data with many variations can be obtained for the same word, so that a better learning result can be obtained.

When there are a plurality of standard speakers, a standard speaker may be selected for each learning speaker.
That is, if the number of standard speakers is, for example, 5 and the number of learning speakers is, for example, 200, pseudo-learning word data having 1000 variations for one word to be newly created in the database is created. Become. This is effective in that many variations can be obtained, but it takes a lot of processing time to learn, so a standard speaker with a short distance is selected and the utterance of the selected standard speaker is selected. Learning efficiency may be improved by creating learning data based on pseudo-speaker word data obtained by mapping the data.

For example, considering the first learning speaker,
The first learning speaker is the first standard speaker if the utterance data of the first standard speaker among the five standard speakers is closer to the utterance data of the other standard speakers. Learning data is created based on pseudo-speaker word data obtained by mapping the utterance data of only the speaker.

As described above, one standard speaker data closest to one learning speaker is selected, and pseudo learning speaker word data for the selected standard speaker is created. If this is the case, the subsequent processing can be made much more efficient.

The present invention can be applied not only to learning of a DRNN speech model, but also to learning of a speech model for another neural network or a speech model in another speech recognition means such as a speech model based on a hidden Markov method (HMM method). Of course, it can also be applied. Further, the present invention is not limited to the embodiments described above, and various modifications can be made without departing from the gist of the present invention.

The processing program for performing the processing of the present invention described above can be stored in a storage medium such as a floppy disk, an optical disk, or a hard disk. The present invention also includes the storage medium. Also,
A format for obtaining data from a network may be used.

[0074]

As described above, according to the present invention,
From the utterance data obtained from a large number of speakers previously held as a database, the utterance data of a small number of speakers are selected as standard speaker data, and the other as learning speaker data, a new When creating learning data for a word, the data uttered by the standard speaker for the new word is converted to a learning speaker data space using a predetermined conversion function, and the learning data for the new word is converted. Because it was created, there is no need to collect hundreds of utterance data and create training data to create a speech model for a new word, as in the past, and a speech model can be created in a short period of time at low cost. Can be created. So, for example, if you want to create a new voice model for a child, you can select a nearby adult woman as a standard speaker and already have it as a database without collecting hundreds of children's data. By mapping voice data of an adult female selected as a standard speaker to child data, learning data for children of new words is created, and a voice model of a new word is created based on the learning data. Can also be created.

As described above, according to the present invention, in order to create a speech model for a new word, it is not necessary to collect hundreds of utterance data to create learning data, which is a short time. A new voice model can be created at a low cost, whereby when a voice model of a word that does not have as a database is ordered from a user, it can be provided inexpensively with a short delivery date.

[Brief description of the drawings]

FIG. 1 is a flowchart illustrating a schematic process of the present invention.

FIG. 2 is a diagram illustrating a conversion function from standard speaker data to learning speaker data according to the first embodiment of the present invention.

FIG. 3 is a view for explaining a process of creating pseudo-learning word data according to the first embodiment;

FIG. 4 is a block diagram illustrating an example of a device configuration according to the first embodiment;

FIG. 5 is a diagram illustrating code mapping from standard speaker data to learning speaker data according to the second embodiment.

FIG. 6 is a view for explaining processing of vector quantization of input data and mapping of the code vector to a learning speaker codebook in the second embodiment.

FIG. 7 is a block diagram illustrating an example of a device configuration according to a second embodiment;

[Description of Signs] 1 Voice input unit 2 A / D conversion unit 3 Frequency analysis unit 4 Word data creation unit for pseudo learning 5 Input data storage unit 6 Learning data creation unit 41 Standard speaker data storage unit 42 Learning speaker data storage Part 43 Data conversion processing part Vk1, Vk2,... Standard speaker characteristic vector Vt1, Vt2,... Learning speaker characteristic vector V1, V2,... Difference vector N New word Vkn1, Vkn2,. New word feature vector Vtn1, Vtn2,... Pseudo-learning word data Ck1, Ck2,... Standard speaker code vector Ct1, Ct2,.

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Hiroo Hasegawa 3-5-5 Yamato, Suwa City, Nagano Prefecture Seiko Epson Corporation

Claims (6)

[Claims] 1. A speech model learning data creating method for creating learning data for learning a speech model for speech recognition, wherein at least one of utterance data obtained from a large number of speakers previously stored as a database. The speaker's utterance data is used as standard speaker data, and the others are used as learning speaker data.
When a conversion function from the standard speaker data space to the learning speaker data space is created based on the word data held in advance, and learning data for a new word is created, the standard language is used for the new word. A method for generating speech model learning data, wherein data obtained by uttering a speaker is converted into learning speaker data space using the conversion function to generate learning data for a new word. 2. The data existing in the standard speaker data space and the learning speaker data space is a feature vector for each word obtained by frequency analysis of a speech signal, and the new word The process of converting the data obtained by uttering the standard speaker into the learning speaker data space by using the conversion function is performed by using the feature vector and the learning speaker constituting each word in the standard speaker data space. The method according to claim 1, wherein the method is performed using a difference vector from a feature vector constituting each word in the data space. 3. The data present in the standard speaker data space and the learning speaker data space is code data obtained by vector-quantizing a feature vector for each word obtained by frequency analysis of a speech signal. Further, the process of converting the data obtained by the standard speaker uttering the new word into the learning speaker data space using the conversion function includes converting the new word data into the standard speaker data space. The method according to claim 1, wherein code data is obtained by vector quantization, and the code data is mapped to the learning speaker data space to convert the data from the standard speaker data space to the learning speaker data space. How to create speech model training data. 4. A speech model learning data creating apparatus for creating learning data for learning a speech model for speech recognition, wherein at least one of utterance data obtained from a large number of speakers, which is previously stored as a database. A standard speaker data storage unit for storing speaker data as standard speaker data, a learning speaker data storage unit for storing speech data other than the standard speaker as learning speaker data, and the standard speaker data Data conversion from space to learning speaker data space,
A pseudo-learning word data creation unit having a data conversion unit that performs a conversion using a conversion function obtained in advance; and a learning data storage unit that stores data created by the pseudo-learning word data creation unit. When creating learning data for a new word, the data obtained by uttering the standard speaker for the new word is converted into a learning speaker data space using the conversion function, and a new word is created. A speech model learning data creating apparatus, which creates pseudo-learning word data for and generates learning data using the pseudo-learning word data. 5. The standard speaker data stored in the standard speaker data storage unit and the learning speaker data stored in the learning speaker data storage unit are words obtained by frequency analysis of a speech signal. The feature vector of each
The process of converting the data obtained by the standard speaker uttering the new word into the learning speaker data space using the conversion function includes a feature vector constituting each word in the standard speaker data space. 5. The speech model learning data creating apparatus according to claim 4, wherein the processing is performed using a difference vector between the feature vector and a feature vector constituting each word in the learning speaker data space. 6. The standard speaker data stored in the standard speaker data storage unit and the learning speaker data stored in the learning speaker data storage unit are words obtained by frequency analysis of a speech signal. Code data obtained by vector-quantizing a feature vector for each word, and data obtained by the standard speaker uttering the new word is converted into a learning speaker data space using the conversion function. Processing is
New word data is vector-quantized in the standard speaker data space to obtain code data, and the code data is mapped to the learning speaker data space to convert the data from the standard speaker data space to the learning speaker data space. The voice model learning data creating apparatus according to claim 4, wherein