JP2012118441A - Method, device, and program for creating acoustic model - Google Patents

Abstract

PROBLEM TO BE SOLVED: To automatically optimize the size of a voice recognition result lattice including voice recognition errors in a discrimination learning method.SOLUTION: A partial learning data selection unit selects voice data for partial learning from a voice database for learning, and a partial recognition parameter determination unit obtains such a determination recognition parameter that the voice data for partial learning becomes a partial lattice having a prescribed size. Then, a recognition unit for lattice creation uses the determination recognition parameter to generate a voice recognition result lattice. A discrimination learning unit performs discrimination learning by comparison between the voice recognition result lattice and a correct answer symbol sequence to create an acoustic model subjected to discrimination learning.

Description

  The present invention relates to an acoustic model creation method, an apparatus, and a program for creating an acoustic model using a discriminative learning method.

  As a learning method of an acoustic model, an identification learning method that improves the discrimination ability between symbols such as phonemes is increasingly used from the conventional method based on maximum likelihood estimation. The identification learning method improves the learning effect of the identification model by simultaneously using a reference word string that does not include a speech recognition error and a recognition word string that includes a speech recognition error. As the recognition word string, a speech recognition result expressed in a form such as a word lattice (a directed acyclic graph for compactly expressing a plurality of recognition word strings) is used.

  Discriminative learning methods are disclosed in, for example, Patent Documents 1 and 2 and Non-Patent Document 1. A conventional acoustic model creation apparatus 900 based on a discriminative learning method will be briefly described with reference to FIG. The acoustic model creation apparatus 900 includes a learning speech database 90, a language model storage unit 91, a learning acoustic model storage unit 92, a lattice creation recognition unit 93, and an identification learning unit 94.

  The learning speech database 90 stores learning speech data in which speech data and a correct symbol sequence thereof are combined. The language model storage unit 91 stores a grammar (including a pronunciation dictionary) that expresses a connection relationship between words. The learning acoustic model storage unit 92 stores an acoustic model for learning that associates phonemes and speech feature quantities. The lattice creation recognition unit 93 performs speech recognition on all the learning speech data based on the language model, the acoustic model, and the recognition parameters, and performs word recognition (hereinafter simply referred to as “speech recognition result lattice”). Or “word lattice”). The discrimination learning unit 94 performs discrimination learning by comparing the speech recognition result lattice and the correct symbol series to generate an acoustic model having been discriminated and learned.

  FIG. 12 shows an example of a speech recognition result lattice. FIG. 12 is a directed acyclic graph representing a plurality of recognized word strings including an error that has recognized the voice “Thank you for calling.” The acoustic model can be efficiently learned by performing discrimination learning by comparing the speech recognition result lattice with the correct lattice including no error.

JP 2007-322984 A JP 2006-201553 A

Erik McDermott and Atsushi Nakamura, “String and Lattice based Discriminative Training for the Corpus of Spontaneous Japanese Lecture Transcription Task”, INTERSPEECH, pp.2081-2084, Aug. 2007.

  Since the conventional acoustic model creation apparatus 900 uses a large amount of memory for creating word lattices, there is a problem of squeezing the memory capacity (disk capacity). With the recent increase in memory capacity, the amount of speech data stored in the learning speech database is becoming large-scale data such as several hundred hours. When speech recognition is performed on the large-scale speech data and a word lattice is generated, the word lattice contains many possible speech recognition results as described above. Will be enormous. In addition, since the size of the word lattice depends on the sound quality of the learning speech data, the acoustic model, the language model, the speech recognition parameters, etc., it is difficult to predict in advance.

  For this reason, if an attempt is made to generate a word lattice directly from large-scale audio data, the acoustic model creation apparatus 900 becomes inoperable because the disk capacity is used up and the memory is insufficient. For example, if the search beam width, which is one of the recognition parameters for speech recognition, is narrowed for the purpose of preventing this, speech data in which a word lattice cannot be obtained also appears, and the accuracy of the acoustic model after learning decreases. appear.

  The present invention has been made in view of the above points, and can automatically optimize a speech recognition parameter when generating a speech recognition result lattice to generate a word lattice of an appropriate size. An object of the present invention is to provide an acoustic model creation method, apparatus and program thereof.

  The acoustic model creation method of the present invention includes a partial learning data selection process, a partial lattice creation recognition process, a partial recognition parameter determination process, a lattice creation recognition process, and an identification learning process. In the partial learning data selection process, partial learning speech data is selected from the speech data stored in the learning speech database and the learning speech data that is a combination of the correct answer symbol series. The recognition process for creating the partial lattice is obtained by the partial learning speech data, the language model stored in the language model storage unit, the learning acoustic model stored in the learning acoustic model storage unit, and the partial recognition parameter determination process. A partial lattice is generated by performing speech recognition using the recognition parameter for control. In the partial recognition parameter determination process, the capacity of the partial lattice is evaluated to control the recognition parameter for control, and the control recognition parameter from which the partial lattice having a predetermined capacity is obtained is output as the determination recognition parameter. The lattice creation recognition process performs speech recognition on all learning speech data based on the language model, the learning acoustic model, and the stagnation recognition parameter, and generates a speech recognition result lattice. In the discriminative learning process, discriminative learning is performed by comparing the speech recognition result lattice and the correct symbol sequence to generate an identified acoustic model.

  According to the acoustic model creation method of the present invention, partial learning speech data is selected from the learning speech database, and a decision recognition parameter is obtained in which the partial learning speech data is a partial lattice having a predetermined size. Since the speech recognition result lattice is generated using the decision recognition parameter, the size of the speech recognition result lattice can be automatically made appropriate.

The figure which shows the function structural example of the acoustic model production apparatus 100 of this invention. The figure which shows the operation | movement flow of the acoustic model production apparatus 100. The figure which shows the function structural example of the acoustic model production apparatus 200 of this invention. The figure which shows the operation | movement flow of the acoustic model production apparatus 200. The figure which shows the function structural example of the acoustic model production apparatus 300 of this invention. The figure which shows the operation | movement flow of the acoustic model production apparatus 300. The figure which shows the function structural example of the acoustic model production apparatus 400 of this invention. The figure which shows the operation | movement flow of the acoustic model production apparatus 400. The figure which shows the function structural example of the acoustic model production apparatus 500 of this invention. The figure which shows the function structural example of the acoustic model production apparatus 600. FIG. The figure which shows the function structure of the conventional acoustic model production apparatus 900. The figure which shows an example of the speech recognition result lattice.

    Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

  FIG. 1 shows an example of the functional configuration of an acoustic model creation device 100 of the present invention. The operation flow is shown in FIG. The acoustic model creation device 100 includes a learning speech database 90, a partial learning data selection unit 11, a partial lattice creation recognition unit 12, a partial recognition parameter determination unit 13, a language model storage unit 91, and a learning acoustic model. A storage unit 92, a lattice creation recognition unit 93, and an identification learning unit 94 are provided. The functions of the units other than the database and the storage unit are realized by reading a predetermined program into a computer including, for example, a ROM, a RAM, and a CPU and executing the program by the CPU.

  The acoustic model creation device 100 is new to the conventional acoustic model creation device 900 in that it includes a partial learning data selection unit 11, a partial lattice creation recognition unit 12, and a partial recognition parameter determination unit 13.

  The acoustic model creation apparatus 100 performs speech recognition processing on speech digital signals, and the speech data 90 converted into digital signals is recorded in the learning speech database 90 as a plurality of speech files. The voice data is subjected to voice recognition processing for each frame with a time interval of, for example, 20 ms as one frame.

  The learning voice database 90 records learning voice data in which the voice data and its correct symbol series are combined. The partial learning data selection unit 11 selects partial learning voice data from the learning voice database 90 (step S11). For the selection, the size of the learning audio data, that is, an audio file with a large amount of audio file data may be used as partial learning audio data, or an audio file selected at random with the data amount being known is partially learned. Audio data may be used.

  The partial lattice creation recognizing unit 12 uses the partial learning speech data as a language model stored in the language model storage unit 91, a learning acoustic model stored in the learning acoustic model storage unit 92, and partial recognition parameter determination. Speech recognition is performed using the control recognition parameters input from the unit 13 to generate a partial lattice (step S12).

  Note that since the recognition parameter for control cannot be determined when the partial lattice is generated for the first time, the initial recognition parameter 120 set in advance in the partial lattice creation recognition unit 12 is used as the recognition parameter. The recognition parameters include search beam width and language weight. The search beam width is a cut-off width of a hypothesis of a speech recognition result. As the initial recognition parameter 120, for example, 1000 hypotheses are searched. The language weight is a weight by which the language score is multiplied when the reliability score is expressed by the sum of the acoustic score and the language score. The initial recognition parameter 120 is set to a value such as 10, for example. The word lattice is also described in Non-Patent Document 1 described above and is a general one. Since the generation of the partial lattice itself is not a main part of the present invention, detailed description thereof is omitted.

The partial recognition parameter determination unit 13 evaluates the capacity of the partial lattice output by the partial lattice creation recognition unit 12 and controls the recognition parameter for control, and determines the recognition parameter for control from which the partial lattice having a predetermined capacity is obtained. It outputs as a recognition parameter (step S13). Partial recognition parameter determining section 13, a portion Lattice capacity L _E as compared to the target lattice volume L _T, adjusting the control recognition parameter so as to approach the target.

When the partial Lattice capacity L _E is smaller than the target lattice volume L _T as used search beam width B as the control recognition parameters, expanding the search beam width B, it is larger conversely reduced. For example, the search beam width B of the control recognition parameters 'is based on the ratio r = _{L E} / _{L T} to the target lattice volume _{L T} B' may be obtained in = B / r. Then, the control recognition parameter from which a partial lattice having a predetermined capacity is obtained is output as a decision recognition parameter.

Target Lattice capacity L _T, so varies with the size of the audio data for partial learning, it sets a value in consideration of the number of frames that part training speech data (Equation (1)).

Here, L _E is the capacitance of the portion lattice generated from the audio partial learning data, N _E is the number of frames that part training speech data (file length). N is the total number of frames of all learning speech data.

The goal Lattice capacity L _T is, and the remaining capacity of the disk to write the word lattice, may be obtained from the ratio of the total capacity of the training speech data. Also, if you put the correct label to the input may be determined target lattice volume L _T based on the lattice capacity when given the true label as a language model (e.g., the reference 10-fold, etc.).

Partial recognition parameter determination unit 13 compares the capacitance L _E and the target lattice volume L _T parts lattice, substantially equal to the target lattice volume L _T (or the difference is below a certain value, for example, the difference is less than 1%, etc.) outputting a recognition parameter control the capacity L _E was obtained as a decision recognition parameter. In addition, you may provide the frequency limit etc. for adjustment here.

  The lattice creation recognizing unit 93 converts the language model stored in the language model storage unit 91, the learning acoustic model stored in the learning acoustic model storage unit 92, and all learning speech data based on the decision recognition parameters. Then, speech recognition is performed and a speech recognition result lattice is generated (step S93).

  The discrimination learning unit 94 performs discrimination learning by comparing the speech recognition result lattice generated by the lattice creation recognition unit 93 and the correct symbol sequence stored in the learning speech database 90 to generate a discriminatively learned acoustic model ( Step S94).

  As described above, the acoustic model creation apparatus 100 generates a partial lattice from a part of the speech data of the learning speech data, and determines the recognition parameters for speech recognition so that the size of the partial lattice becomes a predetermined size. Therefore, a speech recognition result lattice having an appropriate size can be automatically generated.

  Note that the speech recognition result lattice can be replaced with the speech recognition result N-best. That is, the recognition unit for partial lattice creation 12 and the partial recognition parameter determination unit 13 determine the determination recognition parameter by controlling the recognition parameter for control so that an N-best of an appropriate size can be obtained instead of the word lattice. Then, the lattice creation recognizing unit 93 generates the N-best of the speech recognition result by using the determination parameter, thereby automatically generating the N-best of the speech recognition result having an appropriate size. .

FIG. 3 shows a functional configuration example of the acoustic model creation device 200 of the present invention. The operation flow is shown in FIG. The acoustic model creation apparatus 200 is different from the above-described acoustic model creation apparatus 100 only in the partial learning data selection unit 20 and the partial lattice creation recognition unit 12 ′.

  The partial learning data selection unit 20 includes an initial lattice capacity calculation unit 201. The initial lattice capacity calculation means 201 uses the initial recognition parameters 120 (not shown, may be the same as that shown in FIG. 1) for the audio file of the learning audio data recorded in the learning audio database 90. A recognition process is performed, and the initial lattice capacity of each audio file is calculated (step S201). At this time, the initial lattice capacity may be obtained for all of the learning speech data, or may be obtained by limiting to a plurality of files. Then, the partial learning data selection unit 20 selects an audio file having a large initial lattice capacity as partial learning audio data (step S202).

  The partial lattice creation recognition unit 12 ′ transmits the initial lattice capacity magnitude information to the partial recognition parameter determination unit 13 and then uses the control recognition parameters input from the partial recognition parameter determination unit 13 for partial learning. A partial lattice of the audio data is created (step S12 '). Then, the partial recognition parameter determination unit 13 controls the recognition parameter for control so that the partial lattice has a predetermined size, and outputs the determination recognition parameter (step S13).

In this way, because the decision recognition parameters are set for a voice file with a large initial lattice capacity, it is possible to reduce the risk of exceeding the disk capacity when generating a speech recognition result lattice, and more efficiently A speech recognition result lattice can be generated.
[Modification]
The partial learning data selection unit 20 ′ may select an audio file having a large change in partial lattice capacity with respect to a change in recognition parameters as partial learning audio data. The partial learning data selection unit 20 ′ includes an initial lattice capacity calculation unit 201 ′.

  The initial lattice capacity calculating unit 201 ′ performs speech recognition processing using the initial recognition parameter 120 to calculate the initial lattice capacity of the audio file, and uses the second recognition parameter obtained by changing the initial recognition parameter 120 to A second lattice capacity is calculated (step S201 ').

  The partial learning data selection unit 20 ′ selects an audio file having a large difference between the initial lattice capacity and the second lattice capacity as partial learning audio data (step S20 ′). Here, by setting the second recognition parameter to a value slightly changed with respect to the initial recognition parameter (for example, a value obtained by changing the search beam width by about 10%), the partial learning data selection unit 20 ′ determines the recognition parameter. It is possible to select a highly sensitive audio file with a large change in lattice capacity with respect to the change.

  The partial recognition parameter determination unit 13 sets a decision recognition parameter using an audio file that is highly sensitive to changes in the recognition parameter. Therefore, the lattice creation recognition unit 93 can generate a speech recognition result lattice having a more appropriate size.

  FIG. 5 shows a functional configuration example of the acoustic model creation device 300 of the present invention. The operation flow is shown in FIG. The acoustic model creation apparatus 300 differs from the acoustic model creation apparatuses 100 and 200 described above only in the partial learning data selection unit 30.

The partial learning data selection unit 30 includes a feature amount extraction unit 301 and a reliability score calculation unit 302. Feature calculating unit 301 converts the audio feature o _t audio data of the audio file training speech data for each frame (step S301). The audio feature _{o t,} for example, is used and 12 yuan MFCC (Mel-Frequency Cepstrum Coefficient) , and dynamic parameters ΔMFCC like its variation, the power and Δ power like. Also, processing such as cepstrum average normalization (CMN) may be performed.

Confidence score calculating means 302 converts the reliability estimated from the acoustic score and the language score of speech recognition results for a sequence of speech features o _t, files of the file the reliability of each audio file from the reliability A reliability score normalized by the length is calculated (step S302). The reliability score is estimated from the speech recognition result, and it is preferable to use the reliability score conventionally used in the speech recognition apparatus.

  The partial learning data selection unit 30 selects an audio file having a small reliability score as partial learning audio data (step S303). At this time, the reliability scores of all of the learning speech data may be calculated, or the partial learning speech data may be selected by comparing the reliability scores obtained by limiting to a plurality.

  If a speech recognition result lattice is generated from an audio file having a small reliability score, the number of word conflict candidates increases, and the lattice capacity increases. Therefore, in order to suppress the total amount of the speech recognition result lattice capacity, the risk of the total lattice capacity exceeding the disk capacity can be reduced by obtaining the decision recognition parameter from the sound file having a small reliability score.

  FIG. 7 shows a functional configuration example of the acoustic model creation device 400 of the present invention. The operation flow is shown in FIG. The acoustic model creation apparatus 400 differs from the acoustic model creation apparatus 300 in that the partial learning data selection unit 40 includes a pre-fast reliability score calculation unit 401.

The prior high-speed reliability score calculation means 401 calculates the highest monophonic product having the highest product of the output probability b _s obtained from a monophone GMM (Gaussian Mixture Model) for the speech feature value for each frame and the appearance probability of the state to which the GMM belongs. The likelihood value P (s ^) bs ^ (o _t ) is obtained, and the difference between the logarithm of the monophone maximum likelihood value and the logarithm of the voice / pause maximum likelihood value bg ^ (o _t ) is averaged for each audio file. A high-speed prior reliability score is calculated (step S401). Represented by the formula (2) high-speed pre confidence score c of each frame (o _t), Equation (3) is a fast advance confidence score C averaged voice file basis.

  Here, T is the total frame length of each audio file. In addition, the calculation of the high-speed prior reliability score C is described in Reference 1: “Kohashikawa, Asami, Yamaguchi, Masatsugu, Takahashi“ Speech recognition target data selection based on prior reliability estimation ”, Proc. Month or Reference 2: “Kobashikawa, Asami, Yamaguchi, Masataki, Takahashi,“ Efficient Data Selection for Speech Recognition Based on Prior Confidence Estimation Using Speech and Context Independent Models ”, INTERSPEECH 2010, pp.238-241 September 2010 May be used.

  The partial learning data selection unit 40 selects an audio file having a small high-speed prior reliability score C as partial learning audio data (step S403). At this time, the high-speed pre-reliability score C of all the learning speech data may be calculated, or the partial pre-speech speech data is selected by comparing the high-speed pre-reliability scores C obtained by limiting to a plurality. May be.

  Thus, by calculating the reliability score from the output probability obtained from the mophone GMM, it is possible to speed up the processing compared to the case where an acoustic model such as triphone or biphone is used. The acoustic model creation device 400 can generate a speech recognition result lattice faster than the acoustic model creation device 300.

  FIG. 9 shows a functional configuration example of the acoustic model creation device 500 of the present invention. The acoustic model creation device 500 is a lattice creation language model storage unit 50 in which the language model storage unit stores a lattice creation language model created from the speech file of the learning speech database. Only the point is different.

  By creating a language model used when creating a word lattice from an acoustic model for learning, unknown words are eliminated and the number of conflict candidates is reduced. Therefore, the capacity of the speech recognition result lattice can be reduced. As a result, the disk can be used efficiently.

  The acoustic model creation device 500 has been described based on the acoustic model creation device 100, but the language model storage unit of the other acoustic model creation devices 200 to 400 is changed to a lattice creation language model storage unit 50. Similar effects can be expected.

  In FIG. 9, a lattice creation language model creation unit 51 that creates a language model from the correct symbol sequence recorded in the learning speech database 90 is indicated by a broken line. In this manner, the lattice creation language model creation unit 51 may be integrated as the acoustic model creation device 500 to sequentially generate the lattice creation language model. Alternatively, the lattice creation language model creation unit 51 may be a separate unit, and only the lattice creation language model storage unit 50 that stores a previously created lattice creation language model may be used.

  FIG. 10 shows a functional configuration example of the acoustic model creation device 600 of the present invention. The acoustic model creation device 600 differs from the acoustic model creation device 500 in that it further includes a language model creation data maintenance unit 60.

  The language model creation data maintenance unit 60 extracts katakana or hiragana reading labels corresponding to the correct symbol series from the learning speech database 90. The lattice creation language model creation unit 51 creates a language model from the correct reading label.

  By using correct reading labels, notation fluctuation can be reduced and the increase in vocabulary size can be suppressed, so that the lattice capacity can be reduced.

  Depending on the word separation of the reading label, there may be a problem that the sound is divided as a word boundary where a double vowel should be used as a sound. The double vowels are “ei”, “ou”, “iu”, etc., and each is one sound. When this double vowel is divided into “e / i”, the display shakes, causing an increase in vocabulary size.

  In view of this, the section that should be a double vowel is subjected to a frame concatenation process so that voices having the same acoustic feature amount are not assigned to different phonemes. By doing so, confusion between phonemes can be reduced and the accuracy of the acoustic model can be increased. It is also possible to reduce the capacity of the speech recognition result lattice.

  Specifically, the double vowels to be connected are registered in the double vowel list, and it is determined whether or not the double vowel is included in the list. In addition, a method of creating a language model by performing pose insertion determination for comparing likelihoods between the case of dividing and inserting a pose and the case of connecting poses is conceivable.

  As described above, the acoustic model creation apparatuses 100 to 500 according to the present invention can suppress the capacity of the generated speech recognition result lattice. Since the capacity of the speech recognition result lattice can be estimated in advance, it is possible to appropriately determine the free capacity of the disk to be prepared. In addition, it is possible to automatically determine an appropriate speech recognition parameter, and it is possible to automatically generate a discriminatively learned acoustic model with high accuracy by adapting to the disk capacity. Furthermore, since the acoustic model creation apparatus 600 of the present invention creates a language model from the learning speech database 90, the acoustic model creation device 600 has excellent effects such as no need to prepare a language model.

  In addition, the above-described word lattice of the speech recognition result may be replaced with the N-best of the speech recognition result, and even in that case, it is possible to provide an acoustic model creation method that exhibits the same effect.

  When the processing means in the above apparatus is realized by a computer, the processing contents of functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  Further, the processes described in the above method and apparatus are not only executed in time series according to the order of description, but also may be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. Good.

  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 media, MO (Magneto Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  This 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. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

Claims (15)

A partial learning data selection process for selecting partial learning voice data from learning voice data that is a combination of the voice data stored in the learning voice database and the correct answer symbol sequence;
Using the partial learning speech data, the language model stored in the language model storage unit, the learning acoustic model stored in the learning acoustic model storage unit, and the control recognition parameters obtained in the partial recognition parameter determination process A recognition process for creating a partial lattice by generating a partial lattice by speech recognition;
A partial recognition parameter determination process for controlling the recognition parameter for control by evaluating the capacity of the partial lattice and outputting the recognition parameter for control from which the partial lattice of a predetermined capacity is obtained as a determination recognition parameter;
A recognition process for creating a lattice that performs speech recognition on all the learning speech data based on the language model, the learning acoustic model, and the decision recognition parameters, and generates a speech recognition result lattice;
A discriminative learning process in which discriminative learning is performed by comparing the speech recognition result lattice with the correct symbol sequence to generate an identified acoustic model;
An acoustic model creation method comprising: The acoustic model creation method according to claim 1,
The partial learning data selection process is as follows:
Including an initial lattice capacity calculation step of performing speech recognition processing using an initial recognition parameter for the speech file of the learning speech data and calculating an initial lattice capacity of each of the speech files;
A method for creating an acoustic model, characterized by being a process of selecting an audio file having a large initial lattice capacity as the partial learning audio data. In the acoustic model creation method according to claim 2,
The initial lattice capacity calculating step further calculates a second lattice capacity using a second recognition parameter obtained by changing the initial recognition parameter.
The acoustic model creation method, wherein the partial learning data selection process is a process of selecting an audio file having a large difference between the initial lattice capacity and the second lattice capacity as the partial learning audio data. The acoustic model creation method according to claim 1,
The partial learning data selection process is as follows:
A feature amount analyzing step of converting the sound data of the sound file of the learning sound data into a sound feature amount for each frame;
A reliability score calculation step of converting the voice feature amount into a reliability composed of an acoustic score and a language score, and calculating a reliability score obtained by normalizing the reliability of each audio file from the reliability by the file length of the file; Including
A method for creating an acoustic model, characterized by being a process of selecting an audio file having a small reliability score as the partial learning audio data. The acoustic model creation method according to claim 1,
The partial learning data selection process is as follows:
A feature amount analyzing step of converting the sound data of the sound file of the learning sound data into a sound feature amount for each frame;
The product having the highest product of the output probability obtained from the monophone GMM with respect to the speech feature value for each frame and the appearance probability of the state to which the GMM belongs is obtained as the monophone maximum likelihood value, and the logarithm of the monophone maximum likelihood value and the speech / pause A high-speed pre-reliability score calculating step of calculating a high-speed pre-reliability score by averaging the difference from the logarithm of the maximum likelihood value for each audio file;
A method for creating an acoustic model, characterized by being a process of selecting an audio file having a small high-speed prior reliability score as the audio data for partial learning. The acoustic model creation method according to any one of claims 1 to 5,
The acoustic model creation method, wherein the language model stores a lattice creation language model created from an audio file of the learning speech database. In the acoustic model creation method according to claim 6,
The acoustic model creation method, wherein the lattice creation language model is a language model created from a correct reading label. In the acoustic model creation method according to claim 7,
An acoustic model creation method, wherein the double vowel of the correct reading label is not a word boundary. A speech database for learning that records speech data for training that includes speech data and a correct symbol sequence thereof;
A language model storage unit storing a grammar expressing a connection relation between words as a language model;
An acoustic model storage unit for learning that stores an acoustic model for learning that associates phonemes and feature quantities of speech;
A partial learning data selection unit for selecting partial learning voice data from the learning voice data;
A recognition unit for creating a partial lattice by recognizing the speech data for partial learning using the recognition parameter for control input from the language model, the acoustic model, and a partial recognition parameter determination unit;
A partial recognition parameter determination unit that evaluates the capacity of the partial lattice to control the recognition parameter for control and outputs the recognition parameter for control from which the partial lattice of a predetermined capacity is obtained as a determination recognition parameter;
A lattice creation recognizing unit that performs speech recognition on all the learning speech data based on the language model, the learning acoustic model, and the determination recognition parameter, and generates a speech recognition result lattice;
A discrimination learning unit that performs discrimination learning by comparing the speech recognition result lattice and the correct symbol series to generate a discriminated learning acoustic model;
An acoustic model creation device comprising: In the acoustic model creation device according to claim 9,
The partial learning data selection unit includes an initial lattice capacity calculation unit that performs a speech recognition process using an initial recognition parameter for the sound file of the learning sound data and calculates an initial lattice capacity of each of the sound files,
An acoustic model generation apparatus, wherein an audio file having a large initial lattice capacity is selected as the partial learning audio data. The acoustic model creation device according to claim 10,
The initial lattice capacity calculating means calculates a second lattice capacity using a second recognition parameter obtained by changing the initial recognition parameter.
The acoustic model creation device, wherein the partial learning data selection unit selects a voice file having a large difference between the initial lattice capacity and the second lattice capacity as the partial learning voice data. In the acoustic model creation device according to claim 9,
The partial learning data selection unit
Feature amount analyzing means for converting the sound data of the sound file of the learning sound data into a sound feature amount for each frame;
A reliability score calculation means for converting the above-mentioned voice feature amount into a reliability composed of an acoustic score and a language score, and calculating a reliability score obtained by normalizing the reliability of each audio file by the file length of the file from the reliability. Prepared,
An acoustic model generation apparatus, wherein an audio file having a small reliability score is selected as the partial learning audio data. In the acoustic model creation device according to claim 9,
The partial learning data selection unit
Feature amount analyzing means for converting the sound data of the sound file of the learning sound data into a sound feature amount for each frame;
The product having the highest product of the output probability obtained from the monophone GMM with respect to the speech feature value for each frame and the appearance probability of the state to which the GMM belongs is obtained as the monophone maximum likelihood value, and the logarithm of the monophone maximum likelihood value and the speech / pause A high-speed pre-reliability score calculation unit that calculates a high-speed pre-reliability score that averages the difference from the logarithm of the maximum likelihood value for each audio file,
An acoustic model generation apparatus, wherein an audio file having a small high-speed prior reliability score is selected as the partial learning audio data. The acoustic model creation device according to any one of claims 9 to 13,
The language model storage unit
An acoustic model creation apparatus characterized by storing a lattice creation language model created from the learning speech database.   A program for causing a computer to execute the acoustic model creation method according to any one of claims 1 to 8.

Images