JP2007248730A - Sound model adaptive apparatus, method, and program, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To construct a sound model with high accuracy, by easily selecting a voice recognition result with high accuracy, suitable for adaptation without a teacher of the sound model, and by using the selected voice recognition result. <P>SOLUTION: A reliability degree imparting section 150 calculates a reliability degree which is the estimation value of a recognition rate for each utterance sequence in which the word sequence of the voice recognition result is divided by using the voice recognition result. An utterance selecting section 160 selects an utterance sequence which is used for adaptation of the sound model by using the recognition rate of the sound model and the reliability degree for each utterance sequence. A sound model adaptive section 170 performs adaptation of the sound model by using the utterance sequence selected by the utterance selecting section 160 and a featured value corresponding to the utterance sequence. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a technique for adapting an acoustic model, and more particularly, to a technique for performing unsupervised adaptation of an acoustic model using reliability of a speech recognition result.

  In general, in speech recognition, an acoustic model is adapted using a speech file and correct text representing the utterance content of the speech file as learning data. “Adaptation of the acoustic model” means a process of optimizing the parameters of the acoustic model so that as many cases as possible in the learning data are established by the learning process. In addition, the adaptation of the acoustic model is performed as supervised adaptation using correct text created as a learning data by human writing and the like corresponding to an audio file, and learning data using correct text as a speech recognition result. It is roughly divided into the unsupervised adaptation used.

Here, when the acoustic model is adapted by unsupervised adaptation, it is necessary to use a speech recognition result with high recognition accuracy as the correct text. This is because if the speech recognition result with low recognition accuracy is used as the correct text, the accuracy of the acoustic model may be reduced due to incorrect adaptation of the acoustic model.
For such a problem, a method may be considered in which reliability is given to the speech recognition result, the speech recognition result is selected according to the high reliability, and the acoustic model is adapted using the selected speech recognition result. . Thereby, it can avoid that the speech recognition result with low recognition accuracy is used as a correct text, and the accuracy of an acoustic model falls.

For example, Non-Patent Document 1 discloses a technique for assigning a reliability based on a phoneme posterior probability to a speech recognition result and adapting an acoustic model using the speech recognition result to which a reliability equal to or higher than a threshold is assigned. Has been. In this method, a threshold value is set in a range from 0 to 1, a plurality of data selection models having different values as threshold values are prepared, and an acoustic model is adapted and evaluated.
Satoshi Ogata, Yasuo Ariki, “Unsupervised adaptation of acoustic models using reliability based on phoneme posterior probabilities”, IEICE Technical Report NLC2001-70, pp.19-24

However, in the existing technology such as Non-Patent Document 1, it is very difficult to determine which value should be set as a threshold value and to select a speech recognition result to be used for acoustic model adaptation. .
The present invention has been made in view of such points, and easily selects a highly accurate speech recognition result suitable for unsupervised adaptation of an acoustic model, and uses the selected speech recognition result to provide a highly accurate acoustic model. The purpose is to provide technology that can build

  In the present invention, in order to solve the above-described problem, the reliability providing unit calculates the reliability that is the estimated value of the recognition rate for each utterance sequence obtained by dividing the word sequence of the speech recognition result using the speech recognition result. The utterance selection unit uses the recognition rate of the acoustic model and the reliability of each utterance sequence to select an utterance sequence to be used for adaptation of the acoustic model, and the acoustic model adaptation unit selects the utterance sequence selected by the utterance selection unit and the utterance sequence The acoustic model is adapted using the feature quantity corresponding to the utterance series. The “utterance sequence” means each sequence obtained by dividing a word sequence (reading word sequence) of a speech recognition result according to a predetermined standard. Further, the “utterance series” is composed of one or more words. The “reliability” is an estimated value of the recognition rate, and this includes not only a value that estimates the recognition rate itself (for example, α when the recognition rate is estimated as α%) but also the recognition rate. A concept that includes a value that estimates a range (for example, α when the recognition rate is estimated to be α% or more, and α and β when the recognition rate is estimated to be α% or more and less than β%). is there.

Here, in the present invention, the reliability of the utterance sequence is evaluated with reference to the recognition rate of the acoustic model, and the utterance sequence used for adaptation of the acoustic model is selected. As a result, it is possible to prevent an utterance sequence that would reduce the recognition rate of the acoustic model from being selected. In addition, since the selection of the utterance sequence used for the adaptation of the acoustic model is performed based on the recognition rate of the acoustic model, there is no need for trial and error for setting an appropriate threshold as in the prior art.
Preferably, in the present invention, the utterance selection unit compares the reference value set to a value equal to or higher than the recognition rate of the acoustic model and the reliability for each utterance sequence, and the utterance sequence whose reliability is equal to or higher than the reference value. Or an utterance sequence whose reliability exceeds the reference value is selected.

By selecting the utterance sequence in this way, it is possible to prevent the selection of the utterance sequence that reduces the recognition rate of the acoustic model due to adaptation.
Preferably, in the present invention, the supervised correct text is input to the adaptive data input unit, and the acoustic model adaptation unit selects the utterance sequence selected by the utterance selection unit, the feature amount corresponding to the utterance sequence, and the adaptive data. The acoustic model is adapted using the supervised correct text input to the input unit and the feature amount corresponding to the supervised correct text. The “supervised correct answer text” means a correct answer text created or corrected by a person writing up a reading corresponding to an audio file. Preferably, the correct text selection unit selects a supervised correct text corresponding to at least a part of an utterance sequence that is not selected by the utterance selection unit, and the correct text output unit selects the selected supervised correct text. Output. Preferably, the supervised correct text input to the adaptive data input unit is a supervised correct text output from the correct text output unit. By replacing the utterance sequence with low reliability in this way with the supervised correct text and performing model adaptation, the accuracy of the acoustic model can be further improved while maintaining the advantages of unsupervised adaptation.

  More preferably, the correct text selection unit selects a supervised correct text that corresponds to an utterance sequence that has not been selected by the utterance selection unit and that has a reliability that satisfies a predetermined criterion. As a result, it is possible to prevent an audio file having extremely low reliability and possibly having a problem with the data itself from being used for adaptation of the acoustic model and adversely affecting the accuracy of the acoustic model.

  As described above, according to the present invention, it is possible to easily select a highly accurate speech recognition result suitable for unsupervised adaptation of an acoustic model, and to construct a highly accurate acoustic model using the selected speech recognition result. Become.

The best mode for carrying out the present invention will be described below with reference to the drawings.
[First Embodiment]
<Hardware configuration>
FIG. 1 is a block diagram illustrating a hardware configuration of an acoustic model adaptation device 1 according to the first embodiment.
As illustrated in FIG. 1, the acoustic model adaptation device 1 of this example includes a CPU (Central Processing Unit) 11, an input unit 12, an output unit 13, an auxiliary storage device 14, a ROM (Read Only Memory) 15, and a RAM (Random Access Memory) 16 and a bus 17.

  The CPU 11 in this example includes a control unit 11a, a calculation unit 11b, and a register 11c, and executes various calculation processes according to various programs read into the register 11c. In this example, the input unit 12 is an input port for inputting data, a keyboard, a mouse, and the like, and the output unit 13 is an output port for outputting data, a display, and the like. The auxiliary storage device 14 is, for example, a hard disk, an MO (Magneto-Optical disc), a semiconductor memory, or the like, and stores various data such as a program area 14a storing a program for executing the processing of this embodiment and tag output information. The data area 14b is provided. The RAM 16 is, for example, an SRAM (Static Random Access Memory), a DRAM (Dynamic Random Access Memory), or the like, and has a program area 16a in which the above program is written and a data area 16b in which various data are written. The bus 17 in this example connects the CPU 11, the input unit 12, the output unit 13, the auxiliary storage device 14, the ROM 15 and the RAM 16 so that data can be exchanged.

<Cooperation between hardware and software>
The CPU 11 in this example writes a program stored in the program area 14 a of the auxiliary storage device 14 in the program area 16 a of the RAM 16 in accordance with the read OS (Operating System) program. Similarly, the CPU 11 writes various data stored in the data area 14 b of the auxiliary storage device 14 into the data area 16 b of the RAM 16. Further, the CPU 11 stores the address on the RAM 16 where the program and various data are written in the register 11c. Then, the control unit 11a of the CPU 11 sequentially reads these addresses stored in the register 11c, reads a program and data from the area on the RAM 16 indicated by the read address, and sequentially executes the calculation indicated by the program to the calculation unit 11b. The calculation result is stored in the register 11c.

FIG. 2 is an example of a block diagram of the acoustic model adaptation apparatus 1 configured by reading the program into the CPU 11 as described above. The arrows in FIG. 2 indicate the flow of data, but the description of the flow of data input to and output from the control unit 190 is omitted.
As shown in FIG. 2, the acoustic model adaptation apparatus 1 according to the present embodiment includes a memory 110, a speech recognition result input unit 130, an information conversion unit 140, a reliability assignment unit 150, an utterance selection unit 160, an acoustic model adaptation unit 170, a temporary A memory 180 and a control unit 190 are included. Here, the memory 110 has storage units 111 to 119 for storing various data. In addition, the reliability assigning unit 150 includes a feature vector generation unit 151 and a feature vector evaluation unit 152. Note that the memory 110 and the temporary memory 180 correspond to, for example, the register 11c, the auxiliary storage device 14, the RAM 16, or a storage area obtained by combining at least a part of them as illustrated in FIG. In addition, the information conversion unit 140, the reliability assignment unit 150, the utterance selection unit 160, the acoustic model adaptation unit 170, and the control unit 190 are configured, for example, by reading a program into the CPU 11 illustrated in FIG. . Furthermore, the speech recognition result input unit 130 is, for example, the input unit 12 that operates under the control of the CPU 11 into which a program has been read. In addition, the acoustic model adaptation device 1 executes each process under the control of the control unit 190. Unless otherwise specified, the data of each process is read from and written to the temporary memory 180 one by one.

<Processing>
Next, the process of the acoustic model adaptation apparatus 1 of this form is demonstrated.
FIG. 3 is a flowchart for explaining processing of the acoustic model adaptation apparatus 1 according to the first embodiment. FIG. 4 is a flowchart for explaining details of the process in step S3 in FIG. Hereinafter, the processing of this embodiment will be described with reference to these drawings.
[Preprocessing]
As preprocessing, the identification model is stored in the storage unit 114 of the memory 110, the audio file is stored in the storage unit 118, the acoustic model is stored in the storage unit 119, and the acoustic model is recognized in the storage unit 116 (corresponding to the “recognition rate storage unit”). Each rate is stored. The identification model means a model for obtaining an estimated value (reliability) of a recognition rate using a feature amount obtained from a speech recognition result (details will be described later). The acoustic model is a model that expresses the statistical properties of speech, and examples thereof include a hidden Markov model (HMM). The recognition rate of the acoustic model is obtained by performing speech recognition of actual evaluation data using the acoustic model and calculating the recognition rate.

[Acoustic model adaptation processing]
The acoustic model adaptation process is executed on the premise of the preprocessing as described above.
First, a voice recognition unit (not shown) performs voice recognition of a voice file stored in the storage unit 118 using an acoustic model stored in the storage unit 119 of the memory 110. The voice recognition result is input to the voice recognition result input unit 130, and is associated with each corresponding voice file and stored in the storage unit 111 of the memory 110 (step S1). The speech recognition result includes a reading word sequence obtained by speech recognition and additional information given to each word by speech recognition (for example, part-of-speech information, acoustic likelihood score, language likelihood score, Word likelihood score, word duration, phoneme number, phoneme duration, etc.).

Next, the information conversion unit 140 reads the speech recognition result from the storage unit 111 of the memory 110, classifies the word sequence of the speech recognition result for each utterance sequence based on a certain standard, and obtains each word sequence obtained Are stored in the storage unit 112 of the memory 110 in association with each additional information of each voice file and voice recognition result (step S2). The definition of “utterance sequence” is as described above. Examples of criteria for dividing a word sequence include the length of a silent interval between words and the part of speech information of a word. A specific example of the utterance sequence is as follows.
“I think there will be a lot of profits in that area. ]
"I see. ]
“Now that tour to go shopping in Korea is very popular,”
"Hmm"
Next, the reliability assigning unit 150 calculates the reliability that is the estimated value of the recognition rate for each utterance sequence using the speech recognition result. Each calculated reliability is associated with the corresponding utterance sequence and stored in the storage unit 115 of the memory 110 (step S3). Note that the “reliability” is an estimated value of the recognition rate, but this includes not only a value that estimates the recognition rate itself (for example, α when the recognition rate is estimated as α%) but also the recognition rate. A concept that includes a value that estimates a range (for example, α when the recognition rate is estimated to be α% or more, and α and β when the recognition rate is estimated to be α% or more and less than β%). is there. Details of this processing will be described later.

  Next, the utterance selection unit 160 reads the reliability of each utterance sequence from the storage unit 115 of the memory 110 and reads the recognition rate of the acoustic model from the storage unit 116. And the utterance selection part 160 selects the utterance series used for adaptation of an acoustic model using these, and stores the selection information which shows the selection content in the storage part 117 (step S4). Preferably, the utterance selection unit 160 compares a reference value set based on the recognition rate of the acoustic model with the reliability of each utterance sequence, and selects an utterance sequence whose reliability is equal to or higher than the reference value, An utterance sequence whose degree exceeds the reference value is selected. More preferably, the reference value is a value set to be equal to or higher than the acoustic model recognition rate. Specifically, for example, an utterance sequence is selected as follows.

[Example 1]
The reference value is the recognition rate of the acoustic model, and an utterance sequence whose reliability is equal to or higher than the recognition rate of the acoustic model is selected, or an utterance sequence whose reliability exceeds the recognition rate of the acoustic model is selected.
[Example 2]
A value obtained by adding or multiplying the recognition rate of the acoustic model by a constant is used as a reference value, and an utterance sequence having a reliability equal to or higher than the reference value is selected, or an utterance sequence having a reliability exceeding the reference value is selected.
[Example 3]
A value obtained by subtracting a constant from the recognition rate of the acoustic model is used as a reference value, and an utterance sequence having a reliability higher than the reference value is selected, or an utterance sequence having a reliability higher than the reference value is selected.
[Example 4]
A function value obtained by substituting the recognition rate of the acoustic model into a predetermined function is used as a reference value, and an utterance sequence having a reliability greater than or equal to the reference value is selected, or an utterance sequence having a reliability higher than the reference value is selected.

  Next, the acoustic model adaptation unit 170 reads the selection information from the storage unit 117 of the memory 110 and specifies the utterance sequence selected by the utterance selection unit 160 using the selection information. After that, the acoustic model adaptation unit 170 reads the identified utterance sequence from the storage unit 112 and reads an audio file corresponding to the read utterance sequence from the storage unit 118. Then, the acoustic model adaptation unit 170 adapts the acoustic model using the existing acoustic model adaptation method using the feature amount and the utterance sequence of the read voice file (step S5). At this time, the utterance sequence functions as an unsupervised correct text. The acoustic model adaptation method is not limited. For example, the Baum-Weltch algorithm may be used. However, the adaptive accuracy can be improved by selecting the optimal acoustic model adaptation method according to the amount of data. Can be improved. The acoustic model that has been adapted in this way is stored in the storage unit 119 of the memory 110.

[Details of Step S3 Processing]
Next, details of the processing in step S3 described above will be described.
First, the feature vector generation unit 151 of the reliability assigning unit 150 reads one utterance series from the storage unit 112 of the memory 110 and stores it in the temporary memory 180 (step S11). Next, the feature vector generation unit 151 reads the utterance sequence from the temporary memory 180 and reads additional information associated with the utterance sequence from the storage unit 111. Then, the feature quantity vector generation unit 151 generates a feature quantity vector for each utterance sequence using the read additional information, and stores this in the storage unit 113 in association with the utterance sequence (step S12). For each element of the feature vector, information useful for estimating the recognition rate by the feature vector evaluation unit 152 is used among the additional information. For example, all or part of part-of-speech information, acoustic likelihood score, language likelihood score, word likelihood score, word duration length, phoneme number, phoneme duration length of each word included in the utterance series is a feature vector. Element.

FIG. 5 is a conceptual diagram illustrating the configuration of the feature quantity vector 200 generated in this way.
The feature quantity vector 200 in the example of FIG. 5 includes part-of-speech information 210, an acoustic likelihood score 220,..., And a phoneme duration length 230. Here, the part-of-speech information 210 is a feature amount representing a plurality of words included in the utterance series by one symbol. The part-of-speech information 210 in the example of FIG. 5 includes m elements (0 or 1) corresponding to each part-of-speech 211-1 to m. The element corresponding to the part of speech of the word included in the utterance series is set to 1, and the elements corresponding to other parts of speech are set to 0. In addition, the acoustic likelihood score 220,..., Phoneme duration length 230 in the example of FIG. 5 is the acoustic likelihood score assigned to each word included in the utterance series,..., Statistical information for each phoneme duration length (this In the example, average values 221 and 231, variance values 222 and 232, maximum values 223 and 233, minimum values 224 and 234) are normalized from 0 to 1 (S1 to S4,..., S5 to S8), respectively. Become. For example, part-of-speech information having 37 types (m = 37) of parts-of-speech, and utterances for each of acoustic likelihood score, language likelihood score, word likelihood score, word duration length, phoneme number, and phoneme duration length When a feature vector is constituted by the average, variance, maximum, and minimum elements for each series, the feature vector has 61 {= 37 + (6 × 4)} dimensions. Note that the feature amount vector is not limited to the configuration illustrated in FIG. 5 as long as it is information obtained by converting information in units of words into units of utterance sequences.

Next, the feature vector evaluation unit 152 reads a feature vector from the storage unit 113 of the memory 110 and reads an identification model from the storage unit 114. Then, the feature vector evaluation unit 152 performs statistical evaluation using the feature vector and the identification model, and calculates the reliability (estimated value of the recognition rate) of the utterance sequence corresponding to the feature vector. The calculated reliability is stored in the storage unit 115 of the memory 110 in association with the corresponding utterance sequence (step S13). Hereinafter, details of the process of step S13 will be exemplified.
[Details of processing in step S13]
First, the identification model will be described. The identification model of this embodiment is a model for obtaining the reliability of a corresponding utterance sequence using a feature vector. That is, by substituting each element of the feature vector into the identification model, information for specifying the reliability of the corresponding utterance sequence can be calculated. Such an identification model is generated using learning data (comprising a feature vector and information for specifying the reliability of an utterance sequence). In other words, model parameters are set so that more cases can be established in the learning data by learning, and an identification model is configured. Examples of such an identification model include those based on machine learning such as SVM (support vector machine) and boosting, those based on a probability model such as a maximum likelihood estimation method and a maximum entropy method, and those based on a neural network.

Normally, when the number of dimensions of the feature vector is very large, learning of a statistical identification model requires a large amount of learning data, and if there is little learning data, an overlearning problem often occurs. On the other hand, since SVM automatically selects only a small number of learning samples near the identification plane based on the criterion of “maximizing margin” to form the identification plane, relatively good identification performance can be obtained even with a small amount of learning data. . For this reason, SVM is suitable for the present invention.
The identification model based on SVM is a model that performs two-class pattern recognition of whether or not the recognition rate for an input feature vector is equal to or greater than a threshold value (n%). Such an identification model is generated by preparing learning data (a feature vector whose class membership is known) in advance and learning a probabilistic correspondence between the feature vector and the class. Further, the identification model based on the SVM can only be estimated whether or not the recognition rate for the feature vector is equal to or higher than a threshold value (n%). Therefore, such an identification model is created with a required density in the range of 0 ≦ n ≦ 100. For example, when it is necessary to determine to which range the estimated value of the recognition rate belongs with an accuracy of 10% intervals (for example, the estimated value of the recognition rate is 70 to 80%), the 11 identification models (n = 0,10, ..., 100) must be created. On the other hand, when only the information indicating whether or not the estimated value of the recognition rate is n% or more (for example, whether or not the estimated value of the recognition rate is 70% or more), one identification model (n = 70) need only be created (end of description of [Details of processing in step S13]).

Next, the control unit 190 refers to the utterance sequences and the reliability stored in the storage units 112 and 115 of the memory 110, and determines whether or not the reliability of all the utterance sequences has been calculated (step S14). ). Here, when the reliability of all the utterance sequences has not been calculated, the control unit 190 returns the process to step S11. On the other hand, when the reliability of all the utterance sequences has been calculated, the control unit 190 ends the process of step S3 (end of description of [details of process of step S3]).
[Second Embodiment]
Next, a second embodiment of the present invention will be described.

The second embodiment is a modification of the first embodiment, in which an acoustic model adaptation is performed using a supervised correct text for an utterance sequence with low reliability. Below, it demonstrates centering around difference with 1st Embodiment, and abbreviate | omits description about the matter which is common in 1st Embodiment.
<Configuration>
FIG. 6 is an example of a block diagram of an acoustic model adaptation apparatus 301 configured by reading a predetermined program into a known computer similar to the first embodiment. The arrows in FIG. 6 indicate the flow of data, but the description of the flow of data input to and output from the control unit 190 is omitted. Further, in FIG. 6, portions common to FIG. 2 are denoted by the same reference numerals as those in FIG. 2 to simplify the description.

  As shown in FIG. 6, the acoustic model adaptation apparatus 301 of this embodiment includes a memory 110, a speech recognition result input unit 130, an information conversion unit 140, a reliability assignment unit 150, an utterance selection unit 160, an acoustic model adaptation unit 170, a temporary A memory 180, a control unit 190, a correct text selection unit 330, a correct text output unit 340, and an adaptive data input unit 350 are provided. Here, the memory 110 includes storage units 311 and 312 in addition to the storage units 111 to 119 for storing various data. The correct text selection unit 330 is configured by reading a program into the CPU 11 of FIG. The correct text output unit 340 and the adaptive data input unit 350 are configured by, for example, a program being read into the CPU 11 in FIG. 1 or an output unit that operates under the control of the CPU 11 that has read the program. 13 and the input unit 12. The acoustic model adaptation apparatus 301 executes each process under the control of the control unit 190. Unless otherwise specified, the data of each process is read from and written to the temporary memory 180 one by one.

<Processing>
Next, the process of the acoustic model adaptation apparatus 301 of this form is demonstrated.
FIG. 7 is a flowchart for explaining the processing of the acoustic model adaptation apparatus 301 in the second embodiment. Hereinafter, the processing of this embodiment will be described with reference to FIG.
[Preprocessing]
As preprocessing, an identification model is stored in the storage unit 114 of the memory 110, an audio file is stored in the storage unit 118, an acoustic model is stored in the storage unit 119, and a recognition rate of the acoustic model is stored in the storage unit 116. The storage unit 311 stores a supervised correct text file that is a set of supervised correct texts corresponding to the audio files stored in the storage unit 118.

[Acoustic model adaptation processing]
The acoustic model adaptation process is executed on the premise of the preprocessing as described above.
Steps S21 to S24 are the same as steps 1 to S4 of the first embodiment. That is, first, a speech recognition result is input to the speech recognition result input unit 130, associated with each corresponding speech file, and stored in the storage unit 111 of the memory 110 (step S21). Next, the information conversion unit 140 classifies the word sequence of the speech recognition result for each utterance sequence based on a certain criterion, and associates each obtained word sequence with each additional information of the speech file and the speech recognition result. And stored in the storage unit 112 of the memory 110 (step S22). Then, the reliability level assigning unit 150 uses the speech recognition result to calculate the reliability level that is the estimated value of the recognition rate for each utterance sequence, and associates each calculated reliability level with the corresponding utterance sequence, 110 is stored in the storage unit 115 (step S23). Next, the utterance selection unit 160 selects the utterance sequence used for adaptation of the acoustic model using the reliability for each utterance sequence and the recognition rate of the acoustic model, and stores selection information indicating the selection contents in the storage unit 117. (Step S24).

Next, the correct text selection unit 330 reads each selection information from the storage unit 117 of the memory 110 and reads each reliability from the storage unit 115. Then, the correct text selection unit 330 stores a supervised correct text corresponding to an utterance sequence that has not been selected by the utterance selection unit 160 and has a reliability that satisfies a predetermined criterion. Selection is made from 311 supervised correct text files (step S25). Note that “an utterance sequence whose reliability is satisfactory enough to satisfy a predetermined criterion” is selected as follows, for example.
[Example 1]
The utterance sequences not selected by the utterance selection unit 160 are rearranged in the order of high reliability, and a predetermined number of utterance sequences are selected in descending order of reliability.

[Example 2]
The threshold value is a value smaller than the reference value used by the utterance selection unit 160, and an utterance sequence having a reliability higher than the threshold value is selected.
The selected supervised correct text is output from the correct text output unit 340 and stored in the storage unit 312 of the memory 110. Next, the supervised correct text stored in the storage unit 312 is input to the adaptive data input unit 350 and sent to the acoustic model adaptation unit 170. The acoustic model adaptation unit 170 reads an audio file corresponding to the sent supervised correct text from the storage unit 118.

  Furthermore, the acoustic model adaptation unit 170 reads the selection information from the storage unit 117 of the memory 110, identifies the utterance sequence selected by the utterance selection unit 160 using the selection information, reads the identified utterance sequence from the storage unit 112, An audio file corresponding to the read utterance series is read from the storage unit 118. Then, the acoustic model adaptation unit 170 uses the feature amount, utterance sequence, and supervised correct text of the read audio file (that is, the utterance sequence selected by the utterance selection unit 160 and the feature amount corresponding to the utterance sequence, and Then, the acoustic model is adapted (using the supervised correct text input to the adaptive data input unit 350 and the feature amount corresponding to the supervised correct text) (step S26). The acoustic model that has been adapted in this way is stored in the storage unit 119 of the memory 110.

[Modifications, etc.]
The present invention is not limited to the embodiment described above. For example, in the above-described embodiment, the acoustic model adaptation apparatus is configured by reading a program into one computer. However, each function of the acoustic model apparatus may be distributed to a plurality of computers and CPUs. Good. For example, the correct text selection unit 330 in the second embodiment may be realized by another computer (separate apparatus), or a plurality of correct text selection units 330 each configured by a plurality of computers may be used. Good. The supervised correct text selected by another apparatus is input from an adaptive data input unit 350 (in this case, corresponding to the input unit 12 operating under the control of the CPU 11 into which the program has been read).

Further, in each of the above-described embodiments, the audio file is stored in the storage unit 118, and the acoustic model adaptation unit 170 extracts the feature amount from the acoustic file and adapts the acoustic model. However, the storage unit 118 may store the feature quantity itself, and the acoustic model adaptation unit 170 may directly use the feature quantity read from the storage unit 118.
In the second embodiment described above, the correct text selection unit 330 is an utterance sequence that is not selected by the utterance selection unit 160, and the utterance sequence is satisfactory only if the reliability satisfies a predetermined criterion. The corresponding supervised correct text was selected. However, the correct text selection unit 330 may arbitrarily select the supervised correct text corresponding to at least a part of the utterance series that the utterance selection unit 160 did not select. Furthermore, the correct text selection unit 330 may arbitrarily select the correct text with the teacher regardless of the selection contents of the utterance selection unit 160.

In addition, speech recognition is performed using an acoustic model that has been adapted as in each of the above-described embodiments, the speech recognition result is input to the speech recognition result input unit 130 again, and similar processing is repeated. Also good. Thereby, highly accurate model adaptation is attained.
In addition, the various processes described above are not only executed in time series according to the description, but may be executed in parallel or individually according to the processing capability of the apparatus that executes the processes or as necessary. Needless to say, other modifications are possible without departing from the spirit of the present invention.
Further, when the above-described configuration is realized by a computer, processing contents of functions that each device should have are described by a program. The processing functions are realized on the computer by executing the program on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. The computer-readable recording medium may be any medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, or a semiconductor memory. Specifically, for example, the magnetic recording device may be a hard disk device or a flexible Discs, magnetic tapes, etc. as optical discs, DVD (Digital Versatile Disc), DVD-RAM (Random Access Memory), CD-ROM (Compact Disc Read Only Memory), CD-R (Recordable) / RW (ReWritable), etc. As the magneto-optical recording medium, MO (Magneto-Optical disc) or the like can be used, and as the semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory) or the like 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.
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 each embodiment, the apparatus is configured by executing a predetermined program on a computer. However, at least a part of the processing contents may be realized by hardware.

  As an industrial application field of the present invention, for example, a voice dialogue system in which a computer and a person communicate by voice dialogue can be exemplified. In such a spoken dialogue system, a computer collects, selects and learns speech while interacting with humans, and performs sequential autonomous adaptation. According to the present invention, since a highly accurate acoustic model can be constructed easily and efficiently with a small amount of adaptation data, that is, with a short adaptation time, a highly accurate spoken dialogue system can be configured easily.

FIG. 1 is a block diagram illustrating a hardware configuration of the acoustic model adaptation apparatus according to the first embodiment. FIG. 2 is an example of a block diagram of the acoustic model adaptation apparatus according to the first embodiment. FIG. 3 is a flowchart for explaining processing of the acoustic model adaptation apparatus according to the first embodiment. FIG. 4 is a flowchart for explaining details of the processing in step S3 in FIG. FIG. 5 is a conceptual diagram illustrating the configuration of the feature vector. FIG. 6 is an illustration of a block diagram of the acoustic model adaptation apparatus in the second exemplary embodiment. FIG. 7 is a flowchart for explaining the process of the acoustic model adaptation apparatus according to the second embodiment.

Explanation of symbols

1,301 Acoustic model adaptation device

Claims (9)

An acoustic model adaptation device for adapting an acoustic model,
A recognition rate storage for storing the recognition rate of the acoustic model;
A speech recognition result input unit for inputting a speech recognition result using the acoustic model;
Using the speech recognition result, for each utterance sequence obtained by dividing the word sequence of the speech recognition result, a reliability providing unit that calculates a reliability that is an estimated value of the recognition rate;
An utterance selection unit that selects an utterance sequence to be used for adaptation of the acoustic model, using the recognition rate of the acoustic model and the reliability of each utterance sequence;
An acoustic model adaptation unit that adapts the acoustic model using the utterance sequence selected by the utterance selection unit and a feature amount corresponding to the utterance sequence;
An acoustic model adaptation device characterized by comprising: The acoustic model adaptation device according to claim 1,
The utterance selection unit
Compare the reference value set to a value equal to or higher than the recognition rate of the acoustic model and the reliability for each utterance sequence, and select an utterance sequence whose reliability is equal to or higher than the reference value, or the reliability is the reference value Select an utterance sequence that exceeds
An acoustic model adaptation device characterized by that. The acoustic model adaptation device according to claim 1,
An adaptive data input unit for inputting a supervised correct text;
The acoustic model adaptation unit is
Using the utterance sequence selected by the utterance selection unit and the feature amount corresponding to the utterance sequence, and the supervised correct text input to the adaptive data input unit and the feature amount corresponding to the supervised correct text, the sound Adapt the model,
An acoustic model adaptation device characterized by that. The acoustic model adaptation device according to claim 3,
A correct text selection unit that selects a supervised correct text corresponding to at least a part of an utterance sequence that is not selected by the utterance selection unit;
A correct text output unit for outputting the supervised correct text;
An acoustic model adaptation device characterized by comprising: The acoustic model adaptation device according to claim 4,
The correct text selection part
Selecting a supervised correct text corresponding to an utterance sequence that is not selected by the utterance selection unit and having a reliability that satisfies a predetermined criterion;
An acoustic model adaptation device characterized by that. The acoustic model adaptation device according to claim 4,
The supervised correct text input to the adaptive data input unit is
This is a supervised correct text output from the correct text output unit.
An acoustic model adaptation device characterized by that. An acoustic model adaptation method for adapting an acoustic model,
A process in which a speech recognition result using the acoustic model is input to the speech recognition result input unit;
A process of calculating a reliability that is an estimated value of a recognition rate for each utterance sequence obtained by dividing the word sequence of the speech recognition result using the speech recognition result,
The utterance selection unit uses the recognition rate of the acoustic model and the reliability for each utterance sequence to select an utterance sequence to be used for adaptation of the acoustic model;
The acoustic model adaptation unit adapts the acoustic model using the utterance sequence selected by the utterance selection unit and the feature amount corresponding to the utterance sequence;
An acoustic model adaptation method characterized by comprising:   The acoustic model adaptation program for functioning a computer as an acoustic model adaptation apparatus in any one of Claim 1 to 6.   A computer-readable recording medium storing the acoustic model adaptation program according to claim 8.

Images