JP2000259169A - Voice recognition device and its recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the precision of voice recognition in the case of generating an acoustic model using the voice sample of an unspecified speaker. SOLUTION: An estimation sample selecting section 112 computes the distance between the voice sample of an unspecified speaker of a voice database 109 and the voice sample of a specific speaker of a task voice database, selects the voice sample of a voice database that has a smaller distance to the voice sample of the specific speaker and transmits the sample and a writing start text to a parameter estimating section 111.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus for performing speech recognition using the parameters of an acoustic model, and a recording medium therefor.

[0002]

2. Description of the Related Art Conventionally, applications corresponding to types of speakers, such as specific speakers and unspecified speakers, and applications corresponding to types of devices used for speech recognition (hereinafter, applications are collectively referred to as tasks). A speech recognition device that can be limited has been proposed. In order to improve the recognition accuracy of such a speech recognition device, it is effective to use an acoustic model limited to a recognition task. Estimation of the parameters of the acoustic model requires a large number of speech samples and a text describing the utterance content of the speech sample (referred to as transcribed text). For example, it is necessary for a specific speaker to be recognized to input speech corresponding to a task, and for a speech recognition device to generate samples sufficient for estimating parameters.

However, for a specific speaker, inputting a voice according to a task imposes a heavy burden. Therefore, a voice sample that does not specify a task, for example, a voice sample input by a plurality of speakers is used. Will be.

[0004] Therefore, a speech recognition method has been proposed in which utterance samples of a relatively small number of persons to be recognized are used for a task ("Maximum Likelihood Lin").
iearregression for Speaker Adaptation of Continuou
s Density Hidden Markov Moders ”CJ Leggettec and
PC Woodland, Computer Speech and Language 1995 (9) 1
71-185). In this proposal, the above-described model is used as an acoustic model.

[0005]

Since speech samples used for acoustic parameter estimation are collected from a plurality of speakers, speech samples having characteristics relatively close to the characteristics of a specific speaker to be recognized and adversely affecting recognition accuracy. And voice samples with speaker characteristics that give

The acoustic model estimated including samples having speaker characteristics different from the target speaker causes a reduction in recognition accuracy.

In view of the above, it is an object of the present invention to provide a method for estimating acoustic parameters from speech samples of a plurality of speakers for speech recognition of a specific speaker, which has an adverse effect on recognition accuracy. An object of the present invention is to provide a speech recognition device capable of excluding a sample and a recording medium for the device.

[0008]

In order to achieve the above object, according to the first aspect of the present invention, an acoustic model is created from a speech sample and a transcribed text, and an input speech is created based on the created acoustic model. In voice recognition for voice recognition, a first storage unit storing a plurality of first voice samples that are not bound by a speaker and a plurality of transcribed texts corresponding to the first voice samples, and a second voice sample corresponding to a task. A stored second storage unit, and a first audio sample closer to the second audio sample recorded in the second storage unit from among the plurality of first audio samples stored in the first storage unit. And a data processing unit for extracting a transcript text corresponding to the selected first audio sample from the first storage unit. Ri by the issued transcription text and first audio samples corresponding to the transcription text, characterized in that to create the acoustic model.

According to a second aspect of the present invention, in the voice recognition apparatus according to the first aspect, there are a plurality of the second voice samples, and the selecting means includes a plurality of the second voice samples and a plurality of the second voice samples. Calculating a distance between each of the one audio samples and selecting the closer first audio sample based on the calculation result.

According to a third aspect of the present invention, in the speech recognition apparatus according to the second aspect, the selecting unit determines whether the calculation result is within an allowable range, and determines that the calculation result is within the allowable range. Is obtained, the first voice sample used for the calculation result is selected as a voice sample used for the acoustic model.

According to a fourth aspect of the present invention, in the speech recognition apparatus according to the second aspect, the selecting means rearranges the calculation results in ascending order of distance, and selects the one of the rearranged calculation results having the shortest distance. , A selected number of calculation results are selected, and the first voice sample used for the selected calculation result is selected as a voice sample used for the acoustic model.

According to a fifth aspect of the present invention, in the voice recognition apparatus according to the first aspect, clustering is performed on a plurality of voice samples independent of a speaker, and a result of the clustering is used as the second voice sample as the second voice sample. It is characterized by further comprising clustering means for storing in the storage means.

According to a sixth aspect of the present invention, in accordance with the voice recognition apparatus of any one of the first to fifth aspects, the second voice sample corresponding to the task is a voice sample for a specific speaker. It is characterized by the following.

According to a seventh aspect of the present invention, there is provided a storage medium in which an acoustic model is created from a speech sample and a transcribed text, and a program executed in speech recognition for performing speech recognition of an input speech based on the created acoustic model. In the above, the voice recognition device may include a first storage unit storing a plurality of first voice samples that are not bound by a speaker and a plurality of transcribed texts corresponding thereto, and a second voice sample corresponding to a task. And a second storage means for storing the second sound recorded in the second storage means from among the plurality of first sound samples stored in the first storage means. The first, closer to the sample
And an information processing step of extracting a transcript text corresponding to the selected first audio sample from the first storage means. The acoustic model is created from a first audio sample corresponding to the transcribed text.

According to an eighth aspect of the present invention, in the recording medium of the voice recognition device according to the seventh aspect, there are a plurality of the second voice samples, and in the selecting step, each of the plurality of second voice samples is A distance between each of the plurality of first audio samples is calculated, and the closer first audio sample is selected based on the calculation result.

According to a ninth aspect of the present invention, in the recording medium of the voice recognition device according to the eighth aspect, the selecting step includes:
Determining whether or not the calculation result is within an allowable range; if a determination result indicating that the calculation result is within the allowable range is obtained, the first voice sample used for the calculation result is used for an acoustic model; It is characterized in that it is selected as a sample.

According to a tenth aspect of the present invention, in the recording medium of the speech recognition apparatus according to the eighth aspect, in the selecting step, the calculation results are rearranged in ascending order of the distance, and the most distant of the rearranged calculation results is obtained. Is selected from the closest one, and a first voice sample used for the selected calculation result is selected as a voice sample to be used for the acoustic model.

[0018] According to an eleventh aspect of the present invention, in the recording medium of the voice recognition device according to the seventh aspect, clustering is performed on a plurality of voice samples independent of a speaker, and a result of the clustering is used as the second voice sample. Second
And further comprising a clustering step of storing in the storage means.

According to a twelfth aspect of the present invention, there is provided the seventh to first aspects.
2. The method according to claim 1, wherein the second voice sample corresponding to the task is a voice sample for a specific speaker, according to the recording medium of the voice recognition device.

[0020]

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

Prior to the description of the speech recognition apparatus to which the present invention is applied, a speech recognition method of the present invention, more specifically, a method of estimating acoustic parameters will be described.

In this embodiment, a relatively small number of utterance samples (voice samples) of the recognition target are used as references (samples used for comparison), and voice samples close to the recognition target are obtained from voice samples of different speakers. Select a sample. Thus, the parameters of the acoustic model specialized for the task are estimated. In order to select a nearby voice sample, the distance between the voice sample used as a reference and each of a number of voice samples of different speakers (distance between speech samples) is used.

In this embodiment, a plurality of reference voice samples are further clustered (similar ones are collected into one group), and the estimated number of samples of each acoustic model is secured by the above-described method. Create an acoustic model with.

Hereinafter, a method for selecting a speech sample close to the speech sample to be recognized from a large number of speech samples from different speakers will be described.

The sample used for estimation is selected based on the distance (inter-speech sample distance) between each of a plurality of speech samples of different speakers and each of the speech samples of the recognition target person used as a reference. . Since the sample selection method has nonlinearity in the relationship between the distance between speech samples and the speech recognition accuracy, one or both of the following selection methods of Abs and RT are used.

The Abs selection method is a method of selecting, from a plurality of voice samples from different speakers, a voice sample whose distance from any reference voice sample is equal to or less than a predetermined value x.

The Rt selection method is a method of selecting from a plurality of different audio samples by the number of r% of the samples whose distance is closer to the reference sample.

When calculating the distances between a plurality of samples for different speakers and different utterance contents and a plurality of samples for reference individually, and selecting a voice sample having a short distance by the above-described method, the present embodiment is used. Then, a plurality of audio samples used for reference are clustered.
In this clustering, for example, a method called complete chain division hierarchy method (Toshio Nishida, thorough study of multivariate analysis,
Is divided into a plurality of clusters (groups). Prepare speech samples selected from multiple samples with different speakers for each clustered reference, and estimate the parameters of the acoustic model to perform speech recognition according to the task,
Improve the speech recognition accuracy compared to the past.

Several embodiments to which the present invention is applied will be described.

(First Embodiment) FIG. 1 shows a system configuration according to a first embodiment of the present invention. The first embodiment is an example in which an estimation data selection unit is arranged in an acoustic model parameter estimation unit in a speech recognition device. In FIG. 1, a recognition voice input unit 101 inputs a voice to be recognized. The acoustic analysis unit 102 analyzes the input speech and outputs an acoustic feature vector sequence. The acoustic features include information indicating the frequency characteristics of the input speech. The speech recognition decoder 103 performs speech recognition on a word basis using the language model 104 and the acoustic model 106, and outputs a speech recognition result 105. For speech recognition, a speech recognition method (Ro
nald Rosenfeld, “The CMU Statistical Language Mod
eling Toolkit and its use in the 1994 ARPA CSR Eva
luation ”, Proceedings of the Spoken Language Syst
ems Technology Workshop, pp. 47-50 (1995.1) can be used.

More specifically, the language model 104 uses a built-in word string that is grammatically (or semantically) combined and a word that has been speech-recognized so far to be recognized. The linguistic certainty (language score) of the word speech that is present is given to the speech recognition decoder 103. The acoustic model 106 compares the candidate word phoneme string from the word pronunciation dictionary 107 with the input speech (the form of the acoustic feature vector string), calculates the likelihood (acoustic score) that the input speech is the word, and calculates the acoustic score. This is given to the speech recognition decoder 103.

The parameter estimating unit 111 estimates acoustic model parameters for calculating an acoustic score based on the acoustic parameters of the audio sample selected by the estimated sample selecting unit 112 and the transcript indicating the utterance content of the audio sample. The estimated sample selection unit 112 includes a task voice database 108 and a voice database 109.
And an estimated sample determination unit 110. The task voice database 108 is a database that stores voice samples in the form of acoustic parameters, which specify the task.
One of the tasks is to store a voice sample of a specific speaker in the task voice database 108.

The speech database 109 is a database that stores speech samples (acoustic parameter forms) of unspecified speakers (independent of speakers) and transcribed texts thereof. The estimated sample determination unit 110 determines a voice sample whose speaker characteristic is close to a voice sample of a specific task (for example, a specific speaker) previously specified from the outside in the task voice database by the Abs selection method or the Rt selection method. Select from the audio database 109.

(First Embodiment of Estimated Sample Selector) FIG. 2
2 shows a circuit configuration of the estimated sample selection unit 112 when the Abs selection method is used. 2, the same parts as those in FIG. 1 are denoted by the same reference numerals, and the detailed description will be omitted. The utterance distance matrix calculation unit 206 stores the speech database 109
Are input and m voice samples 204 of the task voice database 108 are input, and a distance between the voice samples is calculated by a predetermined utterance sample distance calculation formula. The memory is temporarily stored in a matrix form of xn (referred to as an utterance distance matrix). The value of each row of the utterance distance matrix is the shortest distance search unit 2
07.

The shortest distance search unit 207 detects the value of the shortest distance for each row, and gives the comparison operation unit 209 the value of the shortest distance and the identification name (identification) of the voice sample in the voice database 109 used for calculating the value. ID) is output. The comparison operation unit 209 compares the value of the shortest distance of each row with the value x of an externally provided parameter, and determines the identification ID whose value of the shortest distance is within an allowable range (in this case, a value smaller than x) in the estimated sample list. Output to the creation output unit 211.

The estimation sample list creation output unit 211
Search the voice database 109 and enter the identification ID
Is read out and output to the parameter estimating unit 111 in FIG. 1, and a voice sample close to the voice sample in the voice database 109 is selected.

Any of the following formulas can be calculated as the formula used for calculating the distance between speech samples.

(1) Log Likelihood Ratio

[0039]

(Equation 1)

(Reference: “Segregation of Speakers
for Speech Recognition and SpeakerIdentification
Herbert Gish, Man-Hung Siu, and Robin Rohliek.
ICA SSP (91) 873-172) (2) Arithmetic Harmonic Sp
hericity

[0041]

(Equation 2)

(Reference: Text-Free Spe
aker Recognition Using an
Arithmetic-harmonic Sphe
risk Measure "Bimbot. F and Mat
han L. Eurospeech 93 (1) 169-172) (3) Gaussian Divergence

[0043]

(Equation 3)

(Reference: "Speaker Clustering Using
Direct Maximization of the MLLR-Adapted Likelihoo
d ”SEJohnson and PC Woodland. ICSLP-98 # 726) where X is a reference sample, Y is a sample in the database, μ _X , μ _Y , Σ _X , Σ _Y are X,
It is a mean vector and a covariance matrix of Y samples. N
_X , N _Y , N are the number of vectors in each sample and X,
The sum of the number of vectors at Y, D is the number of dimensions of the vector, and I is the unit matrix. Also,

[0045]

(Equation 4)

Is as follows.

(Second Embodiment of Estimated Sample Selector) FIG.
9 shows a circuit configuration of the estimated sample selection unit 112 using the Rt sample selection method. 3, the same parts as those in FIG. 1 are denoted by the same reference numerals, and detailed description thereof will be omitted.

The utterance distance matrix calculation unit 305 receives n speech samples (sound parameter form) 303 in the speech database 109 and m speech samples (sound parameter form) in the task speech database 108, and Calculate the distance between the two audio samples. The calculation result is stored in the form of m × n matrices together with the identification ID of the voice sample, and the utterance distance sorting unit 30
6 is output. The utterance distance matrix is rearranged in the utterance distance sorter 306 in ascending order of distance (from smaller to larger) for each column of the matrix. In the estimated sample list extracting unit 307, the rearranged utterance distance matrix has a fixed number, that is, m × n × r%
The audio samples, in this case, the identification IDs, are selected in ascending order of the number corresponding to the number of, and the overlapping distance values and IDs are removed for each column. The selected identification ID is output to the estimated sample list creation output unit 308. The estimated sample list creation / output unit 306 searches the voice database 109 and extracts the acoustic parameters corresponding to the identification ID and the transcribed text thereof.
Is output to the parameter estimating unit 111.

(Second Embodiment) Next, a second embodiment in which a plurality of acoustic models are created for each task will be described. FIG. 4 shows a system configuration of the second embodiment. 4, a voice is input from a recognition voice input unit 401. The input speech is converted into acoustic parameters by the acoustic analysis unit. The speech decoder 403 performs speech recognition of a recognition target acoustic parameter (input speech) using the language model 404 and the plurality of acoustic models 406, and a recognition result 405.
Is output. The word pronunciation dictionary 407 gives the acoustic model 406 a candidate word phoneme sequence. The above components are the same as those of the first embodiment except that there are a plurality of acoustic models.

In this embodiment, the same number of estimated sample determination units 110 and parameter estimation units 111 as the number of acoustic models are prepared. Also, the audio samples in the task audio database 108 (408 in FIG. 4) of FIG. 1 are clustered into four groups (ones with short distances are
Into one group (cluster)), and each group of the clustered audio samples is stored in a plurality of databases 416. The voice samples in the individual clustered database 416 and the voice samples in the voice database 408 (corresponding to the voice database 109 in FIG. 1) are compared (distance calculation between them), and the voice samples in the voice database having similar acoustic characteristics correspond. Is selected by the estimated sample determination unit 412. A corresponding parameter estimating unit 41 is output from each estimated sample determining unit 412.
The selected voice sample and the transcribed text are output to No. 4.

The second embodiment has an advantage that the recognition accuracy is improved because a more detailed task-limited acoustic model is used as compared with the first embodiment.

(Third Embodiment) A third embodiment in which a general-purpose computer (personal computer) executes a speech recognition program to realize a speech recognition device will be described.

FIG. 5 shows the system configuration of the third embodiment. In FIG. 5, a CPU 500, a system memory 51
0, input device, disk reading device 530, hard disk storage device (hereinafter abbreviated as hard disk) 5
40, a display 550 and a voice input device 560 are connected to the bus.

CPU 500 performs voice processing according to a voice recognition program loaded in system memory 510. In the present embodiment, it is possible to perform processing for creating an acoustic model and a language model, and perform speech recognition processing using the acoustic model and the language model.

The CPU 500 further includes a hard disk 54
System control is performed according to the operating system stored in “0”.

The system memory 510 includes a ROM and an RA.
M, and stores a program executed by the CPU 500, data used for calculation, and the like. The input device 520 has a keyboard and a mouse, and inputs various commands and data to the CPU 500 using the keyboard and the mouse. The input device 520 can also input transcript text to be registered in the audio database.

The disk reader 530 reads recorded data from a portable recording medium such as a floppy or a CDROM. In the present embodiment, the above-described speech recognition program is read from a recording medium.

The hard disk 540 includes an operating system used for system control, a speech recognition program, a language model used for the speech recognition program, an acoustic model, and a speech database (the speech database 40 shown in FIG. 4).
8), a task voice database (corresponding to the task voice database 410 in FIG. 4), and other data required for voice recognition. The display 550 displays data input from the input device 520, a voice recognition result, and the like under the control of the CPU 500. Voice input device 5
Reference numeral 60 has a microphone and an analog-to-digital converter, and outputs an input voice in the form of a digital voice signal that can be processed by the CPU 500. The input voice may be a target for voice recognition or a target for registration in the task voice database.

The speech recognition program is divided into a program portion for performing speech recognition itself and a program portion for creating an acoustic model. As the speech recognition program portion, a program for performing speech recognition using a language model and an acoustic model as in the related art can be used. In addition, since the program other than the audio sample selection process according to the present invention in the program for creating the acoustic model can be the same as the conventional one, the program for the audio sample selection process will be described.

FIG. 6 shows an example of a program for executing the audio sample selecting process. This program forms a part of a speech recognition program, and is executed by a CPU when an instruction to execute creation of an acoustic model is input from the input device 520.
500. Also, the hard disk 540
It is assumed that audio samples and transcription samples are registered in the upper audio database, and audio samples are registered in the task audio database in advance.

In FIG. 6, CPU 500 clusters speech samples in a speech database (a method called complete chain division hierarchy method (see Toshio Nishida, thorough study of multivariate analysis, published by Hyundai Mathematics) n clustering). The result of clustering is stored in the form of a database (corresponding to the clustered task voice database 416 in FIG. 4) on the hard disk 540 (step 1000). For example, if the audio samples are classified into four clusters, four databases are constructed.

The CPU 500 designates the first cluster, reads the first voice sample from the corresponding database (referred to as a cluster-specific database), and temporarily stores it in the system memory 510. Next, the first voice sample (with identification ID) is read from the voice database and temporarily stored in the system memory 510. The distance between the two read voice samples is calculated by the CPU 500, and the calculation result is stored in the system memory 510. The area of the system memory 510 for storing the calculation result has a matrix configuration. Then C
The PU 500 reads the second audio sample from the audio database and temporarily stores it in the system memory 510 for distance calculation. The distance between the first voice sample read from the cluster-specific database and the second voice sample read from the voice database is calculated and written to the calculation result storage area of the system memory.

In this way, the distance between the first voice sample in the task-specific database and all voice samples in the voice database is calculated, and the calculation result is written to the system memory 510. After calculating the distance for the first audio sample in the task-specific database,
The CPU 500 calculates the distance between all the audio samples in the audio database and the second audio sample in the task-specific database, and writes the calculation result in the system memory 510.

When such processing is performed, the distance between all voice samples in the task-specific database and all voice samples in the task-specific database is calculated (step 1).
020).

CPU 500 executes a voice sample selection process described later on the distance calculation result stored in system memory 510, and stores the identification ID of the selected voice sample, for example, the storage address of the voice database, in system memory 510 or hard disk 540. Save (step 1030). The storage format can be a document file format (so-called list).

When the processing of selecting the voice sample for one cluster is completed, the CPU 500 switches the cluster to another cluster and calculates the distance between the voice sample in the voice database and the voice sample in the cluster database in the same manner as described above. Perform audio sample selection processing.
(Step 1040 → Steps 1020 to 1030).
When the distance calculation and the selection process of the voice sample are performed for all the cluster-specific databases, the identification ID is obtained from the voice database based on the identification ID of the selected voice sample.
And a system memory 510 for extracting a transcribed text corresponding to and transferring a speech sample (acoustic parameter) to a program for parameter estimation processing.
And the processing procedure of FIG. 6 ends (step 1050).

FIG. 7 shows the details of the audio sample selection process in FIG. 6 when the Abs selection method is employed. In FIG. 7, CPU 500 sequentially reads distance calculation results (matrix form) stored in system memory 510 from the first position (step 1100). The CPU 500 compares the value of the read distance calculation result with the threshold (x described above) in magnitude (step 1110).

If the result of the distance calculation is smaller than the threshold value, that is, if the voice sample in the voice database and the voice sample in the task-specific database are acoustically close, the read voice sample in the voice database and its associated The identification ID of the
0 (step 1120).

On the other hand, if the value of the result of the distance calculation is larger than the threshold value, the procedure returns to step 1100, and the next distance calculation result is read and compared with the threshold value (step 113).
0 → 1100). Hereinafter, for all distance calculation results,
By repeatedly executing the processing of steps 110 to 1130, it is possible to select a voice sample whose distance calculation result value is smaller than the threshold value. The above is the process of selecting the audio sample for one cluster, and the same process is repeated for each cluster.

FIG. 8 shows a processing procedure for selecting a sound sample by the Rt selection method. In FIG. 8, CP
U500 performs a sorting process on all distance calculation results stored in system memory 510 (step 1200). A well-known program may be used for the process of rearranging the values of a plurality of data in ascending order.

Next, the CPU 500 sets the system memory 510
Is obtained, and the number N of audio samples to be selected is calculated by multiplying by a certain ratio r% (step 1210 → 1220).

The CPU 500 retrieves from the system memory 510 the speech sample of the speech database and the identification ID thereof used in the calculation of the value from the smallest value among the distance calculation values arranged in ascending order, and stores them in another area. (Step 1240). At this time, the number I of taking out audio samples is counted (step 1260), and when it is detected in step 1250 that the number N of audio samples to be selected has been reached, the processing procedure of FIG. 8 ends.

The following embodiment can be carried out in addition to the above embodiment.

1) In the above-described embodiment, the speech recognition apparatus that performs speech recognition using both the acoustic model and the language model has been described. However, the present invention can be applied to a speech recognition device that performs speech recognition using only an acoustic model, although the speech recognition accuracy is slightly reduced.

2) In the above-described second and third embodiments, the speech recognition apparatus has the function of clustering speech samples. However, another apparatus performs clustering, and the clustered speech samples are offline or online. Of course, the data may be transferred to the voice recognition device.

3) In the above embodiment, a stand-alone (single) speech recognition apparatus is introduced. However, it is possible to input a speech of a telephone or a speech extracted from a television image as a recognition target.

4) The recording medium referred to in the present invention is not limited to a recording medium such as a floppy disk, CDROM and the like. Any medium that can record (store) a program may be used. For example, an IC memory, a hard disk storage device, or the like can be used as a recording medium. Further, such a recording medium does not need to be installed in the voice recognition device, and a voice recognition program can be downloaded from a storage medium installed in another device to a storage device in the voice recognition device via wireless or wired communication. (Transferring) may of course be performed.

5) Threshold x or R used in Abs selection method
The value of r% used in the t selection method does not need to be fixed, and can be varied according to the type of acoustic model to be created. In this case, either a method of manually inputting a numerical value from the input device 520 or a method of manually inputting only the type of the acoustic model and extracting a value corresponding to the type of the acoustic model from a table or the like may be adopted.

6) In the above-described embodiment, a voice sample of a specific speaker is extracted from the voice samples stored in the task voice database, and is used for selecting a voice sample from the voice database. However, speech samples may be selected from the speech database using other tasks depending on the application of speech recognition, for example, speech samples according to machine characteristics such as speech recognition decoders, recognized speech input devices, and the like.

[0080]

As described above, according to the first and seventh aspects of the present invention, a task, for example, a voice sample close to a voice sample of a specific speaker is selected from voice samples of different speakers. Speech samples that adversely affect the recognition characteristics are excluded from the speech samples used to create the model.

According to the second and eighth aspects of the present invention, by calculating the distance between the first and second voice samples, it is possible to extract as many voice samples as possible for use in the acoustic model.

According to the third and ninth aspects of the present invention, it is possible to maintain the accuracy of the closeness of the first audio sample to be selected by determining whether the result of the distance calculation is within the allowable range.

According to the fourth and tenth aspects of the present invention, it is possible to reduce the influence of the nonlinearity between the distance between speech samples to be used and the speech recognition accuracy.

The fifth, sixth, eleventh and twelfth aspects of the present invention
When there are a large number of audio samples, a more detailed task limitation model can be used by performing clustering. In addition, by using the second voice sample for each task of a specific speaker or the like, it is possible to select a second voice sample of a reference (reference) according to a use from a plurality of second voice samples.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a system configuration according to a first embodiment of the present invention.

FIG. 2 is a block diagram illustrating a circuit configuration of an estimated sample selection unit.

FIG. 3 is a block diagram illustrating another circuit configuration of the estimated sample selection unit.

FIG. 4 is a block diagram showing a system configuration according to a second embodiment of the present invention.

FIG. 5 is a block diagram illustrating a system configuration according to a third embodiment of the present invention.

FIG. 6 is a flowchart illustrating a procedure of a voice sample selection process according to a third embodiment of the present invention.

FIG. 7 is a flowchart showing a processing procedure of audio sample selection processing according to the Abs selection method.

FIG. 8 is a flowchart showing a processing procedure of audio sample selection processing according to an Rt selection method.

[Explanation of symbols]

 Reference Signs List 101 recognition speech input unit 102 sound analysis unit 103 speech recognition decoder 104 language model 105 speech recognition result 106 acoustic model 107 word pronunciation dictionary 108 task speech database 109 speech database 110 estimation sample determination unit 111 parameter estimation unit

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Akio Ando 1-10-11 Kinuta, Setagaya-ku, Tokyo Japan Broadcasting Corporation Broadcasting Research Institute F-term (reference) 5D015 AA03 BB01 GG01 GG03

Claims (12)

[Claims] In a speech recognition for creating an acoustic model from a speech sample and a transcribed text, and performing speech recognition of an input speech based on the created acoustic model, a plurality of first speech samples independent of a speaker and First storage means for storing a plurality of transcribed texts corresponding thereto, second storage means for storing a second voice sample corresponding to a task, and a plurality of storage means for storing a plurality of transcripts stored in the first storage means. Selecting means for selecting a first audio sample closer to the second audio sample recorded in the second storage means from among the first audio samples, corresponding to the selected first audio sample And information processing means for extracting the transcribed text from the first storage means. Speech recognition apparatus characterized by a first audio samples to create the acoustic model. 2. The speech recognition device according to claim 1, wherein said plurality of second speech samples are provided, and said selecting means selects each of said plurality of second speech samples and said plurality of first speech samples. A speech recognition apparatus comprising: calculating a distance between each of them; and selecting the closer first speech sample based on the calculation result. 3. The speech recognition apparatus according to claim 2, wherein said selecting means determines whether or not said calculation result is within an allowable range, and when a result of determination that said calculation result is within an allowable range is obtained. A first speech sample used for the calculation result is selected as a speech sample used for an acoustic model. 4. The speech recognition apparatus according to claim 2, wherein the selecting unit rearranges the calculation results in the order of the shortest distance, and calculates a certain number of the calculation results starting from the shortest distance in the rearranged calculation results. A speech recognition apparatus, wherein a result is selected, and a first speech sample used for the selected calculation result is selected as a speech sample used for an acoustic model. 5. The voice recognition device according to claim 1, wherein clustering is performed on a plurality of voice samples independent of a speaker, and a result of the clustering is stored as the second voice sample in the second storage unit. A speech recognition device, further comprising clustering means. 6. The speech recognition apparatus according to claim 1, wherein the second speech sample corresponding to the task is a speech sample for a specific speaker. Voice recognition device. 7. A storage medium in which an acoustic model is created from a speech sample and a transcribed text and a program executed in speech recognition for performing speech recognition of an input speech based on the created acoustic model is recorded. The apparatus has first storage means for storing a plurality of first voice samples independent of a speaker and a plurality of transcripts corresponding thereto, and second storage for storing a second voice sample corresponding to a task. Means, and wherein the program includes a second sound sample closer to the second sound sample recorded in the second storage means from among the plurality of first sound samples stored in the first storage means. A selection step of selecting one audio sample; and information for extracting a transcript text corresponding to the selected first audio sample from the first storage means. Processing medium, wherein the acoustic model is created from the extracted transcribed text and a first speech sample corresponding to the transcribed text. 8. The recording medium of the speech recognition device according to claim 7, wherein there are a plurality of said second speech samples, and in said selecting step, each of said plurality of second speech samples and said plurality of first speech samples are included. A recording medium for a speech recognition device, comprising: calculating a distance between each of voice samples; and selecting the closer first voice sample based on the calculation result. 9. The recording medium of the speech recognition device according to claim 8, wherein in said selecting step, it is determined whether said calculation result is within an allowable range, and a result of determination that said calculation result is within an allowable range is obtained. A recording medium for a speech recognition apparatus, wherein, when the first speech sample is used, the first speech sample used for the calculation result is selected as a speech sample used for an acoustic model. 10. The recording medium of the speech recognition device according to claim 8, wherein, in the selecting step, the calculation results are rearranged in ascending order of distance, and the calculated results are rearranged from the shortest distance to a constant distance. A recording medium for a speech recognition apparatus, wherein a number of calculation results are selected, and a first voice sample used for the selected calculation result is selected as a voice sample used for an acoustic model. 11. The recording medium of the speech recognition apparatus according to claim 7, wherein clustering is performed on a plurality of speech samples independent of a speaker, and a result of the clustering is used as the second speech sample in the second storage unit. A storage medium for a speech recognition device, further comprising a clustering step of storing the information in a storage device. 12. The method according to claim 7, wherein the second voice sample corresponding to the task is a voice sample for a specific speaker. The recording medium of the speech recognition device, which is a feature.