JP4829871B2 - Learning data selection device, learning data selection method, program and recording medium, acoustic model creation device, acoustic model creation method, program and recording medium - Google Patents

Description

  The present invention relates to selection of data (learning data) used for creating an acoustic model and a technique for creating an acoustic model using the learning data.

  In conventional speech recognition, an acoustic model is created by modeling with a Hidden Markov Model (hereinafter referred to as “HMM”) for each category of speech units such as phonemes, syllables, and words constituting the recognition result candidate. The method has high recognition performance and has become the mainstream of current speech recognition technology.

  An acoustic model typified by an HMM is generated using sufficient statistics learned and accumulated from learning data. In recent years, the amount of learning data has become enormous, and learning data having a learning data amount exceeding 500 hours has been used for creating an acoustic model.

  By the way, creation of an acoustic model requires a learning time corresponding to the amount of learning data. In recent years, with the increase in the amount of learning data, the time cost for acoustic model learning has become enormous. In addition, there is interference data that does not contribute to improvement of recognition performance in the learning data, and the recognition performance may be deteriorated due to the presence of the interference data.

Therefore, in the technique disclosed in Patent Document 1, a high-accuracy acoustic model is created as follows. Using the base acoustic model and a plurality of learning data clusters, sufficient statistics (cluster sufficient statistics) corresponding to each learning data cluster are obtained. Then, an acoustic model is created from one or a plurality of combinations among the sufficient statistics of each cluster, and each acoustic model is evaluated using the evaluation data and the evaluation language model. An acoustic model giving a predetermined evaluation result is selected from the evaluation results of each acoustic model.
JP 2007-249051 A

  Conventionally, the sound that is acoustically close to the scheduled speech recognition target (collecting tasks such as environment, such as tasks, speech styles, speakers, etc. as Merckmar) is collected from scratch or existing Learning data (specifically, voice data and a label based on a voice unit category associated with the voice data) used to create an acoustic model is obtained by manually selecting from a voice database. In order to create a high-accuracy acoustic model, a large amount of learning data corresponding to a task can be collected from scratch, and the speech recognition target scheduled from a speech database with a large amount of learning data can be acoustically Choosing a close voice is very labor intensive.

  Further, as described above, the creation of the acoustic model requires a learning time corresponding to the learning data amount, and the time cost for the acoustic model learning becomes enormous as the learning data amount increases. In addition, the learning data includes interference data that does not contribute to the improvement of the recognition performance, and the recognition performance deteriorates due to the presence of the interference data.

  In the technology of Patent Document 1, in one aspect, a large amount of learning data is subdivided to create a plurality of acoustic models, and a high-accuracy acoustic model is obtained based on these evaluation results using evaluation data and the like. However, since a plurality of acoustic models are created, the calculation cost is high.

  In view of such a problem, the present invention provides a technique for selecting learning data useful for creating a highly accurate acoustic model that realizes high recognition performance from learning data. In addition, a technique for creating a highly accurate acoustic model in a short time is provided.

In order to solve the above-mentioned problem, the present invention selects learning data from learning data including voice data and a label associated with the voice data as follows. An acoustic model (adaptive acoustic model) adapted to a task is created by learning a base acoustic model that is an initial acoustic model from task adaptive learning data that is learning data adapted to the task. Then, speech recognition is performed on the speech data, and the recognition score (adaptive recognition score) in the case of speech recognition using the grammar and adaptive acoustic model obtained from the label of the learning data, and the grammar and base acoustic model are used. A recognition score (base recognition score) in the case of speech recognition is obtained. Of the learning data, one that passes the comparison judgment between the adaptive recognition score and the base recognition score is selected.
In this way, learning data that passes the comparison determination between the adaptive recognition score and the base recognition score is regarded as learning data suitable for the task, and is selected.

  In order to solve the above problems, the present invention creates a sound model by learning a basic sound model that is an initial sound model using selected learning data.

  In addition, the computer can be operated as a learning data selection device by a learning data selection program that causes the computer to function as the learning data selection device of the present invention. Similarly, the computer can be operated as an acoustic model creation device by an acoustic model creation program that causes the computer to function as the acoustic model creation device of the present invention. A computer-readable program recording medium that records such a program makes it possible for another computer to function as a learning data selection device and an acoustic model creation device, or to distribute the program.

  According to the present invention, an adaptive recognition score in the case of speech recognition using an adaptive acoustic model and a base recognition score in the case of speech recognition using a base acoustic model are obtained, Learning data useful for creating a high-accuracy acoustic model can be selected by selecting the one that passes the comparison and determination with the base recognition score. In addition, since the amount of learning data selected is usually smaller than the amount of learning data itself, the basic acoustic model is learned with such selected learning data, thereby achieving high recognition performance. A simple acoustic model can be obtained in a short time.

<< First Embodiment >>
A first embodiment of the present invention will be described with reference to the drawings.
The learning data selection device 1 according to the first embodiment of the present invention is a device that creates an acoustic model using selected learning data (in the first embodiment of the present invention, rather than being independently present alone). It is practical to exist as a component constituting an acoustic model creation device 2). Furthermore, the learning data selection device 1 is not a component constituting the acoustic model creation device 2 so as to be easily separable from the acoustic model creation device 2, but focuses on a certain function of the acoustic model creation device 2 itself. It can also be said that it was evaluated on one side. In short, it is practically practical that the learning data selection device 1 is the acoustic model creation device 2 itself.
However, this does not exclude the fact that the learning data selection device 1 exists as a single independent component and that the learning data selection device 1 is a component constituting the acoustic model creation device 2 so as to be easily separable from the acoustic model creation device 2. . For example, if the purpose is to select learning data itself, there is no obstacle to realizing the learning data selection device 1 as a single independent component.
Here, for example, the acoustic model creation device 2 is realized by a computer such as a dedicated machine configured by dedicated hardware or a general-purpose machine such as a personal computer, and the learning data selection device 1 is realized as a single independent component. The same applies to the case.

  A hardware configuration example will be described in which the acoustic model creation device 2 is a single component and is realized by a computer (general-purpose machine). The learning data selection device 1 is a component constituting the acoustic model creation device 2.

<Hardware configuration example of acoustic model creation device 2>
The acoustic model creation device 2 may include an input unit to which a keyboard or the like can be connected, an output unit to which a liquid crystal display or the like can be connected, a CPU (Central Processing Unit) [cache memory, or the like. ] RAM (Random Access Memory), ROM (Read Only Memory), external storage device as a hard disk, and exchange of data between these input unit, output unit, CPU, RAM, ROM, external storage device It has a bus that can be connected. Further, if necessary, the acoustic model creation device 2 may be provided with a device (drive) that can read and write a storage medium such as a CD-ROM. A physical entity having such hardware resources includes a general-purpose computer.

  The external storage device of the acoustic model creation device 2 stores a program for creating an acoustic model, data necessary for processing of this program, etc. [not limited to the external storage device, for example, a program for read-only storage device It may be stored in a ROM. ]. Data obtained by the processing of these programs is appropriately stored in a RAM or an external storage device. Hereinafter, a storage device that stores data, addresses of storage areas, and the like is simply referred to as a “storage unit”.

  In the present embodiment, learning data 200 is stored as data in a predetermined storage area of the storage unit. The learning data 200 is learning data included in a general-purpose speech database used for speech recognition and the like. This general-purpose voice database can use an existing one, and has a data amount exceeding 500 hours, for example.

  In addition, base learning data 100 is stored as data in a predetermined storage area of the storage unit. The base learning data 100 is learning data used to create a base acoustic model 141 that is an initial acoustic model. The base learning data 100 does not have to be learning data adapted to a task. For example, the base learning model 100 may be the same as the learning data 200 or may be collected from scratch.

  Furthermore, task adaptive learning data 120 is stored as data in a predetermined storage area of the storage unit. The task adaptive learning data 120 is learning data adapted to a task in which an acoustic model is used, and is prepared in advance. For example, when the task is known in advance, the task adaptive learning data 120 can be prepared in advance by collecting the voices in the task from scratch or selecting the learning data to be adapted to the task from the existing voice database. . In addition, even when the voice in the task cannot be collected, learning data obtained by using at least one of the recording environment, the speaker, the utterance content, and the utterance style may be adopted as the task adaptive learning data 120 (described later). See example).

  The task adaptive learning data 120 is not guaranteed to have a sufficient amount of data depending on the task, and the amount of data is often smaller than that of the learning data 200. In the learning data selection technique of the present invention, as will be described later, the selection learning data 131 can be obtained by selecting the learning data appropriate for the task adaptation learning data 120 from the learning data 200. Therefore, the task adaptation learning data 120 is selected. The learning data 131 can be augmented. From this point of view, it is desirable that the learning data 200 includes voice data close to a task.

  The base learning data 100, the task adaptive learning data 120, and the learning data 200 are each utterance units, and voice data (analog data of real voice) and a voice unit category (for example, phonemes, syllables, semi-syllables, etc.) associated with the voice data. ). However, the present invention is not limited to such a configuration. For example, the acoustic analysis result and a label based on a voice unit category associated with the acoustic analysis result may be used, or digitized voice data and It may be composed of a label based on a voice unit category associated with the voice data.

  The storage unit of the acoustic model creation device 2 stores a program for creating the base acoustic model 141, a program for creating the adaptive acoustic model 151, a program for digitizing the learning data 200, and a grammar from the learning data 200. A program for generating, a program for performing speech recognition, a program for selecting learning data based on a speech recognition result, and a program for creating an acoustic model using at least the selected learning data are stored Yes.

  In the acoustic model creation device 2, each program stored in the storage unit and data necessary for processing each program are read into the RAM as necessary, and are interpreted and processed by the CPU. As a result, the CPU realizes predetermined functions (base acoustic model creation unit, adaptive acoustic model creation unit, digitization unit, grammar generation unit, speech recognition unit, learning data selection unit, acoustic model creation unit), and learning data Selection and creation of an acoustic model is realized. The base acoustic model creation unit and the grammar generation unit are not essential components of the acoustic model creation device 2. The learning data selection device 1 of the first embodiment includes a base acoustic model creation unit, an adaptive acoustic model creation unit, a digitization unit, a grammar generation unit, a speech recognition unit, and a learning data selection unit. The base acoustic model creation unit and the grammar generation unit are not essential components of the learning data selection device 1.

  Next, as a first embodiment, a flow of acoustic model creation processing by the acoustic model creation device 2 including learning data selection processing by the learning data selection device 1 will be described descriptively with reference to FIGS. 1 and 2.

  First, the grammar generation unit 16 generates a grammar 161 using a label included in the learning data 200 (step S1). The grammar 161 is created, for example, as a grammar that allows a short pause between words in a label. The process for generating the grammar 161 is not essential, and a description grammar prepared in advance can be used as the grammar 161.

  Next, the base acoustic model creation unit 14 creates a base acoustic model 141 using the base learning data 100 (step S2). The base acoustic model 141 creates, for example, a monophone label by labeling the phonetic data of the base learning data 100 with, for example, a phoneme corresponding to the transcription form (kana) of the transcription text (forcing a short pause). It is assumed that alignment is performed.) A probability model created with a three-state left-to-right type HMM structure. The process of creating the base acoustic model 141 is not essential, and an unlearned acoustic model prepared in advance can be used as the base acoustic model 141.

  Next, the adaptive acoustic model creating unit 15 learns the base acoustic model 141 with the task adaptive learning data 120 and creates an acoustic model (adaptive acoustic model 151) (step S3).

  Next, the digitizing unit 11 converts the speech data included in the learning data 200 into a digital speech signal in units of utterances (step S4). This digitization is performed by applying a known A / D conversion or the like. This digital voice signal is input to the voice recognition unit 12. It should be noted that this process is not necessary when the audio data included in the learning data 200 is a digital audio signal.

  Next, the speech recognition unit 12 receives the digital speech signal, the grammar 161, the base acoustic model 141, and the adaptive acoustic model 151, performs speech recognition in units of utterances, and performs speech recognition using the grammar 161 and the base acoustic model 141. A recognition score (base recognition score) with respect to the digital speech signal based on it and a recognition score (adaptive recognition score) with respect to the digital speech signal based on speech recognition using the grammar 161 and the adaptive acoustic model 151 are obtained (step S5). The calculation method of each recognition score may be based on a known method.

  Next, the learning data selection unit 13 performs a comparison determination between the adaptive recognition score and the base recognition score, and learning data that has passed the determination from the learning data 200 (usually a part) is selected learning data 131. Select (step S6). For example, learning data composed of speech data corresponding to an utterance having an adaptive recognition score equal to or higher than the base recognition score and a label corresponding to the speech data is selected as the selection learning data 131. That is, the selected learning data 131 is a subset of the learning data 200.

  Subsequently, the acoustic model creation unit 17 creates the acoustic model 171 by learning the base acoustic model 141 using the selection learning data 131 (step S7).

  Supplementary explanation will be given for the creation of the acoustic models of the adaptive acoustic model 151 and the acoustic model 171. By replacing the model creation unit of this supplementary explanation with the adaptive acoustic model creation unit 15 and the acoustic model creation unit 17, and by replacing the learning data of this supplementary explanation with the task adaptive learning data 120 and the selection learning data 131, each process can be understood. can get.

  The model creation unit calculates sufficient statistics using the base acoustic model 141 and the learning data.

  An example of learning by the model creation unit is an acoustic analysis of the label-corresponding speech data constituting the learning data, and regarding the acoustic analysis result as an output signal sequence from the state of the HMM, the Baum-Welch algorithm A statistic for calculating the maximum likelihood parameter of the HMM (under the base acoustic model 141) is calculated for each category. This statistic is a sufficient statistic.

  Sufficient statistics are parameters that characterize the HMM, and a specific example will be described. When the probability distribution that gives the relationship between the acoustic feature quantity of the voice data corresponding to the label and the speech unit category is expressed as a mixed normal distribution, this mixed normal distribution is a probability distribution obtained by mixing one or a plurality of multidimensional normal distributions. is there. The multidimensional normal distribution here generally includes the i-th dimensional cepstrum [LPC cepstrum, MFCC (Mel frequency cepstrum coefficient), and the like. The same applies hereinafter. ], Cepstrum coefficients and logarithmic power such as i-th dimension Δ cepstrum (primary difference of cepstrum coefficients), i-th dimension ΔΔ cepstrum (primary difference of Δ cepstrum coefficients), Δ logarithmic power (primary difference of logarithmic power), Each normal distribution is composed of ΔΔ log power (first difference of Δ log power), and each normal distribution is characterized by mean and variance. The mixed normal distribution is generally obtained by weighting and mixing each multidimensional normal distribution. The statistics for calculating the average, variance, mixture weight, and state transition probability of each multidimensional normal distribution listed here are sufficient statistics.

The model creation unit synthesizes an acoustic model from sufficient statistics. Methods for synthesizing acoustic models from sufficient statistics are detailed in the references.
(Reference) Co-authored by Lawrence Rabiner and Biing-Hwang Juang, translated by Sadahiro Furui, “Basics of Speech Recognition (below)”, NTT Advanced Technology, 1995

<< Second Embodiment >>
The second embodiment is a modification regarding selection of the selection learning data 131. In the second embodiment, the process (step S6) of the first embodiment is performed such that the learning data selection unit 13 obtains a score (difference recognition score) obtained by subtracting the base recognition score from the adaptive recognition score in the learning data 200. In this embodiment, the processing is changed to processing (step S6a) for selecting learning data composed of speech data corresponding to an utterance that is equal to or greater than a predetermined threshold and a label corresponding to the speech (step S6a) (see FIG. 3). An example of the first embodiment is a case where the threshold is zero. The threshold value may be a negative value. Except for the changes described here, the second embodiment is the same as the first embodiment.

  When the threshold value is set to be large, it is possible to select the learning data including the speech data corresponding to the utterance close to the task adaptive learning data 120 and the label corresponding thereto as the selection learning data 131. On the other hand, if the threshold value is set small (including a negative case), the data amount of the selection learning data 131 can be increased, although it is not necessarily specialized for a task.

«Third embodiment»
The third embodiment is a modified example related to creation of the acoustic model 171. In the third embodiment, the process of step S7 of the first embodiment is performed by using the learning model in which the acoustic model creation unit 17 uses the combination of the selected learning data 131 and the basic learning data 101 as learning data. This is an embodiment in which the processing is changed to the process of creating the acoustic model 171 (step S7a) (see FIG. 4). Here, as the basic learning data 101, the base learning data 100, the task adaptive learning data 120, or a combination of the base learning data 100 and the task adaptive learning data 120 can be adopted. Other than this change, the second embodiment is the same as the first embodiment. Although not shown, the third embodiment can be applied to the second embodiment.

  By using not only the selected learning data 131 but also the task adaptive learning data 120 as learning data, it is possible to obtain an acoustic model 171 trained with learning data having a sufficient amount of data while being specialized for the task. Further, by using not only the selected learning data 131 but also the base learning data 100 as learning data, the acoustic model 171 trained with learning data having a sufficient amount of data is obtained, although not necessarily specialized for a task. it can.

≪Supplementary notes≫
In the first embodiment, when only the learning data selection process is performed, the process of step S7 can be omitted. Even in this case, the processes in steps S1 and S2 are not essential. The same applies to the second embodiment. In the third embodiment, when only the learning data selection process is performed, the process of step S7a can be omitted. Even in this case, the processes in steps S1 and S2 are not essential.

  In the case where the learning data selection device 1 and the acoustic model creation device 2 are configured as separate devices, the selection learning data 131 output from the learning data selection unit 13 of the learning data selection device 1 (for example, via the recording medium 33). ) It is stored in the storage unit of the acoustic model creation device 2, and the acoustic model creation unit 17 can learn the basic acoustic model 142 from the stored selection learning data 131 to obtain the acoustic model 171 (see FIG. 5). In this case, the basic acoustic model 142 used by the acoustic model creation device 2 is preferably the same as the base acoustic model 141 used in the learning data selection device 1, but it is not always necessary to use the same base acoustic model. Absent.

  Moreover, as in the third embodiment, the acoustic model creation unit 17 uses the learning data as a combination of the selected learning data 131 stored in the storage unit of the acoustic model creation device 2 and the basic learning data 101 as learning data. The acoustic model 171 can be obtained by learning the basic acoustic model 142 (see FIG. 5). In this case, the basic learning data 101 is the base learning data 100 used in the learning data selection device 1, the task adaptive learning data 120, or a combination of the base learning data 100 and the task adaptive learning data 120. It is preferable that the same is used, but it is not always necessary to use the same.

  In addition to the above embodiments, the learning data selection device / method and the acoustic model creation device / method according to the present invention are not limited to the above-described embodiments, and can be appropriately changed without departing from the spirit of the present invention. It is. In addition, the processing described in each embodiment may be executed not only in time series according to the description order, but also in parallel or individually as required by the processing capability of the apparatus that executes the processing. .

  When the processing function in the learning data selection device / acoustic model creation device is realized by a computer, the processing content of the function that the learning data selection device / acoustic model creation device should have is described by a program. Then, by executing this program on a computer, the processing function in the learning data selection device / acoustic model creation device is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto-Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the learning data selection device / acoustic model creation device is configured by executing a predetermined program on the computer. However, at least a part of these processing contents is realized by hardware. It is good as well.

  From the learning data 200 with the data amount of 458.36 hours, the selection learning data 131 with the data amount of 2.36 hours is selected according to the present invention. When the base acoustic model 141 obtained from the base learning data 100 with a data amount of 13.58 hours was used, the recognition rate was 78.37% and the recognition accuracy was 74.49%. An acoustic model obtained by learning the base acoustic model 141 using 58-hour base learning data 100 combined with selected learning data 131 having a data amount of 2.36 hours as learning data (data amount is 15.94 hours). When 171 was used, the recognition rate was 79.10%, and the recognition accuracy was 75.33%, both improving.

  Further, although different from the task adaptive learning data 120, when a free speech having the same utterance style as the task adaptive learning data is set as a recognition target, the recognition rate when the acoustic model 171 is used is 67.60%. The accuracy was 66.10%. In contrast, the recognition rate when using the base acoustic model 141 was 67.43%, and the recognition accuracy was 65.81%. Thus, although different from the task adaptive learning data 120, even when the acoustic model 171 is used when speech of the same utterance style is used as a recognition target, the recognition rate and the recognition accuracy are higher than those of the base acoustic model 141. Results were good.

  The present invention is useful for creating an acoustic model used for speech recognition, for example, character input based on speech recognition or speech recognition of a dialogue system.

The block diagram which shows the function structural example of the learning data selection apparatus and acoustic model creation apparatus concerning 1st Embodiment. The figure which shows the processing flow of the learning data selection process and acoustic model creation process concerning 1st Embodiment. The figure which shows the processing flow of the learning data selection process and acoustic model creation process concerning 2nd Embodiment. The figure which shows the processing flow of the learning data selection process and acoustic model creation process concerning 3rd Embodiment. The block diagram which shows the function structural example at the time of using a learning data selection apparatus and an acoustic model production apparatus as another apparatus, respectively.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Learning data selection apparatus 2 Acoustic model creation apparatus 12 Speech recognition part 13 Learning data selection part 15 Adaptive acoustic model creation part 17 Acoustic model creation part 131 Selection learning data 141 Base acoustic model 151 Adaptive acoustic model 161 Grammar 171 Acoustic model 200 Learning data

Claims (9)

A base acoustic model that is an initial acoustic model, learning data composed of sound data and a label associated with the sound data, a grammar obtained from the label of the learning data, and learning data adapted to the task (task Storage means for storing adaptive learning data),
Adaptive acoustic model creating means for learning the base acoustic model from the task adaptive learning data and creating an acoustic model adapted to the task (adaptive acoustic model);
Speech recognition is performed on the speech data, and a recognition score (adaptive recognition score) in speech recognition using the grammar and the adaptive acoustic model, and speech recognition using the grammar and the base acoustic model. Speech recognition means for obtaining a recognition score (base recognition score);
A learning data selecting device comprising learning data selecting means for selecting, from among the learning data, a score obtained by subtracting the base recognition score from the adaptive recognition score is a predetermined threshold value or more . The storage means is adapted to the base acoustic model, which is the initial acoustic model, learning data composed of speech data and a label associated with the speech data, the grammar obtained from the label of the learning data, and the task Learning data (task adaptive learning data) to be stored,
An adaptive acoustic model creating step of learning the base acoustic model from the task adaptive learning data and creating an acoustic model adapted to the task (adaptive acoustic model);
Speech recognition is performed on the speech data, and a recognition score (adaptive recognition score) in speech recognition using the grammar and the adaptive acoustic model, and speech recognition using the grammar and the base acoustic model. A speech recognition step for obtaining a recognition score (base recognition score);
A learning data selection method comprising: a learning data selection step of selecting, from among the learning data, a score obtained by subtracting the base recognition score from the adaptive recognition score is equal to or greater than a predetermined threshold . Adapts to the basic acoustic model and the base acoustic model , which are initial acoustic models, learning data composed of speech data and a label associated with the speech data, grammar obtained from the label of this training data, and task Storage means for storing learning data (task adaptive learning data) ;
Adaptive acoustic model creating means for learning the base acoustic model from the task adaptive learning data and creating an acoustic model adapted to the task (adaptive acoustic model);
Speech recognition is performed on the speech data, and a recognition score (adaptive recognition score) in speech recognition using the grammar and the adaptive acoustic model, and speech recognition using the grammar and the base acoustic model. Speech recognition means for obtaining a recognition score (base recognition score);
Learning data selection means for selecting, from among the learning data, a score obtained by subtracting the base recognition score from the adaptive recognition score (selection learning data) equal to or greater than a predetermined threshold;
An acoustic model creation device comprising acoustic model creation means for learning the basic acoustic model using the selection learning data and creating an acoustic model. Upper Symbol acoustic model creating means, those combined with the selected learning data different from the basic learned data and the selected learning data as learning data, creating an acoustic model in the learning data to learn the basic acoustic model The acoustic model creation apparatus according to claim 3 . The storage means includes a basic acoustic model and a base acoustic model , which are initial acoustic models, learning data including speech data and a label associated with the speech data, and a grammar obtained from the label of the learning data. , Learning data to be adapted to the task (task adaptive learning data) is stored,
An adaptive acoustic model creating step of learning the base acoustic model from the task adaptive learning data and creating an acoustic model adapted to the task (adaptive acoustic model);
Speech recognition is performed on the speech data, and a recognition score (adaptive recognition score) in speech recognition using the grammar and the adaptive acoustic model, and speech recognition using the grammar and the base acoustic model. A speech recognition step for obtaining a recognition score (base recognition score);
A learning data selection step for selecting, from among the learning data, a score obtained by subtracting the base recognition score from the adaptive recognition score being equal to or greater than a predetermined threshold (selection learning data);
An acoustic model creation method including an acoustic model creation step of creating an acoustic model by learning the basic acoustic model using the selection learning data. A program for causing a computer to function as the learning data selection device according to claim 1 . A computer-readable recording medium on which the program according to claim 6 is recorded. Program for causing a computer to function as the acoustic model generating apparatus according to claim 3 or claim 4. A computer-readable recording medium on which the program according to claim 8 is recorded.

Images