JP2011095680A - Acoustic model adaptation device, acoustic model adaptation method and program for acoustic model adaptation - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To efficiently perform acoustic model adaptation with high accuracy in a limited period of time. <P>SOLUTION: An acoustic model adaptation device includes: a clustering section 12 for clustering a voice signal divided by a dividing section 11 according to acoustic difference; a reliability calculation section 13 for calculating reliability for a voice signal included in a cluster; a label estimation section 14 for obtaining an estimation label by recognizing a voice signal included in the cluster; a presentation section 15 for presenting the voice signal and the estimation label included in a first cluster which is selected based on the reliability; a teacher label obtaining section 16 for obtaining a teacher label to the presented voice signal; a transition instruction section 17 which instructs the presentation section 15 to shift from a state for dealing the first cluster to a state for dealing a second cluster which is different from the first cluster; and an acoustic model adaptation section 18 for adapting the acoustic model to the voice signal in the cluster by using the teacher label. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to an acoustic model adaptation device, an acoustic model adaptation method, and an acoustic model adaptation program, and in particular, an acoustic model adaptation device, an acoustic model adaptation method, and an acoustic model adaptation method for obtaining a highly accurate speech recognition result efficiently in a limited time. The present invention relates to an acoustic model adaptation program.

  Generally, in speech recognition, an acoustic model for expressing phoneme features and a language model for expressing restrictions on how phonemes are arranged are used. Then, “acoustic model adaptation” may be performed to adapt the acoustic model to the speaker and the environment.

  An example of an acoustic model adaptation apparatus is described in Patent Document 1. The system described in Patent Literature 1 presents a plurality of speech recognition result candidates to the user, and causes the user to select a correct answer from the candidates. As a result, the user can efficiently correct an erroneous speech recognition result. In other words, the system described in Patent Document 1 is a system that can adapt an acoustic model with supervision using an efficiently corrected recognition result.

  FIG. 8 is a block diagram showing a configuration of a voice recognition system disclosed in Patent Document 1. As shown in FIG. With reference to FIG. 8, the structure and operation | movement for correcting a recognition error word are demonstrated. The speech recognition system 400 shown in FIG. 8 includes speech input means 403, speech recognition means 405, data storage means 412 that stores a dictionary, word correction means 409, and recognition result display means 407.

  The voice recognition unit 405 includes a continuous sound determination unit 413 and a voice recognition execution unit 411.

  In the word correction unit 409, the competitive word display command unit 415 selects one or more competitive words having a competitive probability close to the competitive probability of the word having the highest competitive probability from among the competitive candidates, and corresponds the selected competitive word. It is displayed on the screen of the recognition result display means 407 adjacent to the word having the highest competition probability. The competing word selection means 417 selects an appropriate correction word from one or more competing words displayed on the screen in accordance with a manual operation by the user. The word replacement command unit 419 commands the speech recognition unit 405 to replace the corrected word selected by the competitive word selection unit 417 with the recognized word having the highest competition probability.

JP 2006-146008 A

  In the acoustic model adaptation device described in Patent Document 1, the user can efficiently correct the recognition error word, and as a result, the acoustic model adaptation device can efficiently adapt the acoustic model.

  However, the acoustic model adaptation apparatus described in Patent Document 1 may not be able to obtain the maximum acoustic model adaptation effect in a limited time. The reason is that it is not considered how much the user should correct the target (speaker, utterance environment, etc.) in applying the acoustic model. In other words, it is not considered what kind of teacher should be given.

  Therefore, an object of the present invention is to provide an acoustic model adaptation apparatus that can perform highly accurate acoustic model adaptation efficiently in a limited time.

  An acoustic model adaptation apparatus according to the present invention includes a dividing unit that divides a speech signal, a clustering unit that clusters the speech signals divided by the dividing unit according to an acoustic difference, and a speech included in a cluster created by the clustering unit. A reliability calculation unit that calculates the acoustic reliability of the signal, a label estimation unit that obtains an estimated label by recognizing a speech signal included in the cluster created by the clustering unit, and a cluster created by the clustering unit A presentation unit for presenting the user with the speech signal included in the first cluster, which is a cluster selected based on the reliability calculated by the degree calculation unit, and the estimated label obtained by the label estimation unit, and presented by the presentation unit A teacher label acquisition unit that obtains a teacher label for an audio signal, and when a predetermined condition is satisfied, From the first state in which the speech signal and the estimated label included in the cluster are presented to the user, the second signal that presents the speech signal and the estimated label included in the second cluster different from the first cluster to the user. A transition instructing unit for instructing transition to a state, and an acoustic model adaptation unit for adapting an acoustic model to an audio signal in a cluster using the teacher label acquired by the teacher label acquisition unit.

  The acoustic model adaptation method according to the present invention divides an audio signal, clusters the divided audio signals according to acoustic differences, and calculates an acoustic reliability of the audio signals included in the cluster created by clustering. Then, an estimated label is obtained by recognizing a speech signal included in the cluster created by clustering, and the speech included in the first cluster that is a cluster selected based on the reliability from the cluster created by clustering. The signal and the estimated label are presented to the user, the teacher label for the speech signal presented to the user is obtained, and when a predetermined condition is satisfied, the speech signal and the estimated label included in the first cluster are presented to the user From the first state, the audio signal included in the second cluster different from the first cluster is estimated. Characterized in that to adapt the acoustic model to the speech signals in the cluster using to transition to a second state which presents a label to the user, the training labels obtained.

  An acoustic model adaptation program according to the present invention includes: a dividing process for dividing an audio signal into a computer; a clustering process for clustering audio signals divided by the dividing process according to an acoustic difference; and a cluster created by the clustering process. A reliability calculation process that calculates the acoustic reliability of the audio signal included in the voice signal, a label estimation process that obtains an estimated label by recognizing the audio signal included in the cluster created by the clustering process, and a clustering process A presenting process for presenting a speech signal included in the first cluster, which is a cluster selected based on the reliability calculated in the reliability calculation process, and the estimated label obtained by the label estimation unit to the user Teacher label acquisition process for obtaining a teacher label for the audio signal presented in the presentation process And when the predetermined condition is satisfied, from the first state in which the audio signal and the estimated label included in the first cluster are presented to the user, the audio signal included in the second cluster different from the first cluster; An acoustic model for adapting an acoustic model to a speech signal in a cluster using a transition instruction process for instructing a transition to a second state in which the estimated label is presented to the user and a teacher label acquired in the teacher label acquisition process And an adaptive process.

  According to the present invention, since the possibility that a user will give a large number of teacher data biased to one cluster is reduced, it is possible to perform adaptive processing of a larger number of acoustic models simply by providing relatively less teacher data. Is possible. As a result, it is possible to realize an acoustic model adaptation apparatus that can efficiently and accurately perform acoustic model adaptation in a limited time.

It is a block diagram which shows an example of embodiment of the acoustic model adaptation apparatus by this invention. It is a flowchart which shows operation | movement of an acoustic model adaptation apparatus. It is a block diagram which shows the structural example of the speech recognition system containing an acoustic model adaptation apparatus. It is a block diagram which shows the structural example of the audio | voice detection system containing an acoustic model adaptation apparatus. It is a block diagram which shows the principal part of the acoustic model adaptation apparatus by this invention. It is a block diagram which shows the structural example of a transition instruction | indication part. It is a block diagram which shows the structural example of a transition instruction | indication part. 1 is a block diagram showing a configuration of a voice recognition system described in Patent Document 1. FIG.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram illustrating a configuration example of the acoustic model adaptation apparatus of the present embodiment. The acoustic model adaptation apparatus shown in FIG. 1 includes a dividing unit 1 that divides an input audio stream into audio data, and an audio data clustering unit that clusters the divided audio data according to acoustic differences such as a speaker and an utterance environment. 2 are provided. The voice data clustering unit 2 stores the created clusters 101 _{1 to} 101 _n in the cluster storage unit 101.

In addition, the acoustic model adaptation apparatus includes a reliability calculation unit 3 that calculates an acoustic reliability of the audio data included in the clusters 101 ₁ to 101 _n using the acoustic model stored in the acoustic model storage unit 102; A label estimation unit 4 that obtains an estimated label by recognizing speech data included in the clusters 101-1 to 10-n using an acoustic model stored in the acoustic model storage unit 102, and an estimated label obtained by the label estimation unit 4 And an estimated label storage unit 103 that temporarily stores.

The acoustic model adaptation apparatus also includes a speech data estimated label presentation unit 5 that presents speech data and estimated labels of a highly reliable cluster (referred to as a first cluster) to the user, and a user for the presented speech data. The teacher label acquisition unit 6 that obtains the teacher label from and stores it in the teacher label storage unit 104, and the transition unit 7 that causes the processing target of the speech data estimated label presentation unit 5 to transition to a second cluster different from the first cluster And an acoustic model adaptation unit 8 that obtains adaptive models 105 _{1 to} 105 _n by adapting the acoustic model to the speech data in the cluster using at least a teacher label from the user.

  Next, the operation of the acoustic model adaptation apparatus of this embodiment will be described. FIG. 2 is a flowchart showing a processing procedure in the acoustic model adaptation apparatus of the present embodiment.

  Let the input audio stream be x (t). However, t is an index of time, and is 0 to T as an example. The dividing unit 1 divides the input audio stream x (t) into m pieces of audio data x1 (t1), x2 (t2),..., Xm (tm) (step S1). However, t1, t2,..., Tm are time indexes, and the range is included in the range 0 to T of t. Note that a unit such as an utterance unit or a predetermined time unit can be considered as a unit of division.

Audio data clustering unit 2, the audio data divided by the dividing unit 1, clustering to n clusters 101 ₁ to 101 _n in accordance with the acoustic differences such as speakers and speech environment (step S2). For example, the voice data clustering unit 2 automatically performs clustering based on the proximity of acoustic features. Specifically, when the characteristics of certain audio data (such as frequency spectrum and cepstrum) are similar to the characteristics of other audio data, the audio data is clustered into the same cluster. If the correspondence between the voice data and the speaker is known, clustering may be performed based on the correspondence information.

The reliability calculation unit 3 calculates the reliability of each of the clusters 101 _{1 to} 101 _n by calculating the acoustic reliability of the speech data included in the cluster for each of the clusters 101 _{1 to} 101 _n ( Step S3). As the reliability, for example, an average value of posterior probabilities and an average value of SN ratio are used.

As an example, the audio data x1 (t1) to the cluster 101 _1, x2 when (t2) are included, the audio data x1 (t1), x2 average value of (reliability) below the posterior probability of (t2) Can be calculated as follows.

  Average value of posterior probabilities = ave_ {x1, x2} <ave_ {t1} <P (k1 | x1 (t1))>, ave_ {t2} <P (k2 | x2 (t2)) >> Formula ( 1)

  In Expression (1), ave_ {a} <b> is an operator for calculating the average of b with respect to a, and P (k1 | x1 (t1)) is an acoustic when audio data x1 (t1) is given. This is the posterior probability for the probability distribution k1 in the model. However, the probability distribution k1 is a distribution with the highest posterior probability at time t1. An index other than the posterior probability and the SN ratio may be used as the reliability.

The label estimation unit 4 obtains an estimated label by recognizing speech data included in each cluster for each of the clusters 101 _{1 to} 101 _n (step S4). As estimated labels, for example, phoneme labels (a, i, u, e, o, k, s, t, n,...), Syllable labels (a, i, u, e, o, ka, sa, t , Na, ...), or a label indicating the difference between voice and noise.

  The voice data estimated label presenting unit 5 presents the voice data included in the first cluster with high reliability and the estimated label in the first cluster to the user (step S5). Specifically, it is preferable to present a voice based on the voice data to the user and display an estimated label on a display device (not shown) for the user to see.

  When the user inputs a teacher label for the presented voice data via an input device (not shown), the teacher label acquisition unit 6 acquires the input teacher label. As a result, the teacher label acquisition unit 6 obtains a correct teacher label from the user for the audio data presented to the user (step S6). The teacher label is desirably the same type of label as the estimation level handled by the label estimation unit 4, but may be one that can be converted into the same type of label. For example, since a syllable label can be converted into a phoneme label, the teacher label may be a syllable label even when the label estimation unit 4 creates a phoneme label as an estimated label.

  When the predetermined condition is satisfied, the transition unit 7 causes the processing target of the speech data estimation label presenting unit 5 to transition to a second cluster different from the first cluster having a relatively high reliability (step S7). That is, an instruction to change the processing target to the second cluster is output. Note that the second cluster is a cluster having the second highest reliability after the first cluster.

<Transition from a cluster with high reliability>
If the cluster is highly reliable, it is highly likely that the estimated label is correctly estimated. Therefore, when a small amount of teacher label is given and the degree of coincidence with the estimated label is high, the remaining speech data can be substituted with the estimated label.

  From the above, it can be said that it is possible to transition from this cluster to another cluster if the degree of coincidence of labels equal to or greater than the threshold can be confirmed. Accordingly, in the process of step S7, each time the transition unit 7 presents the estimated label to the user, the transition unit 7 calculates the degree of coincidence (similarity) between the estimated label and the teacher label obtained by the teacher label acquisition unit 6, When the degree is equal to or greater than a predetermined threshold, the processing target of the speech data estimated label presenting unit 5 is shifted to the second cluster even when there is an unpresented estimated label.

<Transition from a cluster with low reliability>
In the case of a cluster with low reliability, there are more estimation errors in the estimation label than in a cluster with high reliability. However, if the phoneme (label) coverage of the teacher label is high, the acoustic model can be applied without giving a teacher label to all speech data.

  From the above, it can be said that transition from this cluster to another cluster can be made if the phoneme (label) coverage of the teacher label equal to or greater than the threshold can be confirmed. Therefore, in the process of step S7, the transition unit 7 determines that the phoneme coverage of the teacher label obtained by the teacher label acquisition unit 6 is greater than or equal to a predetermined threshold value, even when there is an unpresented estimated label. Then, the processing target of the voice data estimated label presenting unit 5 is changed to the second cluster.

<When transitioning from a cluster that the user has determined to be unnecessary>
A cluster determined by the user that does not need to be recognized (speaker's) can be transitioned from the cluster to another cluster according to an instruction from the user without assigning a teacher label. Therefore, in the process of step S7, the transition unit 7 has entered through the input device that the user does not need to input a teacher label for the cluster that the speech data estimated label presenting unit 5 handles at that time. Even if there is an unpresented estimated label, the processing target of the speech data estimated label presenting unit 5 is shifted to the second cluster. Note that the user determines that there is no need to recognize, for example, when the voice based on the voice data matches the estimated label.

When the voice data estimation label presenting unit 5 receives an instruction from the transition unit 7 to transition the processing target to the second cluster, the voice data estimated label presenting unit 5 performs the process of step S5 with the second cluster as the processing target. Thereafter, the speech data estimated label presenting unit 5, the teacher label obtaining unit 6 and the transition unit 7 repeat the processes of steps S5 to S7, and when the processes of steps S5 and S6 are executed for all the clusters 101 _{1 to} 101 _n. The transition unit 7 determines that there is no need to transition to the next cluster.

In the above example, the cluster having the relatively high reliability (for example, the cluster having the highest reliability) among the clusters 101 _{1 to} 101 _n is set as the first cluster, and the priority is next to the cluster handled. The processing of steps S5 and S6 is executed sequentially for high clusters, but the cluster with the relatively low reliability (for example, the cluster with the lowest reliability) among the clusters 101 _{1 to} 101 _n is selected. The processing of steps S5 and S6 may be executed sequentially for the cluster having the lowest priority after the handled cluster.

The acoustic model adaptation unit 8 adapts the acoustic model 102 to the audio data included in each of the clusters 101 _{1 to} 101 _n using the teacher label 104 acquired by the teacher label acquisition unit 6, thereby adapting the adaptive model 105. _{1 to} 105 _n are obtained (step S8). The adaptive models 105 _{1 to} 105 _n are stored in the adaptive model storage unit 105.

As an algorithm for adapting an acoustic model using speech data and a teacher label, an MLLR (Maximum Likelihood Linear Regression) method, a tree structure adaptation method, or the like may be used. In this embodiment, the acoustic model adaptation is performed after acquiring the teacher labels of all the clusters 101 _{1 to} 101 _n . However, the acoustic model adaptation is performed as soon as the teacher labels of a certain cluster are acquired. Also good.

  In this embodiment, the divided speech data is clustered according to the acoustic difference, and the processing target cluster is changed when it is determined that a necessary teacher label is acquired for a certain cluster. Therefore, it is possible to adapt the acoustic model to obtain a highly accurate speech recognition result efficiently.

The acoustic model adaptation apparatus of the above embodiment can be applied to a speech recognition system. FIG. 3 is a block diagram illustrating a configuration example of a speech recognition system including the acoustic model adaptation device according to the above-described embodiment. As shown in FIG. 3, the speech recognition system 200 includes the acoustic model adaptation device 10 and the speech recognition device 20 of the above embodiment. For example, the speech recognition device 20 detects the feature of the input speech data, selects an adaptation model that matches the feature of the speech data from the adaptation models 105 _{1 to} 105 _n in the acoustic model adaptation device 10, and selects the selected adaptation. Perform speech recognition using the model.

Moreover, the acoustic model adaptation apparatus of said embodiment is applicable to an audio | voice detection system. FIG. 4 is a block diagram illustrating a configuration example of a voice detection system including the acoustic model adaptation apparatus according to the above embodiment. As shown in FIG. 4, the speech detection system 300 includes the acoustic model adaptation device 10 and the speech detection device 30 of the above embodiment. For example, the voice detection device 30 detects the feature of the input voice data, selects an adaptation model that matches the feature of the voice data from the adaptation models 105 _{1 to} 105 _n in the acoustic model adaptation device 10, and selects the selected adaptation. Using the model, a voice detection process is performed for extracting a specific voice portion from the voice data and performing speaker recognition.

  FIG. 5 is a block diagram showing a main part of the acoustic model adaptation apparatus according to the present invention. As shown in FIG. 5, the acoustic model adaptation apparatus responds to an acoustic difference between a dividing unit 11 (corresponding to the dividing unit 1 shown in FIG. 1) that divides an audio signal and an audio signal divided by the dividing unit 11. Clustering unit 12 (corresponding to the speech data clustering unit 2 shown in FIG. 1), and a reliability calculation unit 13 (FIG. 1) that calculates the acoustic reliability of speech signals included in the cluster created by the clustering unit 12. And a label estimation unit 14 (corresponding to the label estimation unit 4 shown in FIG. 1) that obtains an estimated label by recognizing a speech signal included in the cluster created by the clustering unit 12. , Included in the first cluster that is selected from the clusters created by the clustering unit 12 based on the reliability calculated by the reliability calculation unit 13. A presentation unit 15 (corresponding to the speech data estimation label presentation unit 5 shown in FIG. 1) that presents the user with the estimated signal obtained by the label estimation unit 14 and the speech signal presented by the presentation unit 15 When the teacher label acquisition unit 16 (which corresponds to the teacher label acquisition unit 6 shown in FIG. 1) that obtains the teacher label and a predetermined condition are satisfied, the presentation unit 15 receives the audio signal and the estimated label included in the first cluster. A transition instructing unit 17 that instructs to transition from the first state presented to the user to the second state presenting the audio signal and the estimated label included in the second cluster different from the first cluster to the user. (Corresponding to the transition unit 7 shown in FIG. 1) and the acoustic model adaptation unit 18 (sound shown in FIG. 1) that adapts the acoustic model to the audio signal in the cluster using the teacher label acquired by the teacher label acquisition unit 16. And a corresponding) and the model adaptation unit 8.

  The acoustic model adaptation device can also be realized by software. That is, the acoustic model adaptation apparatus has a built-in CPU, and the CPU follows the program according to the division unit 11, clustering unit 12, reliability calculation unit 13, label estimation unit 14, presentation unit 15, and teacher label acquisition shown in FIG. The functions of the unit 16, the transition instruction unit 17, and the acoustic model adaptation unit 18 may be realized.

  Further, as illustrated in FIG. 6, the transition instruction unit 17 includes a coincidence calculation unit 17 </ b> A that calculates the coincidence between the estimated label presented by the presentation unit 15 and the teacher label acquired by the teacher label acquisition unit 16. Even if the degree of coincidence calculated by the degree calculation unit 17A is greater than or equal to a predetermined value, it may be configured to include an instruction unit 17B that instructs to transition to the second state, assuming that a predetermined condition is satisfied. Good. In the case of such a configuration, the processing for the cluster can be terminated before all the estimated labels are presented, and the time required for the acoustic model adaptation can be shortened.

  As shown in FIG. 7, the transition instruction unit 17 includes a phoneme coverage calculation unit 17C that calculates the phoneme coverage of the teacher label acquired by the teacher label acquisition unit 16, and a phoneme calculated by the phoneme coverage calculation unit 17C. When the coverage is equal to or higher than a predetermined value, it may be configured to include an instruction unit 17D that instructs to change to the second state, assuming that a predetermined condition is satisfied. In the case of such a configuration, the processing for the cluster can be terminated before all the estimated labels are presented, and the time required for the acoustic model adaptation can be shortened.

  Moreover, the transition instruction | indication part 17 may be comprised so that it may instruct | indicate to change to a 2nd state according to the instruction | indication input from the user. In the case of such a configuration, the transition from the first state to the second state can be made according to the user's intention, and the time required for the acoustic model adaptation can be further shortened.

  Moreover, the transition instruction | indication part 17 may be comprised so that a cluster with the next highest reliability may be made into a 2nd cluster. In the case of such a configuration, the processing of the presentation unit 15 and the transition instruction unit 17 becomes simple.

  Moreover, the transition instruction | indication part 17 may be comprised so that a cluster with the next lowest reliability after a 1st cluster may be made into a 2nd cluster. In the case of such a configuration, the processing of the presentation unit 15 and the transition instruction unit 17 becomes simple.

  The present invention can be applied to an acoustic model adaptation apparatus that can be mounted on a speech recognition system, a speech detection system, or the like.

DESCRIPTION OF SYMBOLS 1 Dividing part 2 Speech data clustering part 3 Reliability calculation part 4 Label estimation part 5 Speech data estimation label presentation part 6 Teacher label acquisition part 7 Transition part 8 Acoustic model adaptation part 10 Acoustic model adaptation apparatus 11 Dividing part 12 Clustering part 13 Reliability Degree calculation section 14 Label estimation section 15 Presentation section 16 Teacher label acquisition section 17 Transition instruction section 17A Matching degree calculation section 17B Instruction section 17C Phoneme coverage calculation section 17D Instruction section 18 Acoustic model adaptation section 20 Speech recognition apparatus 30 Speech detection apparatus 101 Cluster storage unit 101 _{1 to} 101 _n Cluster 102 Acoustic model storage unit 103 Estimated label storage unit 104 Teacher label storage unit 105 Adaptive model storage unit 105 _{1 to} 105 _n Adaptive model 200 Speech recognition system 300 Speech detection system

Claims (10)

A dividing unit for dividing the audio signal;
A clustering unit that clusters the audio signals divided by the dividing unit according to an acoustic difference;
A reliability calculation unit for calculating an acoustic reliability for the audio signal included in the cluster created by the clustering unit;
A label estimation unit that obtains an estimated label by recognizing a speech signal included in the cluster created by the clustering unit;
The speech signal included in the first cluster, which is a cluster selected based on the reliability calculated by the reliability calculation unit from the clusters created by the clustering unit, and the estimated label obtained by the label estimation unit A presentation unit to present to,
A teacher label obtaining unit for obtaining a teacher label for the audio signal presented by the presenting unit;
When a predetermined condition is satisfied, the presentation unit includes the second state different from the first cluster from the first state in which the audio signal and the estimated label included in the first cluster are presented to the user. A transition instruction unit for instructing transition to the second state in which the audio signal and the estimated label are presented to the user;
An acoustic model adaptation device comprising: an acoustic model adaptation unit adapted to adapt an acoustic model to a speech signal in a cluster using the teacher label acquired by the teacher label acquisition unit. The transition instruction section
A degree-of-match calculator that calculates the degree of match between the estimated label presented by the presenter and the teacher label acquired by the teacher label acquisition unit;
The acoustic model according to claim 1, further comprising: an instruction unit that instructs to transition to the second state when a predetermined condition is satisfied when the degree of coincidence calculated by the coincidence degree calculation unit is equal to or greater than a predetermined value. Adaptive device. The transition instruction section
A phoneme coverage calculation unit that calculates the phoneme coverage of the teacher label acquired by the teacher label acquisition unit;
An instructing unit for instructing transition to a second state when a predetermined condition is satisfied when the phoneme coverage calculated by the phoneme coverage calculation unit is equal to or greater than a predetermined value. Item 3. The acoustic model adaptation device according to Item 2. The acoustic model adaptation device according to any one of claims 1 to 3, wherein the transition instruction unit instructs the transition to the second state in accordance with an instruction input from a user. The acoustic model adaptation device according to any one of claims 1 to 4, wherein the transition instruction unit sets a cluster having the second highest reliability after the first cluster as the second cluster. The acoustic model adaptation device according to any one of claims 1 to 4, wherein the transition instruction unit sets a cluster having the second lowest reliability after the first cluster as a second cluster.   A speech recognition system including the acoustic model adaptation device according to any one of claims 1 to 6.   A speech detection system including the acoustic model adaptation device according to any one of claims 1 to 6. Split the audio signal,
Cluster the divided audio signals according to acoustic differences,
Calculate the acoustic reliability of the audio signal included in the cluster created by clustering,
Obtain an estimated label by recognizing the speech signal contained in the cluster created by clustering,
Presenting the speech signal included in the first cluster, which is a cluster selected based on the reliability, from the cluster created by clustering, and the estimated label to the user,
Obtain a teacher label for the audio signal presented to the user,
When the predetermined condition is satisfied, the voice signal and the estimated label included in the second cluster different from the first cluster are changed from the first state in which the voice signal and the estimated label included in the first cluster are presented to the user. To the second state to present to the user,
An acoustic model adaptation method for adapting an acoustic model to a speech signal in a cluster using the obtained teacher label. On the computer,
A division process for dividing the audio signal;
A clustering process for clustering the audio signals divided by the division process according to an acoustic difference;
A reliability calculation process for calculating an acoustic reliability for an audio signal included in the cluster created by the clustering process;
A label estimation process for obtaining an estimated label by recognizing a speech signal included in the cluster created by the clustering process;
A speech signal included in a first cluster that is a cluster selected based on the reliability calculated in the reliability calculation process from the clusters created in the clustering process, and an estimated label obtained by the label estimation unit; Presenting process to present to the user,
A teacher label acquisition process for obtaining a teacher label for the audio signal presented in the presentation process;
When the predetermined condition is satisfied, the voice signal and the estimated label included in the second cluster different from the first cluster are changed from the first state in which the voice signal and the estimated label included in the first cluster are presented to the user. And a transition instruction process for instructing the transition to the second state presented to the user,
An acoustic model adaptation program for executing an acoustic model adaptation process for adapting an acoustic model to an audio signal in a cluster using the teacher label acquired in the teacher label acquisition process.

Images