JP2013083798A - Sound model adaptation device, sound model adaptation method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a sound model adaptation device capable of contributing well to a parameter modification and performing adaptation to a sound model using data that prevents deterioration of effectiveness of adaptation.SOLUTION: In a sound model adaptation device 10, a voice recognition part 100 outputs a voice recognition result text and reliability from an input voice by means of a pre-adaptation sound model. A voice recognition result registration part 200 stores a voice recognition result consisting of a speaker ID, a voice, the voice recognition result text and the reliability. A struggling-speaker detection part 300 extracts a speaker ID of a struggling speaker with whom voice recognition accuracy is lower than with another speaker. An adapting-data selection part 400 reads a voice recognition result of a speaker ID being the speaker ID of the struggling speaker, the result of which the reliability is equal to or higher than a predetermined reliability threshold and outputs adapting-data. A sound model adapting-part 500 uses a preset adapting-parameter to output a post-adaptation sound model.

Description

  The present invention relates to an acoustic model adaptation apparatus, an acoustic model adaptation method, and a program for performing adaptation of an acoustic model used for speech recognition.

  In general, an unspecified speaker acoustic model is used to recognize speech of an unspecified number of speakers. The unspecified speaker acoustic model is an acoustic model in which parameters are set so that correspondence between speech and phonemes can be correctly obtained in many speakers (ideally all speakers). The parameters of the unspecified speaker acoustic model are determined by a machine learning algorithm from a pair of voices of hundreds or more of speakers and text describing the utterance contents.

  However, even if the speech of many speakers is used for learning, it is impossible to cover all speakers who can use the speech recognition system. In an actual speech recognition system, there are speakers (hereinafter referred to as poor speakers) whose speech recognition accuracy is significantly lower than other speakers even when an unspecified speaker acoustic model is used. There arises a problem of lowering the performance.

  In order to cope with this problem, Japanese Patent Laid-Open No. 2004-151867 discloses that when the speech recognition reliability exceeds a certain threshold among the speeches input during operation of the speech recognition system, the speech and the speech recognition result text of the speech A method of updating parameters of an acoustic model by applying an acoustic model adaptation algorithm to a pair is described (hereinafter, acoustic model adaptation using speech recognition result text is referred to as “unsupervised adaptation”). By using speech with high reliability, it is possible to prevent the adaptation effect from fading due to a recognition error included in the speech recognition result text. For example, by storing the voice input during the operation of the voice recognition system and applying the method of Patent Document 1 at a timing when a certain amount is stored, the parameters of the acoustic model are set so as to match the actually input voice. Can be updated.

JP 2011-75622 A

  However, in the method described in Patent Document 1, when selecting a voice to be used for unsupervised adaptation of an acoustic model, a large amount of data that should not be used for adaptation is selected. Data that should not be used for adaptation is speech that matches the parameters of the acoustic model before adaptation, speech recognition result text including recognition errors, and the like. Many of the voices showing high reliability match the parameters of the acoustic model before adaptation. Such speech has a small contribution to parameter correction by adaptation of the acoustic model. On the other hand, there are a small number of recognition errors in the speech recognition result text with high reliability. When speech recognition result text including recognition errors is used for acoustic model adaptation, the effect of acoustic model adaptation is reduced. When such data is used for adaptation, there is a problem that the effect of unsupervised adaptation of the acoustic model is suppressed.

  The present invention has been made in view of the above points, and an acoustic model that can perform unsupervised adaptation of an acoustic model using data that greatly contributes to parameter correction and does not easily reduce the effect of adaptation. An object is to provide an adaptive device.

  In order to solve the above problems, an acoustic model adaptation device according to the present invention includes an acoustic model storage unit, a speech recognition result storage unit, a speech recognition unit, a speech recognition result registration unit, a poor speaker detection unit, and an adaptation data selection unit. And an acoustic model adaptation unit. The acoustic model storage unit stores a pre-adaptation acoustic model. The voice recognition result is stored in the voice recognition result storage unit. The speech recognition unit outputs at least speech recognition result text and reliability from the input speech using the pre-adaptation acoustic model. The speech recognition result registration unit stores in the speech recognition result storage unit a speech recognition result including at least a speaker ID, speech, speech recognition result text, and reliability. The poor speaker detection unit reads all the speech recognition results from the speech recognition result storage unit, and determines the speaker ID of the poor speaker whose speech recognition accuracy is lower than other speakers based on preset detection conditions. Extract. The adaptation data selection unit reads from the speech recognition result storage unit a speech recognition result whose speaker ID is a speaker ID of a poor speaker and whose reliability is equal to or higher than a predetermined reliability threshold, Data for adaptation consisting of speech and speech recognition result text is extracted. The acoustic model adaptation unit outputs a post-adaptation acoustic model from the pre-adaptation acoustic model and the adaptation data, using preset adaptation parameters.

  According to the present invention, in unsupervised adaptation of an unspecified speaker acoustic model using accumulated speech, adaptation of the acoustic model is performed by using speech with high reliability among speeches of poor speakers as adaptation data. The recognition accuracy improvement effect by can be enhanced.

  Further, the improvement of the recognition accuracy of the unspecified speaker acoustic model reduces the variation in the speech recognition accuracy among speakers, so that it is possible to realize a speech recognition system that is more convenient for more users.

1 is a block diagram illustrating a configuration of an acoustic model adaptation device according to Embodiment 1. FIG. 5 is a flowchart illustrating the operation of the acoustic model adaptation device according to the first embodiment. FIG. 2 is a block diagram illustrating a configuration of an acoustic model adaptation device according to a modification of the first embodiment. 5 is a flowchart showing the operation of the acoustic model adaptation apparatus of the modified example of the first embodiment. FIG. 6 is a block diagram illustrating a configuration of an acoustic model adaptation device according to a second embodiment. 10 is a flowchart showing the operation of the acoustic model adaptation apparatus of the second embodiment. FIG. 6 is a block diagram illustrating a configuration of an acoustic model adaptation device according to a modification of the second embodiment. 10 is a flowchart showing the operation of the acoustic model adaptation apparatus according to the modified example of the second embodiment.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the structure part which has the same function, and duplication description is abbreviate | omitted.

  First, an outline of the present invention will be described. In the first embodiment, speech accumulated in the speech recognition system is analyzed to detect poor speakers. Unsupervised adaptation of the speaker-independent acoustic model is performed by using, as adaptation data, speech with high reliability among the speeches of the weak speakers detected. Since the voice of a poor speaker does not match the parameters of the pre-adaptation acoustic model, it is possible to increase the contribution to parameter correction by the acoustic model adaptation. Furthermore, by selecting speech with high reliability from the speech of poor speakers, it is possible to make it difficult for the effect of acoustic model adaptation to be suppressed by using speech recognition result text with relatively few errors as adaptation data. .

  When the post-adaptation acoustic model that is finally output in the first embodiment is used, there is a possibility that the recognition accuracy is greatly lowered in the speech of a speaker other than the poor speaker who has been detected. Therefore, in Example 2, acoustic model adaptation is performed using a plurality of adaptation parameters, and a plurality of post-adaptation acoustic model candidates are generated. Using each generated acoustic model candidate, the accumulated speech is recognized again, the reliability after adaptation is calculated, and the acoustic model candidate with the highest reliability from the pre-adaptation acoustic model is determined as the after-adaptation acoustic model. adopt. If any of the generated acoustic model candidates is used and the reliability falls below a certain threshold value, the post-adaptation acoustic model is rejected. With this processing, it is possible to output a post-adaptation acoustic model in which the recognition accuracy is improved over the entire accumulated speech (that is, there is no significant reduction in recognition accuracy even in speakers other than poor speakers).

  With reference to FIGS. 1 and 2, the operation of the acoustic model adaptation apparatus 10 according to the first embodiment of the present invention will be described in detail. FIG. 1 is a block diagram showing a configuration of an acoustic model adaptation apparatus 10 according to the first embodiment of the present invention. FIG. 2 is a flowchart showing the operation of the acoustic model adaptation apparatus 10 according to the first embodiment of the present invention.

  In the following, description will be made in the order of procedures actually performed. The acoustic model adaptation apparatus 10 of the present embodiment includes a speech recognition unit 100, a speech recognition result registration unit 200, a poor speaker detection unit 300, an adaptation data selection unit 400, an acoustic model adaptation unit 500, an acoustic model storage unit 800, a speech A recognition result storage unit 900 is provided.

  The acoustic model storage unit 800 stores a pre-adaptation acoustic model.

  The speech recognition unit 100 receives speech and the pre-adaptation acoustic model stored in the acoustic model storage unit 800, and performs speech recognition on the input speech (S100). The reliability is calculated simultaneously with the speech recognition, and the obtained speech recognition result text and the reliability are set and output. The input voice may be a voice document (a series of a plurality of utterances such as a conference call or a lecture) or an utterance (a voice section generated by a breath sandwiched between silence sections). . The output reliability is the speech document recognition reliability when the input speech is a speech document, and the speech recognition reliability when the input speech is an utterance. Speech document recognition reliability is, for example, “Taichi Asami, Satoshi Kobashikawa, Yoshikazu Yamaguchi, Hirokazu Masami, Satoshi Takahashi,“ Estimation of speech document recognition reliability using word context consistency and acoustic likelihood ”, Technical report, SP, 110 (43), pp.43-48, 2010. " The utterance recognition reliability is calculated by a method described in, for example, “JP 2005-148342 A”. The reliability calculation method described above is merely an example, and various other reliability calculation methods can be used.

  The voice recognition result registration unit 200 receives the speaker ID, the voice, the voice recognition result text output by the voice recognition unit 100, and the reliability, and sets the input speaker ID, the voice, the voice recognition result text, and the reliability. The voice recognition result thus set is stored in the voice recognition result storage unit 900 (S200). The input speaker ID may be, for example, the value of the user ID obtained at the time of user authentication by the voice recognition system, or by applying the existing speaker identification technology described in “JP 2000-148187”. It is good also as a value obtained.

  The weak speaker detection unit 300 receives a preset detection condition θ and outputs a speaker ID of a poor speaker including zero or more speaker IDs. The timing at which the weak speaker detection unit 300 executes the process may be, for example, when an execution instruction is received from the system operator, or may be a preset cycle (1st of every month, every Sunday, etc.). Alternatively, it may be performed each time a preset data amount is registered (for example, every 1000 new voices are registered).

  Detection of poor speakers is performed by performing the following processes (1) to (4) for all speaker IDs stored in the speech recognition result storage unit 900. In the following description, it is assumed that the speaker ID to be processed is X.

(1) The reliability included in all speech recognition results for which the speaker ID = X is acquired from the speech recognition result storage unit 900 (S301). Hereinafter, the acquired set of reliability is referred to as ConfListID_X.
(2) The reliability included in all speech recognition results for which speaker ID ≠ X is acquired from the speech recognition result storage unit 900 (S302). Hereinafter, the acquired set of reliability is referred to as ConfListID_notX.
(3) From ConfListID_X and ConfListID_notX, it is determined whether or not speaker ID = X is detected as the speaker ID of the poor speaker based on the detection condition θ. Any of the following patterns A and B may be used as a method for determining whether or not to detect.
(Pattern A) If the value obtained by subtracting the average value m_X of ConfListID_X from the average value m_notX of ConfListID_notX is equal to or greater than the threshold θ, it is detected. In this case, the detection condition θ is a threshold value of the average difference in reliability. As a determination method of θ, for example, there is a method of calculating standard deviation σ of all the reliability stored in the speech recognition result storage unit 900 and setting θ = σ.
(Pattern B) Test whether there is a difference between the average value m_notX of ConfListID_notX and the average value m_X of ConfListID_X, and detect if the hypothesis that “m_X is smaller than m_notX” is supported at the significance level θ%. The t-test is used as the test method. In this case, the detection condition θ is the significance level of the test. For example, θ = 5% may be set.
(4) If it is determined to be detected, speaker ID = X is output as the speaker ID of the poor speaker (S303).

  The speaker ID of the poor speaker output by the weak speaker detection unit 300 is “a speaker biased toward a lower reliability than other speakers (ie, a speaker with low recognition accuracy)”. Represents. In (3), when all the speaker IDs are not detected and no speaker ID of a weak speaker is output, the subsequent processing is not executed.

  The adaptation data selection unit 400 receives the preset reliability threshold δ and the speaker ID of the poor speaker output from the poor speaker detection unit 300, and outputs adaptation data used for adaptation of the acoustic model ( S400). The speech recognition result text is paired with the speech included in the speech recognition result that is the speaker ID of the poor speaker to whom the speaker ID is input from the speech recognition result storage unit 900 and the reliability is the reliability threshold δ or more. And get it. A set of acquired speech and speech recognition result text pairs is output as adaptation data. As a method of determining δ, for example, there is a method of calculating an average value μ and standard deviation σ of all the reliability stored in the speech recognition result storage unit 900 and setting δ = μ−σ.

  Since the adaptation data output from the adaptation data selection unit 400 is limited to speakers who do not match the parameters of the pre-adaptation acoustic model, the contribution to parameter correction during adaptation is large. In addition, since the speech recognition result text with relatively few recognition errors is selected, it is difficult to suppress the effect of adaptation.

  The acoustic model adaptation unit 500 receives a preset adaptation parameter τ, the adaptation data output from the adaptation data selection unit 400, and the pre-adaptation acoustic model stored in the acoustic model storage unit 800. Output (S500). The post-adaptation acoustic model is generated by applying an acoustic model adaptation algorithm to the pre-adaptation acoustic model and the adaptation data using the input adaptation parameter τ. The acoustic model adaptation algorithm is, for example, “J.-L.Gauvain and C.-H.Lee,“ Maximum a Posteriori Estimation for Multivariate Gaussian Mixture Observations of Markov Chains ”, IEEE trans. On Speech and Audio processing, 2 (2 , pp.291-298, 1994 (Reference 1), etc., can be used. Although the meaning of the adaptation parameter τ varies depending on the acoustic model adaptation algorithm to be used, when the acoustic model adaptation algorithm described in Reference Document 1 is used, it is a positive numerical value representing the weight of adaptation data for the pre-adaptation acoustic model. In this case, for example, τ = 50 may be set.

The post-adaptation acoustic model output by the acoustic model adaptation unit 500 is an acoustic model whose parameters have been modified so that it matches even a weak speaker who did not match the pre-adaptation acoustic model.
[Modification example]
With reference to FIG. 3 and FIG. 4, the operation of the acoustic model adaptation apparatus 10 ′ according to the modification of the first embodiment of the present invention will be described in detail. FIG. 3 is a block diagram showing a configuration of an acoustic model adaptation device 10 ′ according to a modification of the first embodiment of the present invention. FIG. 4 is a flowchart showing the operation of the acoustic model adaptation apparatus 10 ′ according to the modification of the first embodiment of the present invention.

  The acoustic model adaptation apparatus 10 ′ of this modification includes a speech recognition unit 100, a speech recognition result registration unit 200, a poor speaker detection unit 300, an adaptation data selection unit 400, an acoustic model adaptation unit 510, an acoustic model storage unit 800, A speech recognition result storage unit 900 is provided.

  The acoustic model adaptation unit 510 performs the same process as the acoustic model adaptation unit 500 of the first embodiment (S511), and then stores the output post-adaptation acoustic model in the acoustic model storage unit 800 (S512). Subsequently, the speech included in all speech recognition results stored in the speech recognition result storage unit 900 is recognized again (S513), and the speech recognition result text and the reliability are updated (S514). Thereafter, the processes after the weak speaker detection unit 300 are repeatedly executed. In the repetitive processing, for example, whether an acoustic model after adaptation is generated a predetermined number of times (usually 2 to 3 times) (S992), or the poor speaker detection unit 300 determines the speaker ID of the poor speaker. If any one of them is satisfied (S991), the process is stopped.

  Next, the operation of the acoustic model adaptation apparatus 20 according to the second embodiment of the present invention will be described in detail with reference to FIGS. FIG. 5 is a block diagram showing the configuration of the acoustic model adaptation apparatus 20 according to the second embodiment of the present invention. FIG. 6 is a flowchart showing the operation of the acoustic model adaptation apparatus 20 according to the second embodiment of the present invention.

  In the following, description will be made in the order of procedures actually performed. The acoustic model adaptation device 20 of the present embodiment includes a speech recognition unit 100, a speech recognition result registration unit 200, a poor speaker detection unit 300, an adaptation data selection unit 400, an acoustic model adaptation unit 550, an acoustic model selection unit 600, an acoustic A model storage unit 800 and a speech recognition result storage unit 900 are provided.

  The acoustic model adaptation unit 550 is supplied with a list of preset adaptation parameters, the pre-adaptation acoustic model stored in the acoustic model storage unit 800, and the adaptation data output from the adaptation data selection unit 400. The acoustic model candidate is output (S550). A plurality of post-adaptation acoustic model candidates are generated corresponding to each adaptation parameter by applying an acoustic model adaptation algorithm using each adaptation parameter included in the input list of adaptation parameters. As the acoustic model adaptation algorithm, the same algorithm as that of the acoustic model adaptation unit 500 of the first embodiment can be used. For example, when the acoustic model adaptation algorithm described in Reference Document 1 is used, the list of adaptation parameters can be set to 10 values set from 10 to 100 in 10 increments. The wider the range of values and the smaller the step size, the more accurate the model can be selected by the processing of the acoustic model selection unit, but the calculation time increases.

  The acoustic model selection unit 600 receives a preset reliability decrease threshold ε and a plurality of post-adaptation acoustic model candidates output from the acoustic model adaptation unit 550, and outputs a post-adaptation acoustic model. First, a pre-adaptation reliability average value BeforeAveConf, which is an average value of all the reliability levels stored in the speech recognition result storage unit 900, is calculated (S601). Next, using each of the plurality of input post-adaptation acoustic model candidates, the speech included in all speech recognition results stored in the speech recognition result storage unit 900 is speech recognized, and the reliability is calculated at the same time. Then, an average value of reliability is obtained for each candidate acoustic model after adaptation (S602). Further, from the average value of reliability for each post-adaptation acoustic model candidate, the post-adaptation reliability average value AfterAveConf, which is the maximum value, and the post-adaptation reliability average value AfterAveConf are selected. . Subsequently, the reliability decrease range is obtained by subtracting the post-adaptation reliability average value AfterAveConf from the pre-adaptation reliability average value BeforeAveConf (S603). If the reliability decrease width is less than the input reliability decrease width threshold value ε, MaxAcou is output as an after-adaptation acoustic model. If the reliability decrease width is equal to or greater than the reliability decrease width threshold value ε, nothing is performed. The process ends without outputting (S604). The reliability lowering threshold ε is a numerical value equal to or greater than 0. For example, the standard deviation σ of all the reliability stored in the speech recognition result storage unit 900 in advance is obtained, and ε = σ can be set. it can. Alternatively, ε = 0 may be set so as to output only when the reliability is passively improved.

The post-adaptation acoustic model output from the acoustic model selection unit 600 is a post-adaptation acoustic model that does not significantly reduce (or necessarily improve) the reliability of the entire speech stored in the speech recognition result storage unit 900.
[Modification example]
With reference to FIG. 7, FIG. 8, operation | movement of acoustic model adaptation apparatus 20 'which concerns on the modification of Example 2 of this invention is demonstrated in detail. FIG. 7 is a block diagram showing a configuration of an acoustic model adaptation apparatus 20 ′ according to a modification of the second embodiment of the present invention. FIG. 8 is a flowchart showing the operation of the acoustic model adaptation apparatus 20 ′ according to the modification of the second embodiment of the present invention.

  The acoustic model adaptation apparatus 20 ′ of this modification includes a speech recognition unit 100, a speech recognition result registration unit 200, a poor speaker detection unit 300, an adaptation data selection unit 400, an acoustic model adaptation unit 550, an acoustic model selection unit 610, An acoustic model storage unit 800 and a speech recognition result storage unit 900 are provided.

  The acoustic model selection unit 610 performs the same processing as the acoustic model selection unit 600 of the second embodiment (S611 to S614), and then stores the output post-adaptation acoustic model in the acoustic model storage unit 800 (S615). Subsequently, the speech included in all speech recognition results stored in the speech recognition result storage unit 900 is recognized again (S616), and the speech recognition result text and the reliability are updated (S617). Thereafter, the processes after the weak speaker detection unit 300 are repeatedly executed. In the repetitive processing, for example, whether an acoustic model after adaptation is generated a predetermined number of times (usually 2 to 3 times) (S992), or the poor speaker detection unit 300 determines the speaker ID of the poor speaker. The process is stopped when either one is not output (S991) or the acoustic model selection unit 610 does not output the post-adaptation acoustic model (S993).

<Program, recording medium>
The various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. 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. 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.

  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 present apparatus is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

  The present invention can be used for unsupervised adaptation of an unspecified speaker acoustic model used for speech recognition.

10, 10 ′, 20, 20 ′ Acoustic model adaptation apparatus 100 Speech recognition unit 200 Speech recognition result registration unit 300 Poor speaker detection unit 400 Adaptive data selection unit 500, 510, 550 Acoustic model adaptation unit 600, 610 Acoustic model selection Unit 800 acoustic model storage unit 900 speech recognition result storage unit

Claims (8)

The acoustic model storage unit stores a pre-adaptation acoustic model,
A speech recognition unit that outputs at least a speech recognition result text and reliability from the input speech using the pre-adaptation acoustic model;
A speech recognition result registration unit, wherein the speech recognition result registration unit stores a speech recognition result including at least a speaker ID, the speech, the speech recognition result text, and the reliability in a speech recognition result storage unit;
The weak speaker detection unit reads all the speech recognition results from the speech recognition result storage unit, and is a speaker of a poor speaker whose speech recognition accuracy is lower than other speakers based on preset detection conditions A weak speaker detection step of extracting an ID;
The adaptive data selection unit reads from the speech recognition result storage unit a speech recognition result whose speaker ID is the speaker ID of the poor speaker and whose reliability is equal to or higher than a predetermined reliability threshold. An adaptation data selection step for extracting adaptation data comprising at least the speech and the speech recognition result text;
An acoustic model adaptation unit that outputs a post-adaptation acoustic model from the pre-adaptation acoustic model and the adaptation data using a preset adaptation parameter;
An acoustic model adaptation method characterized by comprising: The acoustic model adaptation method according to claim 1,
The detection condition includes an average value of reliability included in a speech recognition result whose speaker ID is the speaker ID, and an average value of reliability included in a speech recognition result whose speaker ID is other than the speaker ID. If the subtracted value is equal to or greater than a preset threshold value, the speaker ID is used as the speaker ID of the poor speaker. The acoustic model adaptation method according to claim 1,
The detection condition is a speech recognition result in which an average value of reliability included in a speech recognition result whose speaker ID is the speaker ID is a significance level set in advance and the speaker ID is other than the speaker ID. If the test supports that the reliability is smaller than the average value of reliability, the speaker ID is used as the speaker ID of the weak speaker. The acoustic model adaptation method according to any one of claims 1 to 3,
An acoustic model selection step,
In the acoustic model adaptation step, a plurality of adaptation parameters are preset, and a plurality of post-adaptation acoustic model candidates are output for each adaptation parameter from the pre-adaptation acoustic model and the adaptation data,
In the acoustic model selection step, the acoustic model selection unit reads all of the speech recognition results from the speech recognition result storage unit, and uses all the reliability included in the speech recognition results, so that the reliability before adaptation average value And using all the speech included in the speech recognition result and the post-adaptation acoustic model candidate, obtain an after-adaptation reliability average value, and subtract the pre-adaptation reliability average value from the after-adaptation reliability average value If the reliability decrease range is less than a predetermined reliability decrease range threshold, the post-adaptation acoustic model candidate corresponding to the post-adaptation reliability average value is output as the post-adaptation acoustic model. An acoustic model adaptation method characterized by: The acoustic model adaptation method according to any one of claims 1 to 3,
The acoustic model adaptation step stores the output post-adaptation acoustic model in the acoustic model storage unit, reads all the speech recognition results from the speech recognition result storage unit, and for all the speech included in the speech recognition results , Using the post-adaptation acoustic model, outputting speech recognition result text and reliability, storing the speech recognition result text and the reliability in the speech recognition result storage unit,
The acoustic model adaptation method, wherein the weak speaker detection step, the adaptation data selection step, and the acoustic model adaptation step are repeatedly executed until a predetermined condition is satisfied. The acoustic model adaptation method according to claim 4,
The acoustic model selection step stores the output after-adaptation acoustic model in the acoustic model storage unit, reads all the speech recognition results from the speech recognition result storage unit, and for all the speech included in the speech recognition results , Using the post-adaptation acoustic model, outputting speech recognition result text and reliability, storing the speech recognition result text and the reliability in the speech recognition result storage unit,
The acoustic model adaptation method, wherein the weak speaker detection step, the adaptation data selection step, the acoustic model adaptation step, and the acoustic model selection step are repeatedly executed until a predetermined condition is satisfied. An acoustic model storage unit for storing a pre-adaptation acoustic model;
A voice recognition result storage unit for storing a voice recognition result;
A speech recognition unit that outputs at least speech recognition result text and reliability from the input speech using the pre-adaptation acoustic model;
A speech recognition result registration unit that stores a speech recognition result including at least a speaker ID, the speech, the speech recognition result text, and the reliability in the speech recognition result storage unit;
Read all the speech recognition results from the speech recognition result storage unit, and extract a speaker ID of a speaker who is poor in speech recognition accuracy than other speakers based on preset detection conditions A detection unit;
From the speech recognition result storage unit, a speech recognition result in which the speaker ID is the speaker ID of the poor speaker and the reliability is equal to or higher than a reliability threshold value set in advance is read, and at least the speech and the speech An adaptive data selector for extracting adaptive data consisting of recognition result text;
From the pre-adaptation acoustic model and the adaptation data, an acoustic model adaptation unit that outputs a post-adaptation acoustic model using a preset adaptation parameter;
An acoustic model adaptation device comprising:   The program for functioning a computer as an acoustic model adaptation apparatus of Claim 7.

Images