JP2002358093A - Speech recognition method, speech recognition device, and storage medium therefor - Google Patents

Abstract

(57) [Summary] [PROBLEMS] To provide a speech recognition method that realizes a speaker adaptation method that performs sufficient adaptation even with a small number of data by simple processing. SOLUTION: The speech recognition method of the present invention includes a step S12 of extracting a feature parameter sequence from an input speech, a step S14 of reading a parameter of a first acoustic model when adaptation is performed, and a first learning parameter set. A step of creating S15, a step of creating a second learning parameter set S16, a step of calculating an adaptive parameter from the first learning parameter set and the second learning parameter set S17, and a first sound using the adaptive parameter. Step S18 of creating parameters of the second acoustic model from the parameters of the model, and Step 110 of selecting either the first acoustic model or the second acoustic model when performing recognition.
And a step S112 of collating with a sequence of feature parameters using the selected acoustic model.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an unspecified speaker speech recognition technology capable of recognizing anyone's voice. TECHNICAL FIELD The present invention relates to a speech recognition method and a speech recognition device having a speaker adaptation function for adapting to a speech recognition device and a storage medium thereof.

[0002]

2. Description of the Related Art Conventionally, as a technique of a speaker adaptation method in speech recognition, for example, a technique described in Japanese Patent No. 3035239 is known.

FIG. 9 shows a processing flow chart of a conventional speaker adaptation method.

As shown in FIG. 9, the speaker adaptation method according to the prior art is based on the MLLR method, which is a first known technique (see C.I.
L. Leggetter et al., OchMaximumLikelihood Linear R
egression for Speaker Adaptation of Continuous Den
sity Hidden Markov Models ", Computer Speech and Lan
guage, Vol. 9, pp. 171-185, 1995 ”) and a MAP estimation method which is a second known technique (Reference 2“ CHLee et al., Pasted och ”).
A Study on Speaker Adaptation of the Parameters of
Continuous Density Hidden Markov Models ", IEEE Tra
nsactions on Signal Processing, Vol. 39, No. 4, pp. 806-
814, 1991 "). By this combination, the MLLR method, which is the first known technique, solves a problem that sufficient adaptation is not performed when the number of learning data is small.

[0005]

However, the conventional speaker adaptation method uses two learning algorithms, the MLLR method and the MAP estimation method, so that the processing is extremely complicated. .

SUMMARY OF THE INVENTION The present invention has been made to solve such a conventional problem, and a voice recognition method and a voice recognition apparatus for realizing a speaker adaptation method for sufficiently adapting even a small amount of data by simple processing. And its storage medium.

[0007]

According to the present invention, there is provided a speech recognition method comprising the steps of: analyzing an input speech and extracting a feature parameter sequence; and generating a feature parameter sequence if the speech is input for adaptation. Reading a parameter of a first acoustic model for an unspecified speaker, creating a first learning parameter set from the parameters of the first acoustic model, extracting the feature parameter sequence and the first Generating a second learning parameter set from the parameter series of the acoustic model, calculating an adaptive parameter from the first learning parameter set and the second learning parameter set, and performing multiple regression using the adaptive parameter. The parameters of the second acoustic model are created from the parameters of the first acoustic model by a predetermined mapping function based on the model. Performing a step of selecting either the first acoustic model or the second acoustic model when performing recognition, and performing a comparison with the sequence of the feature parameter using the selected acoustic model, Determining a recognition result from the collation result.

According to this method, even when the number of learning speakers is small, a speech recognition method having stable speaker adaptation can be realized by simple processing.

Further, in the speech recognition method according to the present invention, the acoustic model may be a continuous distribution HMM or an H approximating a continuous HMM.
An acoustic model modeled by MM, wherein parameters of the acoustic model are average vectors in the acoustic model, and individual elements of the learning parameter set are configured in a vector pair form; A learning parameter set is created while determining the vector pair by using a time axis matching method based on the Viterbi method, the DP method, or the Baum-Welch method with the parameter sequence of the acoustic model.

[0010] According to this method, a learning parameter set in which the input feature amount vector sequence and the model parameter sequence are associated on the time axis is obtained.
Accurate adaptation parameter estimation can be performed.

Further, in the speech recognition method according to the present invention, each element of the learning parameter set is constituted by a pair of vectors, and an element of one parameter vector of the pair is used as an objective variable in the multiple regression analysis. The other parameter vector of the pair is used as an explanatory variable, and the adaptive parameter is used as a partial regression coefficient in multiple regression analysis.

According to this method, the adaptive parameter can be obtained by the maximum likelihood estimation, and the adaptive parameter can be estimated with high accuracy.

The speech recognition method of the present invention further comprises the steps of: analyzing an input speech to extract a feature parameter sequence;
Reading the parameters of the first acoustic model for an unspecified speaker prepared in advance when the speech is speech input for adaptation, and determining whether the speech is the first speech in the adaptation, Creating a first learning parameter set from the parameters of the first acoustic model only, and creating a second learning parameter set from the feature parameter series and the parameter series of the first acoustic model. Generating and updating an adaptive parameter from the first learning parameter set and the second learning parameter set; and a predetermined mapping function based on a multiple regression model using the adaptive parameter. Creating the parameters of the second acoustic model from the parameters, and performing the first Selecting one of a sound model and a second sound model, performing a match with the sequence of the feature parameters using the selected sound model, and determining a recognition result from the match result. I decided to have it.

With this configuration, a speech recognition apparatus having a stable speaker adaptation function can be realized even when the number of learning speakers is small. Also, in the speech recognition method of the present invention, the acoustic model has a continuous distribution HMM. Or an acoustic model modeled by an HMM approximating a continuous HMM, wherein the parameters of the acoustic model are average vectors in the acoustic model, and each element of the learning parameter set is configured in the form of a vector pair; Viterbi between the feature parameter sequence and the parameter sequence of the acoustic model
A learning parameter set is created while determining the vector pair using a time axis matching method based on the method, the DP method, or the Baum-Welch method.

According to this configuration, a learning parameter set in which the input feature amount vector sequence and the model parameter sequence are associated on the time axis is obtained.
Accurate adaptation parameter estimation can be performed.

Further, in the speech recognition method according to the present invention, each element of the learning parameter set is constituted in the form of a vector pair, and an element of one parameter vector of the pair is used as an objective variable in the multiple regression analysis. The other parameter vector of the pair is used as an explanatory variable, and the adaptive parameter is used as a partial regression coefficient in multiple regression analysis.

With this configuration, it is possible to realize a storage medium having a stable speaker adaptation function even when the number of learning speakers is small.

The speech recognition method of the present invention further comprises the steps of: analyzing an input speech to extract a feature parameter sequence;
Reading the parameters of the first acoustic model for an unspecified speaker prepared in advance when the speech is speech input for adaptation, and determining whether the speech is the first speech in the adaptation, Creating a first learning parameter set from the parameters of the first acoustic model only, and creating a second learning parameter set from the feature parameter series and the parameter series of the first acoustic model. Generating and updating an adaptive parameter from the first learning parameter set and the second learning parameter set; and a predetermined mapping function based on a multiple regression model using the adaptive parameter. Creating the parameters of the second acoustic model from the parameters, and performing the first Determining whether to use the acoustic model or the second acoustic model for recognition, performing a selection control of the two types of acoustic models based on the determination result, and using the selected acoustic model. The method includes a step of performing matching with a series of feature parameters and a step of determining a recognition result from the matching result.

With this configuration, a learning parameter set in which the input feature vector series and the model parameter series are associated on the time axis is obtained.
Accurate adaptation parameter estimation can be performed.

Further, in the speech recognition method of the present invention, the acoustic model may be a continuous distribution HMM or an H approximating a continuous HMM.
An acoustic model modeled by MM, wherein parameters of the acoustic model are average vectors in the acoustic model, and individual elements of the learning parameter set are configured in a vector pair form; A learning parameter set is created while determining the vector pair by using a time axis matching method based on the Viterbi method, the DP method, or the Baum-Welch method with the parameter sequence of the acoustic model.

With this configuration, the adaptive parameter can be estimated based on the maximum likelihood, and the adaptive parameter can be estimated with high accuracy.

Further, in the speech recognition method according to the present invention, each element of the learning parameter set is formed in the form of a pair of vectors, and an element of one parameter vector of the pair is used as an objective variable in the multiple regression analysis. The other parameter vector of the pair is used as an explanatory variable, and the adaptive parameter is used as a partial regression coefficient in multiple regression analysis.

According to this method, even when the number of learning speakers is small, it is possible to realize a stable sequential adaptive speaker adaptation method.

Further, the speech recognition method of the present invention comprises the steps of: analyzing an input speech and extracting a feature parameter sequence;
When the voice is a voice input for adaptation, a step of reading parameters of a first acoustic model for an unspecified speaker prepared in advance, and determining whether or not this is the first utterance in the adaptation, Creating a first learning parameter set from the parameters of the first acoustic model only, and creating a second learning parameter set from the feature parameter series and the parameter series of the first acoustic model. Generating and updating an adaptive parameter from the first learning parameter set and the second learning parameter set; and a predetermined mapping function based on a multiple regression model using the adaptive parameter. Creating the parameters of the second acoustic model from the parameters, and performing the first A process of performing a sequence and matching of the feature parameters using the sound model and the second acoustic model,
Determining a recognition result from the collation result.

According to this method, a learning parameter set in which a sequence of the input feature amount vectors and a model parameter sequence are associated on the time axis is obtained.
Accurate adaptation parameter estimation can be performed.

Further, in the speech recognition method according to the present invention, the acoustic model may be a continuous distribution HMM or an H approximating a continuous HMM.
An acoustic model modeled by MM, wherein parameters of the acoustic model are average vectors in the acoustic model, and individual elements of the learning parameter set are configured in a vector pair form; A learning parameter set is created while determining the vector pair by using a time axis matching method based on the Viterbi method, the DP method, or the Baum-Welch method with the parameter sequence of the acoustic model.

According to this method, the adaptive parameter can be estimated on the basis of the maximum likelihood, and the adaptive parameter can be estimated with high accuracy.

Further, in the speech recognition method according to the present invention, each element of the learning parameter set is formed in the form of a pair of vectors, and an element of one parameter vector of the pair is used as an objective variable in the multiple regression analysis. The other parameter vector of the pair is used as an explanatory variable, and the adaptive parameter is used as a partial regression coefficient in multiple regression analysis.

With this configuration, a stable speaker adaptation apparatus can be realized even when the number of learning speakers is small.

Further, the speech recognition apparatus of the present invention analyzes the input speech and extracts a feature parameter sequence, and a first sound created beforehand for an unspecified speaker. First storage means for storing model parameters; means for generating a first learning parameter set from the parameters of the first acoustic model; and means for generating a first learning parameter set from the feature parameter series and the parameter series of the first acoustic model. Means for creating a second learning parameter set, means for merging the first learning parameter set and the second parameter set, and adaptation by a multiple regression analysis from the third learning parameter set obtained by the merging means. An adaptive parameter creating means for calculating a parameter, and a predetermined parameter based on a multiple regression model using the adaptive parameter. An acoustic model parameter conversion unit configured to generate a parameter of the second acoustic model from a parameter of the first acoustic model by using an image function; and a second storage unit configured to store the parameter of the second acoustic model. did.

According to this configuration, a learning parameter set in which the input feature vector sequence and the model parameter sequence are associated on the time axis is obtained.
Accurate adaptation parameter estimation can be performed.

Further, in the speech recognition apparatus of the present invention, the acoustic model may be a continuous distribution HMM or an H approximating a continuous HMM.
An acoustic model modeled by MM, wherein parameters of the acoustic model are average vectors in the acoustic model, and individual elements of the learning parameter set are configured in a vector pair form; A learning parameter set is created while determining the vector pair by using a time axis matching method based on the Viterbi method, the DP method, or the Baum-Welch method with the parameter sequence of the acoustic model.

According to this configuration, a storage medium having a stable speaker adaptation function can be provided even when the number of learning speakers is small. The speech recognition device of the present invention analyzes input speech and is characterized by Feature parameter extraction means for extracting a parameter sequence, first storage means for storing parameters of a first acoustic model prepared in advance for an unspecified speaker before adaptation, and parameters of the first acoustic model Means for creating a first learning parameter set only at the time of the first utterance of learning; means for creating a second learning parameter set from the feature parameter series and the parameter series of the first acoustic model; The first learning parameter set and the second
Means for combining the parameter sets of the above, an adaptive parameter creating means for calculating or updating the adaptive parameters from the third learning parameter set obtained by the combining means by a multiple regression analysis method, and a multiple regression model using the adaptive parameters Acoustic model parameter conversion means for creating a parameter of the second acoustic model from a parameter of the first acoustic model by a predetermined mapping function based on
And a second storage means for storing the parameters of the acoustic model.

According to this configuration, a learning parameter set is obtained in which the input feature amount vector sequence and the model parameter sequence are associated on the time axis.
Accurate adaptation parameter estimation can be performed.

Further, in the speech recognition apparatus according to the present invention, the acoustic model may be a continuous distribution HMM or an H approximating a continuous HMM.
An acoustic model modeled by MM, wherein parameters of the acoustic model are average vectors in the acoustic model, and individual elements of the learning parameter set are configured in a vector pair form; A learning parameter set is created while determining the vector pair by using a time axis matching method based on the Viterbi method, the DP method, or the Baum-Welch method with the parameter sequence of the acoustic model.

With this configuration, the adaptive parameter can be estimated with maximum likelihood, and the adaptive parameter can be estimated with high accuracy. The speech recognition apparatus of the present invention analyzes the input speech and extracts a feature parameter sequence. Feature parameter extraction means, first storage means for storing parameters of a first acoustic model prepared in advance for an unspecified speaker before adaptation, and first learning from the parameters of the first acoustic model Means for creating a parameter set only at the time of the first utterance of learning; means for creating a second learning parameter set from the feature parameter sequence and the parameter sequence of the first acoustic model; Means for merging the learning parameter set and the second parameter set, and a third learning parameter set obtained by the merging means. An adaptive parameter creating means for calculating or updating an adaptive parameter from a meter set by a multiple regression analysis method; and a second sound from a parameter of the first acoustic model by a predetermined mapping function based on a multiple regression model using the adaptive parameter. Acoustic model parameter conversion means for creating model parameters, second storage means for storing parameters of the second acoustic model, and either the first acoustic model or the second acoustic model for recognition. And a means for determining whether to recognize.

According to this method, when the number of learning samples is too small, it is possible to provide a speech recognition method which is automatically performed using an acoustic model of an unspecified speaker and has stable performance.

Further, in the speech recognition apparatus according to the present invention, the acoustic model may be a continuous distribution HMM or an H approximating the continuous HMM.
An acoustic model modeled by MM, wherein parameters of the acoustic model are average vectors in the acoustic model, and individual elements of the learning parameter set are configured in a vector pair form; A learning parameter set is created while determining the vector pair by using a time axis matching method based on the Viterbi method, the DP method, or the Baum-Welch method with the parameter sequence of the acoustic model.

According to this method, a learning parameter set is obtained in which the input feature amount vector sequence and the model parameter sequence are associated on the time axis.
Accurate adaptation parameter estimation can be performed.

Further, the speech recognition apparatus of the present invention analyzes the input speech and extracts a feature parameter sequence, and a first sound created beforehand for an unspecified speaker. First storage means for storing model parameters; means for creating a first learning parameter set from the parameters of the first acoustic model only at the time of the first utterance of learning; the feature parameter sequence and the first audio Means for creating a second learning parameter set from the model parameter series, means for merging the first learning parameter set and the second parameter set only at the time of the first utterance of learning, and a second means obtained by the merging means. An adaptive parameter creating means for calculating or updating an adaptive parameter from the learning parameter set by a multiple regression analysis method; Acoustic model parameter conversion means for creating a parameter of the second acoustic model from a parameter of the first acoustic model by a predetermined mapping function based on a multiple regression model using the adaptive parameter, and a parameter of the second acoustic model Second storage means for storing
Upon recognition, a configuration is provided that includes a verification unit that obtains a verification result for the first acoustic model and a verification result for the second acoustic model.

According to this method, the adaptive parameter can be estimated on the basis of the maximum likelihood, and the adaptive parameter can be estimated with high accuracy.

Further, in the speech recognition apparatus of the present invention, the acoustic model may be a continuous distribution HMM or an H approximating a continuous HMM.
An acoustic model modeled by MM, wherein the parameters of the acoustic model are average vectors in the acoustic model, and the individual elements of the learning parameter set are configured in the form of pairs of vectors; Viterbi method or DP between model parameter series
The learning parameter set is created while the vector pair is determined using the time axis matching method by the Baum-Welch method.

According to this configuration, when the number of learning samples is too small, the recognition using the acoustic model of the unspecified speaker is automatically performed, and a speech recognition device with stable performance can be provided.

The voice recognition storage medium of the present invention further comprises a step of analyzing the input voice and extracting a characteristic parameter sequence, and a step of generating an unspecified predetermined voice when the voice is voice input for adaptation. Reading the parameters of the first acoustic model for the speaker, creating a first learning parameter set from the parameters of the first acoustic model, and analyzing the feature parameter sequence and the first acoustic model. A step of creating a second learning parameter set from the parameter sequence, a step of calculating an adaptive parameter from the first learning parameter set and the second learning parameter set, and a predetermined step based on a multiple regression model using the adaptive parameter. Generating parameters of a second acoustic model from parameters of the first acoustic model by a mapping function of Selecting one of the first acoustic model and the second acoustic model when performing the recognition, performing a comparison with the sequence of the feature parameters using the selected acoustic model, And a program for determining a recognition result from the speech recognition method. The program is readable by a computer.

According to this configuration, a learning parameter set is obtained in which the input feature amount vector sequence and the model parameter sequence are associated on the time axis.
Accurate adaptation parameter estimation can be performed.

Further, in the speech recognition storage medium according to the present invention, the acoustic model is an acoustic model modeled by a continuous distribution HMM or an HMM approximating the continuous HMM, and a parameter of the acoustic model is an average vector in the acoustic model. The individual elements of the learning parameter set are configured in the form of pairs of vectors, and the time axis between the feature parameter sequence and the parameter sequence of the acoustic model is calculated using the Viterbi method, the DP method, or the Baum-Welch method. The learning parameter set is created while determining the vector pair using the matching method.

With this configuration, when the number of learning samples is too small, recognition using an acoustic model of an unspecified speaker is automatically performed, and a speech recognition device with stable performance can be provided.

Further, in the speech recognition storage medium of the present invention, each element of the learning parameter set is formed in the form of a vector pair, and the element of one parameter vector of the pair is used as an objective variable in multiple regression analysis. , The other parameter vector of the pair is used as an explanatory variable, and the adaptive parameter is used as a partial regression coefficient in multiple regression analysis.

With this configuration, a learning parameter set in which the input feature amount vector sequence and the model parameter sequence are associated on the time axis is obtained.
Accurate adaptation parameter estimation can be performed.

The voice recognition storage medium of the present invention further comprises a step of analyzing the input voice and extracting a characteristic parameter sequence, and a step of extracting the feature parameter sequence if the voice is input for adaptation. Reading the parameters of the first acoustic model for the speaker, and determining whether or not the first utterance in the adaptation, and determining the first learning parameter set from the parameters of the first acoustic model only at the first utterance. Creating a second learning parameter set from the feature parameter series and the parameter series of the first acoustic model; and adapting an adaptive parameter from the first learning parameter set and the second learning parameter set. Creating or updating, and a predetermined mapping function based on a multiple regression model using the adaptive parameters, The second from the parameters of the acoustic model of
Creating the parameters of the acoustic model, selecting one of the first acoustic model and the second acoustic model when performing the recognition, and using the selected acoustic model, And storing a program of a voice recognition method having a step of performing a check and a step of determining a recognition result from the result of the check, and making the program readable by a computer.

With this configuration, the adaptive parameter can be estimated with maximum likelihood, and the adaptive parameter can be estimated with high accuracy.

Further, in the speech recognition storage medium according to the present invention, the acoustic model is an acoustic model modeled by a continuous distribution HMM or an HMM approximating the continuous HMM, and a parameter of the acoustic model is an average vector in the acoustic model. The individual elements of the learning parameter set are configured in the form of pairs of vectors, and the time axis between the feature parameter sequence and the parameter sequence of the acoustic model is calculated using the Viterbi method, the DP method, or the Baum-Welch method. The learning parameter set is created while determining the vector pair using the matching method.

According to this method, when the speaker is changed by performing the adaptation successively, the acoustic model of the unspecified speaker becomes effective, and a speech recognition method capable of avoiding an extreme decrease in recognition performance is provided. realizable.

Further, in the speech recognition storage medium of the present invention, each element of the learning parameter set is formed in the form of a pair of vectors, and one parameter vector element of the pair is used as an objective variable in multiple regression analysis. , The other parameter vector of the pair is used as an explanatory variable, and the adaptive parameter is used as a partial regression coefficient in multiple regression analysis.

According to this method, a learning parameter set is obtained in which the input feature amount vector sequence and the model parameter sequence are associated on the time axis.
Accurate adaptation parameter estimation can be performed.

Further, the voice recognition storage medium of the present invention analyzes the input voice and extracts a characteristic parameter sequence, and further includes an unspecified predetermined voice if the voice is voice input for adaptation. Reading the parameters of the first acoustic model for the speaker, and determining whether or not the first utterance in the adaptation, and determining the first learning parameter set from the parameters of the first acoustic model only at the first utterance. Creating a second learning parameter set from the feature parameter series and the parameter series of the first acoustic model; and adapting an adaptive parameter from the first learning parameter set and the second learning parameter set. Creating or updating, and a predetermined mapping function based on a multiple regression model using the adaptive parameters, The second from the parameters of the acoustic model of
Creating a parameter of the acoustic model of the first acoustic model, determining whether to use the first acoustic model or the second acoustic model to perform the recognition, and performing the two types of acoustic model based on the determination result. A program for a voice recognition method including a step of performing model selection control, a step of performing matching with the sequence of feature parameters using the selected acoustic model, and a step of determining a recognition result from the matching result is stored. The program can be read by a computer.

According to this method, the adaptive parameter can be obtained by the maximum likelihood estimation, and the adaptive parameter can be estimated with high accuracy.

Further, in the speech recognition storage medium according to the present invention, the acoustic model is an acoustic model modeled by a continuous distribution HMM or an HMM approximating the continuous HMM, and a parameter of the acoustic model is an average vector in the acoustic model. The individual elements of the learning parameter set are configured in the form of pairs of vectors, and the time axis between the feature parameter sequence and the parameter sequence of the acoustic model is calculated using the Viterbi method, the DP method, or the Baum-Welch method. The learning parameter set is created while determining the vector pair using the matching method.

With this configuration, when the speaker is changed while performing the adaptation sequentially, the acoustic model of the unspecified speaker becomes effective, and a speech recognition apparatus capable of avoiding an extreme decrease in recognition performance is provided. realizable.

Further, in the speech recognition storage medium of the present invention, each element of the learning parameter set is formed in the form of a pair of vectors, and an element of one parameter vector of the pair is used as an objective variable in the multiple regression analysis. , The other parameter vector of the pair is used as an explanatory variable, and the adaptive parameter is used as a partial regression coefficient in multiple regression analysis.

According to this configuration, a learning parameter set in which the input feature amount vector sequence and the model parameter sequence are associated on the time axis is obtained.
Accurate adaptation parameter estimation can be performed.

The voice recognition storage medium of the present invention further comprises a step of analyzing the input voice and extracting a feature parameter sequence, and a step of extracting the feature parameter sequence when the voice is voice input for adaptation. Reading the parameters of the first acoustic model for the speaker, and determining whether or not the first utterance in the adaptation, and determining the first learning parameter set from the parameters of the first acoustic model only at the first utterance. Creating a second learning parameter set from the feature parameter series and the parameter series of the first acoustic model; and adapting an adaptive parameter from the first learning parameter set and the second learning parameter set. Creating or updating, and a predetermined mapping function based on a multiple regression model using the adaptive parameters, The second from the parameters of the acoustic model of
Creating the parameters of the acoustic model, performing the recognition using the first acoustic model and the second acoustic model to perform matching with the feature parameter sequence, and determining the recognition result from the matching result. And storing the program of the voice recognition method having the steps of:

According to this configuration, when the speaker is changed after the successive adaptation, the acoustic model of the unspecified speaker becomes effective, and a storage medium capable of avoiding extreme reduction in recognition performance is realized. it can.

In the speech recognition storage medium according to the present invention, the acoustic model is an acoustic model modeled by a continuous distribution HMM or an HMM approximating the continuous HMM, and a parameter of the acoustic model is an average vector in the acoustic model. The individual elements of the learning parameter set are configured in the form of pairs of vectors, and the time axis between the feature parameter sequence and the parameter sequence of the acoustic model is calculated using the Viterbi method, the DP method, or the Baum-Welch method. The learning parameter set is created while determining the vector pair using the matching method.

According to this configuration, a learning parameter set in which the input feature amount vector sequence and the model parameter sequence are associated on the time axis is obtained.
Accurate adaptation parameter estimation can be performed.

Further, in the speech recognition storage medium of the present invention, the individual elements of the learning parameter set are formed in the form of pairs of vectors, and the elements of one of the pairs of parameter vectors are used as objective variables in multiple regression analysis. , The other parameter vector of the pair is used as an explanatory variable,
The adaptation parameters were used as partial regression coefficients in multiple regression analysis.

With this configuration, the adaptive parameter can be obtained by maximum likelihood estimation, and the adaptive parameter can be estimated with high accuracy.

[0068]

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

FIG. 1 shows a flowchart of the voice recognition method according to the first embodiment of the present invention.

In FIG. 1, S11 is a process of inputting voice, and S12 is a process of analyzing the input voice and extracting a feature parameter sequence. In S13, an adaptive mode or a recognition mode is selected. S14 is a step of reading parameters of a first acoustic model, which is a previously created acoustic model for an unspecified speaker, and S15 is a step of extracting all or a part of the parameters of the first acoustic model and extracting the first acoustic model. Is a process of creating a learning parameter set, and S16 is a process of creating a second learning parameter set by extracting parameters of all or a part of the feature parameter sequence and the parameter sequence of the first acoustic model, and S17 is a process of creating a second learning parameter set. This is a process of calculating an adaptive parameter from the first learning parameter set and the second parameter set. In step S18, a predetermined mapping function based on a multiple regression model (described later) using the adaptive parameter is used to calculate a parameter from the first acoustic model. This is a process of creating a second acoustic model, and S19 is a process for creating parameters of the second acoustic model. A process for storing. The adaptive processing is performed by obtaining the second acoustic model created through steps S14 to S19 as parameters after the adaptation.

At the time of recognition, the first acoustic model or the second acoustic model
Step S110 of selecting any one of the acoustic models, Step S111 of performing matching with a sequence of feature parameters using the selected acoustic model, Step S122 of determining a recognition result from the matching result, and outputting the recognition result. Step S133
The voice recognition is performed through the process described above.

At the time of adaptation, an adaptive parameter is obtained by maximum likelihood estimation using a parameter set obtained by combining the first learning parameter set and the second parameter learning parameter set. Since the parameters of the acoustic model of the unspecified speaker are included in the learning sample in advance by the merging of the parameters, even when the number of learning speakers is small, it is possible to stably obtain an accurate adaptive parameter.

FIG. 2 shows a configuration diagram of a speech recognition apparatus according to a second embodiment of the present invention.

In FIG. 2, the input voice is analyzed by the acoustic analysis unit 21 and the feature parameter extraction unit 22 extracts a feature parameter sequence. The switching unit 23 switches between “adaptive mode” and “recognition mode”. First, the case of the “adaptive mode” will be described. First
The storage device 24 stores the parameters of the first acoustic model created in advance as an acoustic model for unspecified speakers before adaptation. The first read from the storage device 24
All or a part of the parameters of the acoustic model are extracted, and the first learning parameter set creating unit 25 creates the first learning parameter set. The parameters of all or a part of the feature parameter series and the parameter series of the first acoustic model are extracted, and the second learning parameter set creating unit 26 creates a second learning parameter set. The learning parameter merging unit 27 merges the first learning parameter set and the second parameter set. From the third learning parameter set obtained by the learning parameter merging unit 27, an adaptive parameter is calculated by an adaptive parameter creating unit 28 by a multiple regression analysis method and stored in a second storage device 29. The acoustic model conversion unit 210 creates parameters of the second acoustic model from parameters of the first acoustic model using a predetermined mapping function based on a multiple regression model using adaptive parameters, and stores the parameters in the third storage unit 211.

Next, the configuration and operation in the “recognition mode” will be described. Feature parameter extraction unit 22
In the feature parameters extracted in step (1), a matching score with the acoustic model selected by the control unit 214 and the acoustic model selecting unit 215 is obtained by the matching unit 212, and the result determining unit 217
Yields the final result.

By providing a learning parameter set merging unit 27 and including an acoustic model for an unspecified speaker in the learning parameter set, an adaptive parameter with high accuracy can be stably obtained even when the learning data is small. be able to.

FIG. 3 shows a flowchart of the voice recognition method according to the third embodiment of the present invention.

In the first embodiment, it is assumed that the recognition is performed after a certain number of presentations of the learning speech are made, but this embodiment shows a configuration in which the learning can be performed sequentially.

In FIG. 3, S31 is a process of inputting a voice, and S32 is a process of analyzing the input voice and extracting a feature parameter sequence. In S33, an adaptive mode or a recognition mode is selected. In step S34, a process of reading parameters of a first acoustic model, which is an acoustic model for an unspecified speaker prepared in advance, is performed. In branch S35, it is determined whether or not learning is the first time. In a learning parameter set creation step S36, all or a part of the parameters of the first acoustic model are extracted to create a first learning parameter set. S37
Is a process of extracting a parameter of the feature parameter sequence and all or a part of the parameter sequence of the first acoustic model to create a second learning parameter set;
Reference numeral 38 denotes a process of creating intermediate parameters required for sequential learning. S39 is a process of calculating an adaptive parameter from the first learning parameter set and the second parameter set, and S310 is a process of calculating a first acoustic model parameter from a parameter of the first acoustic model by a predetermined mapping function based on a multiple regression model using the adaptive parameter. The second acoustic model is a process of creating the second acoustic model, and S311 is a process of storing the second acoustic model. By obtaining the second acoustic model created through the processes from S34 to S311 as a parameter after the adaptation, the adaptive process per utterance in the sequential adaptation is performed.

At the time of recognition, the first acoustic model or the second acoustic model
Step S312 of selecting one of the acoustic models, Step S313 of performing matching with a sequence of feature parameters using the selected acoustic model, Step S314 of determining a recognition result from the matching result, and Step of outputting the recognition result Voice recognition is performed through step S315.

As described above, the creation of the first learning parameter set by the method of the third embodiment is performed only at the time of the first utterance of the adaptation, and the intermediate parameters necessary for the successive adaptation are stored for each utterance. By doing so, it becomes possible to update the adaptive parameter for each utterance, and it is possible to realize sequential learning.

FIG. 4 shows a configuration diagram of a speech recognition apparatus according to a fourth embodiment of the present invention.

This embodiment shows an example of the configuration of a speech recognition apparatus that can perform successive adaptation.

In FIG. 4, an input voice is analyzed by an acoustic analysis unit 41, and a feature parameter extraction unit 42 extracts a feature parameter sequence. The switching unit 43 switches between “adaptive mode” and “recognition mode”.

First, the case of the "adaptive mode" will be described. The first storage device 44 stores the parameters of the first acoustic model created in advance as an acoustic model for unspecified speakers before adaptation. All or a part of the parameters of the first acoustic model read from the first storage device 44 are extracted, and the first learning parameter set creating unit 45 creates the first learning parameter set. This operation is performed by the virtual switches 43 and 481 only at the first utterance of learning. The parameters of all or a part of the feature parameter series and the parameter series of the first acoustic model are extracted, and the second learning parameter set creating unit 46 creates a second learning parameter set. The learning parameter set merging unit 47 merges the first and second learning parameter sets only at the first utterance of learning.

From the second learning parameter set obtained by the learning parameter set merging unit 47 or the second learning parameter set generating unit 46, the adaptive parameter generating unit 48 calculates the adaptive parameters by the multiple regression analysis method.
Is stored in the storage device 49. The acoustic model conversion unit 410 creates parameters of the second acoustic model from parameters of the first acoustic model using a predetermined mapping function based on a multiple regression model using adaptive parameters,
11 is stored.

Next, the configuration and operation in the “recognition mode” will be described. The feature parameter extracted by the feature parameter extracting unit 42 is compared with the control unit 41 by the matching unit 412.
4 and a matching score with the acoustic model selected by the acoustic model selecting unit 415, and the result determining unit 417 obtains the final result.

By providing a learning parameter set merging unit 47 and including an acoustic model for an unspecified speaker in the learning parameter set, it is possible to stably obtain accurate adaptive parameters even when the learning data is small. In addition to the above, the first learning parameter set is added to the second learning parameter set only at the time of the first utterance of learning, and the adaptation parameter is updated for each utterance at the time of the utterance other than the first time, thereby realizing the sequential type adaptation. It becomes possible to do.

FIG. 5 shows a flowchart of the voice recognition method according to the fifth embodiment of the present invention.

In this embodiment, in addition to the sequential learning made possible by the configuration of the third embodiment, a selective control function of the acoustic model is newly added to the recognition processing, so that the number of learning samples is extremely small. 2 shows an example of a method for stabilizing the recognition performance in.

In FIG. 5, S51 is a process of inputting a voice, and S52 is a process of analyzing the input voice and extracting a feature parameter sequence. In S53, the adaptive mode or the recognition mode is selected. S54 is a step of reading the parameters of the first acoustic model, which is an acoustic model for an unspecified speaker, created in advance. In branch S55, it is checked whether or not the learning is the first time. In a learning parameter set creation step S56, all or a part of the parameters of the first acoustic model are extracted to create a first learning parameter set. S57
Is a process of extracting a parameter of the feature parameter sequence and all or a part of the parameter sequence of the first acoustic model to create a second learning parameter set;
Reference numeral 58 denotes a process of creating intermediate parameters required for sequential learning. S59 is a process of calculating an adaptive parameter from the first learning parameter set and the second parameter set. S510 is a first acoustic model based on a predetermined mapping function based on a multiple regression model (described later) using the adaptive parameter. Is a process of creating a second acoustic model from the parameters of the above. S511 is a process of storing the second acoustic model. By obtaining the second acoustic model created through the processes from S54 to S511 as a parameter after the adaptation, the adaptive process per utterance in the sequential adaptation is performed.

At the time of recognition, the first acoustic model or the second acoustic model
Determining which one of the acoustic models is selected S5
12, a step S514 of performing matching with a sequence of feature parameters using the selected acoustic model, a step S515 of determining a recognition result from the matching result, and a step of outputting the recognition result. A recognition process is performed through the process of S516.

As described above, the creation of the first learning parameter set by the method of the fifth embodiment is performed only at the time of the first utterance of the adaptation, and the intermediate parameters necessary for the successive adaptation are stored for each utterance. By doing so, it becomes possible to update the adaptive parameter for each utterance, and it is possible to realize sequential learning.

In the above-described embodiment, by adding the selective control function S512 and the selection control S513 of the acoustic model to the recognition process, the recognition performance can be stabilized when the number of learning samples is extremely small. become.

FIG. 6 shows a configuration diagram of a speech recognition apparatus according to a sixth embodiment of the present invention.

In this embodiment, there is shown an example of a speech recognition apparatus capable of successive adaptation, which is designed to stabilize the recognition performance when the number of learning samples is extremely small.

In FIG. 6, an input voice is analyzed by an acoustic analysis unit 61 and a feature parameter extraction unit 62 extracts a feature parameter sequence. The switching unit 63 switches between “adaptive mode” and “recognition mode”. First, the case of the “adaptive mode” will be described. First
The storage device 64 stores the parameters of the first acoustic model created in advance as an acoustic model for unspecified speakers before adaptation. All or a part of the parameters of the first acoustic model read from the first storage device 64 are extracted, and the first learning parameter set creating unit 65 creates a first learning parameter set. This operation is performed by the virtual switches 63 and 681 only at the first utterance of learning. All or a part of the parameter series of the feature parameter series and the parameter series of the first acoustic model are extracted, and the second learning parameter set creating unit 66 creates a second learning parameter set. The learning parameter set merging unit 67 merges the first and second learning parameter sets only at the first utterance of learning.

From the learning parameter sets obtained by the learning parameter set merging section 67 or the second learning parameter set generating section 66, the adaptive parameters are calculated by the adaptive parameter generating section 68 by the multiple regression analysis method, and the second storage device 69 is provided. To be stored. The acoustic model conversion unit 611 creates parameters of the second acoustic model from parameters of the first acoustic model using a predetermined mapping function based on a multiple regression model using adaptive parameters, and stores the parameters in the third storage device 612.

Next, the case of the "recognition mode" will be described. The feature parameter extracted by the feature parameter extraction unit 62 is used by the matching unit 612 to obtain a matching score with the acoustic model selected by the control unit 614, the determination unit 615, and the acoustic model selection unit 616, and the result determination unit 61
7 gives the final result.

In this embodiment, in addition to the sequential learning made possible by the configuration of the fourth embodiment, a selective control function of the acoustic model is realized by the control unit 614 and the result determination unit 615, and a new processing is performed for the recognition processing. , It is possible to stabilize the recognition performance when the number of learning samples is extremely small.

FIG. 7 shows a flowchart of the voice recognition method according to the seventh embodiment of the present invention.

In this embodiment, in addition to the sequential learning made possible by the configuration of the third embodiment, a configuration for preventing an extreme decrease in recognition performance even when the speaker changes during the sequential learning is realized. I have.

In FIG. 7, S71 is a process of inputting a voice, and S72 is a process of analyzing the input voice and extracting a feature parameter sequence. In S73, an adaptive mode or a recognition mode is selected. Step S74 is a step of reading parameters of a first acoustic model which is an acoustic model for an unspecified speaker prepared in advance. In branch S75, it is checked whether learning is the first time or not. In a parameter set creation step S76, all or a part of the parameters of the first acoustic model are extracted to create a first learning parameter set. Step S77 is a step of extracting all or a part of the parameter series of the feature parameter and the parameter series of the first acoustic model to create a second learning parameter set.
Reference numeral 8 denotes a process of creating intermediate parameters required for sequential learning. S79 is a process of calculating an adaptive parameter from the first learning parameter set and the second parameter set, and S710 is a process of calculating the first parameter by a predetermined mapping function based on a multiple regression model (described later) using the adaptive parameter.
Is a process of creating a second acoustic model from the parameters of the acoustic model of Step S711, and S711 is a process of storing the second acoustic model. By obtaining the second acoustic model created through the processes from S74 to S711 as a parameter after the adaptation, the adaptive process per utterance in the sequential adaptation is performed.

At the time of recognition, the feature parameter series and the first acoustic model are collated in S712, the collation with the second acoustic model is performed in S713, and the collation results for the two types of acoustic models are determined in S714. After determining that the result with the highest score is the final result, the recognition result is output in S715.

In the above embodiment, S712 and S71
By constantly comparing the acoustic model before adaptation with the acoustic model after adaptation in step 3, even if the speaker changes, it is possible to avoid an extreme decrease in the recognition rate.

FIG. 8 is a diagram showing the configuration of a speech recognition apparatus according to an eighth embodiment of the present invention.

This embodiment shows an example of a speech recognition apparatus capable of successive adaptation, which is designed to stabilize the recognition performance when the number of learning samples is extremely small.

In FIG. 8, an input voice is analyzed by an acoustic analysis unit 81, and a feature parameter extraction unit 82 extracts a feature parameter sequence. The switching unit 83 switches between “adaptive mode” and “recognition mode”. First, the case of the “adaptive mode” will be described. First
The storage device 84 stores the parameters of the first acoustic model created in advance as an acoustic model for unspecified speakers before adaptation. All or a part of the parameters of the first acoustic model read from the first storage device 84 are extracted, and the first learning parameter set creating unit 85 creates a first learning parameter set. This operation is performed by the virtual switches 83 and 881 only at the first utterance of learning. The parameters of all or a part of the feature parameter series and the parameter series of the first acoustic model are extracted, and the second learning parameter set creation unit 86 creates a second learning parameter set. The learning parameter set merging unit 87 merges the first and second learning parameter sets only at the first utterance of learning.

From the learning parameter sets obtained by the learning parameter set merging unit 87 or the second learning parameter set generating unit 86, the adaptive parameters are calculated by the adaptive parameter generating unit 88 by the multiple regression analysis method, and the second storage unit 89 To be stored. The acoustic model conversion unit 810 creates parameters of the second acoustic model from parameters of the first acoustic model using a predetermined mapping function based on a multiple regression model using adaptive parameters, and stores the parameters in the third storage unit 811.

Next, the case of the "recognition mode" will be described. The feature parameter extracted by the feature parameter extraction unit 82 is compared with the first acoustic model and the second
The matching score with the acoustic model is determined, and the result determination unit 8
17 gives the final result.

In this embodiment, in addition to the sequential learning made possible by the configuration of the fourth embodiment, it is necessary to always collate both the first acoustic model and the second acoustic model. Accordingly, it is possible to prevent an extremely low recognition rate even when the speaker changes while using the recognition device.

The recording medium according to the ninth embodiment of the present invention stores a program for causing a computer to execute the method according to the first embodiment, and is readable by the computer.

The recording medium according to the tenth embodiment of the present invention stores a program for causing a computer to execute the method according to the third embodiment, and is readable by the computer.

The recording medium according to the eleventh embodiment of the present invention stores a program for causing a computer to execute the method according to the fifth embodiment, and is readable by the computer.

Further, a recording medium according to the twelfth embodiment of the present invention stores a program for causing a computer to execute the method according to the seventh embodiment, and is readable by the computer.

[0116]

As described above, according to the present invention, the first learning parameter set created from the acoustic model of the unspecified speaker, the characteristic parameter of the input signal, and the second training model created from the acoustic model of the unspecified speaker. To provide a speech recognition method that realizes stable speaker adaptation with a simple process even with a small number of learning data by obtaining an adaptation parameter using a third learning parameter set obtained by merging the learning parameter sets. Can be.

[Brief description of the drawings]

FIG. 1 is a diagram showing a flowchart of a speech recognition method according to a first embodiment of the present invention;

FIG. 2 is a configuration diagram of a speech recognition device according to a second embodiment of the present invention;

FIG. 3 is a diagram showing a flowchart of a voice recognition method according to a third embodiment of the present invention.

FIG. 4 is a configuration diagram of a voice recognition device according to a fourth embodiment of the present invention.

FIG. 5 is a diagram showing a flowchart of a speech recognition method according to a fifth embodiment of the present invention.

FIG. 6 is a configuration diagram of a speech recognition device according to a sixth embodiment of the present invention.

FIG. 7 is a diagram showing a flowchart of a speech recognition method according to a seventh embodiment of the present invention.

FIG. 8 is a configuration diagram of a speech recognition device according to an eighth embodiment of the present invention.

FIG. 9 is a processing flowchart of a speaker adaptation method according to the related art.

[Explanation of symbols]

 S11 Voice input process S12 Feature parameter extraction process S13 Mode selection process S14 First acoustic model parameter reading process S15 First learning parameter set creation process S16 Second learning parameter set creation process S17 Adaptive parameter calculation process S18 Second Acoustic model parameter creation process S19 Second acoustic model parameter storage process S110 Acoustic model selection process S111 Matching process S112 Recognition result determination process S113 Recognition result output process 21 Acoustic analysis unit 22 Feature parameter extraction unit 23 Switching switch 24 First Storage device 25 First learning parameter set creation unit 26 Second learning parameter set creation unit 27 Learning parameter set merging unit 28 Adaptive parameter creation unit 29 Second storage device 210 Acoustic model conversion unit 21 1 third storage device 212 collation unit 215 acoustic model selection unit 214 control unit 217 result determination unit

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Nobuyuki Kunieda 4-3-1 Tsunashima Higashi, Kohoku-ku, Yokohama-shi, Kanagawa Prefecture Inside Matsushita Communication Industrial Co., Ltd. (72) Kazuya Nomura Tsunashima-higashi, Kohoku-ku, Yokohama, Kanagawa Prefecture 4-3-1, Matsushita Communication Industrial Co., Ltd. F-term (reference) 5D015 GG03 HH00

Claims (32)

[Claims] 1. A process of analyzing a inputted speech to extract a feature parameter sequence, and a first sound for an unspecified speaker prepared in advance when the speech is a speech inputted for adaptation. Reading the parameters of the model, creating a first learning parameter set from the parameters of the first acoustic model, and acquiring a second learning parameter from the feature parameter series and the parameter series of the first acoustic model. Creating a set, calculating an adaptive parameter from the first learning parameter set and the second learning parameter set, and performing the first mapping by a predetermined mapping function based on a multiple regression model using the adaptive parameter. Second from the parameters of the acoustic model
Creating the parameters of the acoustic model, selecting one of the first acoustic model and the second acoustic model when performing the recognition, and using the selected acoustic model, And a step of determining a recognition result from the comparison result. 2. The acoustic model is an acoustic model modeled by a continuous distribution HMM or an HMM approximating a continuous HMM, a parameter of the acoustic model is an average vector in the acoustic model, and each of the learning parameter sets is Are configured in the form of a vector pair, and the vector pair of the feature parameter sequence and the parameter sequence of the acoustic model are matched using a time axis matching method by the Viterbi method, the DP method, or the Baum-Welch method. 2. The speech recognition method according to claim 1, wherein a learning parameter set is created while being determined. 3. The individual elements of the learning parameter set are formed in the form of pairs of vectors, the elements of one parameter vector of the pair are used as objective variables in multiple regression analysis, and the other of the pair is The speech recognition method according to claim 1, wherein the speech recognition method is used as an explanatory variable, and the adaptation parameter is used as a partial regression coefficient in multiple regression analysis. 4. A process for analyzing an input voice to extract a feature parameter sequence, and a first acoustic model for an unspecified speaker prepared in advance when the voice is a voice input for adaptation. Reading the parameters of
A step of determining whether or not the first utterance in the adaptation and creating a first learning parameter set from the parameters of the first acoustic model only at the time of the first utterance; Creating a second learning parameter set from the parameter series, creating or updating an adaptive parameter from the first learning parameter set and the second learning parameter set, and performing a multiple regression model using the adaptive parameter. A step of creating a parameter of the second acoustic model from a parameter of the first acoustic model by a predetermined mapping function based on the first acoustic model, and selecting one of the first acoustic model and the second acoustic model when performing recognition. Using the selected acoustic model to collate with the sequence of feature parameters And degree, the speech recognition method characterized in that it comprises a step of determining a recognition result from the verification result. 5. The acoustic model is an acoustic model modeled by a continuous distribution HMM or an HMM approximating a continuous HMM, a parameter of the acoustic model is an average vector in the acoustic model, and each of the learning parameter sets is Is configured in the form of a pair of vectors, and between the feature parameter sequence and the parameter sequence of the acoustic model, using the time axis matching method by the Viterbi method, the DP method, or the Baum-Welch method, 5. The speech recognition method according to claim 4, wherein a learning parameter set is created while being determined. 6. An individual element of the learning parameter set is configured in the form of a vector pair, an element of one parameter vector of the pair is used as an objective variable in a multiple regression analysis, and the other parameter vector of the pair is The speech recognition method according to claim 4, wherein the speech recognition method is used as an explanatory variable, and the adaptive parameter is used as a partial regression coefficient in a multiple regression analysis. 7. A step of analyzing an input voice to extract a feature parameter sequence, and a first acoustic model for an unspecified speaker prepared in advance when the voice is input voice for adaptation. Reading the parameters of the
Determining whether or not the first utterance in the adaptation and creating a first learning parameter set from the parameters of the first acoustic model only at the time of the first utterance; Creating a second learning parameter set from the parameter series, creating or updating an adaptive parameter from the first learning parameter set and the second learning parameter set, and performing a multiple regression model using the adaptive parameter. Creating a parameter of the second acoustic model from parameters of the first acoustic model by a predetermined mapping function based on the first acoustic model, and performing recognition using either the first acoustic model or the second acoustic model. And performing the selection control of the two types of acoustic models based on the determination result. Cormorants process and the steps of performing a sequence and matching of the feature parameter using the selected acoustic model, the speech recognition method characterized in that it comprises a step of determining a recognition result from the verification result. 8. The acoustic model modeled by a continuous distribution HMM or an HMM approximating the continuous HMM, wherein the parameters of the acoustic model are average vectors in the acoustic model, and each of the learning parameter sets Is configured in the form of a pair of vectors, and between the feature parameter sequence and the parameter sequence of the acoustic model, using the time axis matching method by the Viterbi method, the DP method, or the Baum-Welch method, The speech recognition method according to claim 7, wherein a learning parameter set is created while being determined. 9. An individual element of the learning parameter set is formed in the form of a vector pair, an element of one parameter vector of the pair is used as an objective variable in the multiple regression analysis, and the other of the pair is defined as the parameter vector. The speech recognition method according to claim 7, wherein the speech recognition method is used as an explanatory variable, and the adaptation parameter is used as a partial regression coefficient in a multiple regression analysis. 10. A process of analyzing an input speech to extract a feature parameter sequence, and a first acoustic model for an unspecified speaker prepared in advance when the speech is input speech for adaptation. Reading a parameter, determining whether or not the first utterance in the adaptation, and generating a first learning parameter set from the parameters of the first acoustic model only at the time of the first utterance; And a step of creating a second learning parameter set from the parameter series of the first acoustic model; a step of creating or updating an adaptation parameter from the first learning parameter set and the second learning parameter set; A predetermined mapping function based on the multiple regression model using the parameters, (2) a step of creating parameters of the acoustic model, (2) a step of performing matching with the feature parameter sequence using the first acoustic model and the second acoustic model when performing recognition, and Determining the speech. 11. The acoustic model is an acoustic model modeled by a continuous distribution HMM or an HMM approximating a continuous HMM, a parameter of the acoustic model is an average vector in the acoustic model, and each of the learning parameter sets is Is configured in the form of a pair of vectors, and between the feature parameter sequence and the parameter sequence of the acoustic model, using the time axis matching method by the Viterbi method, the DP method, or the Baum-Welch method, 11. The speech recognition method according to claim 10, wherein a learning parameter set is created while being determined. 12. An individual element of the learning parameter set is configured in the form of a vector pair, an element of one parameter vector of the pair is used as an objective variable in a multiple regression analysis, and the other parameter vector of the pair is The speech recognition method according to claim 10, wherein the speech recognition method is used as an explanatory variable, and the adaptation parameter is used as a partial regression coefficient in a multiple regression analysis. 13. A feature parameter extracting means for analyzing an input speech to extract a feature parameter sequence, and a first parameter for storing parameters of a first acoustic model prepared in advance for an unspecified speaker before adaptation. Means for generating a first learning parameter set from the parameters of the first acoustic model, and generating a second learning parameter set from the feature parameter series and the parameter series of the first acoustic model. Means, means for merging the first learning parameter set and second parameter set, and adaptive parameter creation means for calculating an adaptive parameter from the third learning parameter set obtained by the merging means by a multiple regression analysis method A first mapping function based on a multiple regression model using the adaptive parameters, A speech recognition apparatus comprising: an acoustic model parameter conversion unit configured to generate a parameter of a second acoustic model from a model parameter; and a second storage unit configured to store the parameter of the second acoustic model. 14. An acoustic model modeled by a continuous distribution HMM or an HMM approximating a continuous HMM, wherein the parameters of the acoustic model are average vectors in the acoustic model, and each of the learning parameter sets Is configured in the form of a pair of vectors, and between the feature parameter sequence and the parameter sequence of the acoustic model, using the time axis matching method by the Viterbi method, the DP method, or the Baum-Welch method, 15. The speech recognition device according to claim 14, wherein a learning parameter set is created while being determined. 15. A feature parameter extracting means for analyzing an input speech to extract a feature parameter sequence, and a first parameter for storing parameters of a first acoustic model prepared in advance for an unspecified speaker before adaptation. Storage means, means for creating a first learning parameter set based on the parameters of the first acoustic model only at the first utterance of learning, and second means based on the feature parameter series and the parameter series of the first acoustic model. Means for creating a learning parameter set, means for merging the first learning parameter set and the second parameter set only at the time of the first utterance of learning, and multiple regression from the third learning parameter set obtained by the merging means. An adaptive parameter generating means for calculating or updating an adaptive parameter by an analysis method; and Acoustic model parameter conversion means for creating a parameter of the second acoustic model from a parameter of the first acoustic model by a predetermined mapping function based on a multiple regression model, and a second storing the parameters of the second acoustic model. A speech recognition device comprising: a storage unit. 16. The acoustic model is an acoustic model modeled by a continuous distribution HMM or an HMM approximating a continuous HMM, a parameter of the acoustic model is an average vector in the acoustic model, and each of the learning parameter sets is Is configured in the form of a pair of vectors, and between the feature parameter sequence and the parameter sequence of the acoustic model, using the time axis matching method by the Viterbi method, the DP method, or the Baum-Welch method, 16. The speech recognition device according to claim 15, wherein a learning parameter set is created while being determined. 17. A feature parameter extracting means for analyzing an input speech to extract a feature parameter sequence, and a first parameter storing first acoustic model parameters prepared in advance for an unspecified speaker before adaptation. Storage means, means for creating a first learning parameter set based on the parameters of the first acoustic model only at the first utterance of learning, and second means based on the feature parameter series and the parameter series of the first acoustic model. Means for creating a learning parameter set, means for merging the first learning parameter set and the second parameter set only at the time of the first utterance of learning, and multiple regression from the third learning parameter set obtained by the merging means. An adaptive parameter generating means for calculating or updating an adaptive parameter by an analysis method; and Acoustic model parameter conversion means for creating a parameter of the second acoustic model from a parameter of the first acoustic model by a predetermined mapping function based on a multiple regression model, and a second storing the parameters of the second acoustic model. A speech recognition apparatus comprising: a storage unit; and a unit configured to determine whether to use the first acoustic model or the second acoustic model for recognition. 18. An acoustic model modeled by a continuous distribution HMM or an HMM approximating a continuous HMM, wherein the parameters of the acoustic model are average vectors in the acoustic model, and each of the learning parameter sets Is configured in the form of a pair of vectors, and between the feature parameter sequence and the parameter sequence of the acoustic model, using the time axis matching method by the Viterbi method, the DP method, or the Baum-Welch method, 18. The speech recognition device according to claim 17, wherein a learning parameter set is created while being determined. 19. A feature parameter extracting means for analyzing an input voice to extract a feature parameter sequence, and a first parameter storing a parameter of a first acoustic model prepared in advance for an unspecified speaker before adaptation. Storage means, means for creating a first learning parameter set based on the parameters of the first acoustic model only at the first utterance of learning, and second means based on the feature parameter series and the parameter series of the first acoustic model. Means for creating a learning parameter set, means for merging the first learning parameter set and the second parameter set only at the time of the first utterance of learning, and multiple regression from the third learning parameter set obtained by the merging means. An adaptive parameter generating means for calculating or updating an adaptive parameter by an analysis method; and Acoustic model parameter conversion means for creating a parameter of the second acoustic model from a parameter of the first acoustic model by a predetermined mapping function based on a multiple regression model, and a second storing the parameters of the second acoustic model. A speech recognition apparatus comprising: a storage unit; and a matching unit that obtains a matching result with respect to the first acoustic model and a matching result with respect to the second acoustic model upon recognition. 20. The acoustic model is an acoustic model modeled by a continuous distribution HMM or an HMM approximating a continuous HMM, a parameter of the acoustic model is an average vector in the acoustic model, and each of the learning parameter sets Is configured in the form of a pair of vectors, and between the feature parameter sequence and the parameter sequence of the acoustic model, using the time axis matching method by the Viterbi method, the DP method, or the Baum-Welch method, 20. The speech recognition apparatus according to claim 19, wherein the learning parameter set is created while being determined. 21. A process for analyzing an input speech to extract a feature parameter sequence, and a first acoustic model for an unspecified speaker prepared in advance when the speech is input speech for adaptation. Reading a parameter, a step of creating a first learning parameter set from the parameters of the first acoustic model, and a second learning parameter set from the feature parameter series and the parameter series of the first acoustic model. , A step of calculating an adaptive parameter from the first learning parameter set and the second learning parameter set, and a first mapping function based on a multiple regression model using the adaptive parameter. Second from model parameters
Creating the parameters of the acoustic model, selecting one of the first acoustic model and the second acoustic model when performing the recognition, and using the selected acoustic model, A voice recognition storage medium storing a program of a voice recognition method including a step of performing a verification and a step of determining a recognition result from the verification result, and making the program readable by a computer. 22. An acoustic model modeled by a continuous distribution HMM or an HMM approximating a continuous HMM, wherein the parameters of the acoustic model are average vectors in the acoustic model, and each of the learning parameter sets Is configured in the form of a pair of vectors, and between the feature parameter sequence and the parameter sequence of the acoustic model, using the time axis matching method by the Viterbi method, the DP method, or the Baum-Welch method, 22. The speech recognition storage medium according to claim 21, wherein a learning parameter set is created while being determined. 23. Each element of the learning parameter set is formed in the form of a vector pair, an element of one parameter vector of the pair is used as an objective variable in multiple regression analysis, and the other of the pair is The speech recognition storage medium according to claim 21, wherein the speech recognition storage medium is used as an explanatory variable, and the adaptation parameter is used as a partial regression coefficient in a multiple regression analysis. 24. A process for analyzing an input speech to extract a characteristic parameter sequence, and a first acoustic model for an unspecified speaker prepared in advance when the speech is input speech for adaptation. Reading a parameter, determining whether or not the first utterance in the adaptation, and generating a first learning parameter set from the parameters of the first acoustic model only at the time of the first utterance; And a step of creating a second learning parameter set from the parameter series of the first acoustic model; a step of creating or updating an adaptation parameter from the first learning parameter set and the second learning parameter set; A predetermined mapping function based on the multiple regression model using the parameters, 2) a step of creating parameters of the acoustic model, a step of selecting either the first acoustic model or the second acoustic model at the time of recognition, and a step of selecting the feature parameter using the selected acoustic model. A speech recognition storage medium storing a program of a speech recognition method including a step of performing a collation with a sequence and a step of determining a recognition result from the comparison result, and making the program readable by a computer. 25. The acoustic model is an acoustic model modeled by a continuous distribution HMM or an HMM approximating a continuous HMM, the parameters of the acoustic model are average vectors in the acoustic model, and each of the learning parameter sets Is configured in the form of a pair of vectors, and between the feature parameter sequence and the parameter sequence of the acoustic model, using the time axis matching method by the Viterbi method, the DP method, or the Baum-Welch method, 25. The speech recognition storage medium according to claim 24, wherein a learning parameter set is created while being determined. 26. Each element of the learning parameter set is formed in the form of a vector pair, the element of one parameter vector of the pair is used as an objective variable in multiple regression analysis, and the other of the pair is The speech recognition storage medium according to claim 24, wherein the speech recognition storage medium is used as an explanatory variable, and the adaptation parameter is used as a partial regression coefficient in multiple regression analysis. 27. A process of analyzing an input voice to extract a feature parameter sequence, and a first acoustic model for an unspecified speaker prepared in advance when the voice is a voice input for adaptation. Reading a parameter, determining whether or not the first utterance in the adaptation, and generating a first learning parameter set from the parameters of the first acoustic model only at the time of the first utterance; And a step of creating a second learning parameter set from the parameter series of the first acoustic model; a step of creating or updating an adaptation parameter from the first learning parameter set and the second learning parameter set; A predetermined mapping function based on the multiple regression model using the parameters, A step of creating parameters of the acoustic model 2; a step of determining whether to use the first acoustic model or the second acoustic model when performing recognition; and a step of determining the two types based on the determination result. A program for a voice recognition method, comprising the steps of: performing a selection control of an acoustic model; performing a comparison with the feature parameter sequence using the selected acoustic model; and determining a recognition result from the comparison result. A voice-recognition storage medium that stores the program and makes the program readable by a computer. 28. The acoustic model is an acoustic model modeled by a continuous distribution HMM or an HMM approximating a continuous HMM, a parameter of the acoustic model is an average vector in the acoustic model, and each of the learning parameter sets is Is configured in the form of a pair of vectors, and between the feature parameter sequence and the parameter sequence of the acoustic model, using the time axis matching method by the Viterbi method, the DP method, or the Baum-Welch method, The speech recognition storage medium according to claim 27, wherein a learning parameter set is created while being determined. 29. The individual elements of the learning parameter set are formed in the form of pairs of vectors, the elements of one parameter vector of the pair are used as objective variables in multiple regression analysis, and the other of the pair is The speech recognition storage medium according to claim 27, wherein the speech recognition storage medium is used as an explanatory variable, and the adaptation parameter is used as a partial regression coefficient in a multiple regression analysis. 30. A process for analyzing an input voice to extract a feature parameter sequence, and a first acoustic model for an unspecified speaker prepared in advance when the voice is a voice input for adaptation. Reading a parameter, determining whether or not the first utterance in the adaptation, and generating a first learning parameter set from the parameters of the first acoustic model only at the time of the first utterance; And a step of creating a second learning parameter set from the parameter series of the first acoustic model; a step of creating or updating an adaptation parameter from the first learning parameter set and the second learning parameter set; A predetermined mapping function based on the multiple regression model using the parameters, (2) a step of creating parameters of the acoustic model, (2) a step of performing matching with the feature parameter sequence using the first acoustic model and the second acoustic model when performing recognition, and Storing a program for a voice recognition method having a step of deciding, and making the program readable by a computer. 31. The acoustic model is an acoustic model modeled by a continuous distribution HMM or an HMM approximating a continuous HMM, the parameter of the acoustic model is an average vector in the acoustic model, and each of the learning parameter sets is Is configured in the form of a pair of vectors, and between the feature parameter sequence and the parameter sequence of the acoustic model, using the time axis matching method by the Viterbi method, the DP method, or the Baum-Welch method, 31. The speech recognition storage medium according to claim 30, wherein a learning parameter set is created while determining. 32. Each element of the learning parameter set is formed in the form of a vector pair, an element of one parameter vector of the pair is used as an objective variable in multiple regression analysis, and the other parameter vector of the pair is 31. The speech recognition storage medium according to claim 30, wherein the speech recognition storage medium is used as an explanatory variable, and the adaptation parameter is used as a partial regression coefficient in a multiple regression analysis.