WO2017166625A1

WO2017166625A1 - Acoustic model training method and apparatus for speech recognition, and electronic device

Info

Publication number: WO2017166625A1
Application number: PCT/CN2016/096672
Authority: WO
Inventors: 张俊博
Original assignee: 乐视控股（北京）有限公司; 乐视致新电子科技（天津）有限公司
Priority date: 2016-03-30
Filing date: 2016-08-25
Publication date: 2017-10-05
Also published as: CN105845130A

Abstract

Provided are an acoustic model training method and apparatus for speech recognition, and an electronic device. The method comprises: acquiring a training sample, wherein the training sample comprises an acoustic feature and a training text corresponding to the acoustic feature; acquiring an original acoustic model, and using the original acoustic model to determine an acoustic state in the original acoustic model corresponding to each training text; according to an acoustic state and an acoustic feature corresponding to each training text, determining an acoustic feature corresponding to each acoustic state; using the acoustic feature corresponding to each acoustic state to conduct retraining to obtain a state description model of the acoustic state; and using the state description model obtained by retraining to update an original state description model in the original acoustic model, thus obtaining an updated acoustic model. According to the embodiments of the present invention, the complexity of model training is reduced, and by updating the original acoustic model, the structure of the original acoustic model is not destroyed, thus ensuring the accuracy of speech recognition.

Description

Acoustic model training method, device and electronic device for speech recognition

cross reference

This application claims priority to Chinese Patent Application No. 201610195612.X, entitled "Acoustic Model Training Method and Apparatus for Speech Recognition", filed on March 30, 2016, the entire contents of which are hereby incorporated by reference. The citations are incorporated herein by reference.

Technical field

The present invention relates to the field of speech recognition technologies, and in particular, to an acoustic model training method, apparatus and electronic device for speech recognition.

Background technique

One purpose of the speech recognition system is to convert speech into text. Specifically, a speech signal is searched for a sequence of words (consisting of words or words) so that it matches the speech signal with the highest degree.

One of the most important parts of the speech recognition system is the Acoustic Modeling. When speech recognition is performed, the speech signal is converted into acoustic features, and then the acoustic model is used to determine the acoustic states corresponding to the acoustic features. Acoustic state combination is the text.

Among them, the acoustic state is the basic unit constituting the pronunciation of a word, and generally refers to a smaller unit obtained by further dividing the phoneme.

The acoustic state corresponding to the acoustic features is obtained by using a state description model in the acoustic model. In the acoustic model, each acoustic state corresponds to a state description model, and the state description model can be used to identify the acoustic state that best matches the acoustic features.

In the prior art, the training process of the acoustic model is very complicated, including not only the state description model. The training also includes the extraction of acoustic features, acoustic feature transformation, decision tree generation, and training of state definition models. In practical applications, as the application scenario changes or the life cycle evolves, the original acoustic model may not be applicable to the current application scenario. This requires retraining a new acoustic model, but retraining a new acoustic model. The complexity is high, and the inventors found in the study that some structures in the original acoustic model may not need to be changed, such as the state definition model. If retrained, it will destroy the acoustic state defined by the state definition model, but will affect it. The accuracy of speech recognition.

Summary of the invention

The solution to be solved by the embodiment of the present invention is to reduce the complexity of model training, not to destroy the structure of the original acoustic model, and to ensure the accuracy of speech recognition.

The embodiment of the invention provides an acoustic model training method and device for speech recognition, which is used to solve the technical problem of reducing the complexity of acoustic model training under the premise of ensuring the accuracy of speech recognition in the prior art.

An embodiment of the present invention provides an acoustic model training method for speech recognition, including:

Obtaining a training sample; the training sample includes an acoustic feature and a training text corresponding to the acoustic feature;

Acquiring an original acoustic model, and using the original acoustic model, determining an acoustic state in the original acoustic model corresponding to each training text;

Determining an acoustic feature corresponding to each acoustic state according to an acoustic state and an acoustic feature corresponding to each training text;

Re-training the state description model of the acoustic state using the acoustic features corresponding to each acoustic state;

Using the state description model obtained by retraining, the original state description model in the original acoustic model is updated to obtain an updated acoustic model.

An embodiment of the present invention provides an acoustic model training apparatus for speech recognition, including:

a sample acquisition module, configured to acquire a training sample; the training sample includes an acoustic feature and a training text corresponding to the acoustic feature;

a first determining module, configured to acquire an original acoustic model, and use the original acoustic model to determine an acoustic state in the original acoustic model corresponding to each training text;

a second determining module, configured to determine an acoustic feature corresponding to each acoustic state according to an acoustic state and an acoustic feature corresponding to each training text;

a training module, configured to re-train a state description model of the acoustic state by using an acoustic feature corresponding to each acoustic state;

And an update module, configured to update the original state description model in the original acoustic model by using a state description model obtained by retraining, and obtain an updated acoustic model.

Embodiments of the invention further disclose an electronic device comprising at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor The instructions are executed by the at least one processor to enable the at least one processor to acquire training samples; the training samples include acoustic features and training text corresponding to the acoustic features; acquiring an original acoustic model, and utilizing The original acoustic model determines an acoustic state in the original acoustic model corresponding to each training text; determining an acoustic feature corresponding to each acoustic state according to an acoustic state and an acoustic feature corresponding to each training text; utilizing each acoustic The acoustic feature corresponding to the state is retrained to obtain a state description model of the acoustic state; the state description model obtained by retraining is used to update the original state description model in the original acoustic model to obtain an updated acoustic model.

The present invention also discloses a non-volatile computer storage medium, wherein the storage medium stores computer-executable instructions for causing the computer to perform the method of claims 1-5 .

The embodiment of the invention further provides a computer program product, the computer program product comprising A computer program stored on a non-transitory computer readable storage medium, the computer program comprising program instructions that, when executed by a computer, cause the computer to perform the method of claims 1-5.

The acoustic model training method and apparatus for speech recognition provided by the embodiments of the present invention determine the acoustic state corresponding to the training text in the training sample by using the original acoustic model, and can determine according to the acoustic state and acoustic characteristics corresponding to each training text. Acoustic features corresponding to each acoustic state. Therefore, the state description model of the acoustic state can be retrained directly by using the acoustic features corresponding to the acoustic state, and the state description model obtained by retraining is used to update the original state description model in the original acoustic model, so that after the update can be obtained Acoustic model. With the embodiment of the present invention, only the state description model in the original acoustic model can be retrained without training a new acoustic model, which not only reduces the training complexity, but also updates the original acoustic model without destroying the original. The structure of the acoustic model ensures the accuracy of speech recognition.

DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, a brief description of the drawings used in the embodiments or the prior art description will be briefly described below. Obviously, the drawings in the following description It is a certain embodiment of the present invention, and other drawings can be obtained from those skilled in the art without any creative work.

1 is a flow chart of an embodiment of an acoustic model training method for speech recognition according to the present invention.

2 is a schematic structural view of an embodiment of an acoustic model training device for speech recognition according to the present invention;

FIG. 3 is a schematic structural diagram of hardware of an electronic device according to an embodiment of the present invention.

detailed description

The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings, obviously, The described embodiments are a part of the embodiments of the invention, and not all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts are within the scope of the present invention.

In the description of the present invention, it is to be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inside", "outside", etc. The orientation or positional relationship of the indications is based on the orientation or positional relationship shown in the drawings, and is merely for the convenience of the description of the invention and the simplified description, rather than indicating or implying that the device or component referred to has a specific orientation, in a specific orientation. The construction and operation are therefore not to be construed as limiting the invention. Moreover, the terms "first," "second," and "third" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance.

In the description of the present invention, it should be noted that the terms "installation", "connected", and "connected" are to be understood broadly, and may be fixed or detachable, for example, unless otherwise explicitly defined and defined. Connection, or integral connection; may be mechanical connection or electrical connection; may be directly connected, may also be indirectly connected through an intermediate medium, or may be internal communication of two components, may be wireless connection, or may be wired connection. The specific meaning of the above terms in the present invention can be understood in a specific case by those skilled in the art.

The technical solution of the present invention is mainly applied to the training of acoustic models in the field of speech recognition. The acoustic model is one of the most important parts of the speech recognition system. For the speech recognition, the acoustic model is used to determine the respective acoustic states corresponding to the acoustic features, and the text is obtained by combining the various acoustic states. The acoustic feature is obtained by extracting the speech signal, and the acoustic feature may be, for example, a MFCC (Mel Frequency Cepstrum Coefficient) feature.

Acoustic models are obtained by modeling models such as Hidden Markov Models, and a large number of training samples are needed to train the modeling models to obtain acoustic models.

The acoustic model includes a plurality of state description models corresponding to the acoustic states, the state description models are used to calculate the probability of the acoustic features on the acoustic state, and the acoustic states that are most matched with the acoustic features are determined, thereby combining the acoustic states You can get the text.

Due to the acoustic model training, not only the training of the state description model, but also the training of the state definition model, through the state definition model, a text or a phoneme in a specific context can be obtained, which state sequence should correspond to; Describe the model, and the probability that the acoustic feature is in a certain acoustic state can be obtained, so that the acoustic state that best matches the acoustic feature can be determined.

If the original acoustic model is not applicable to the current application scenario, this requires retraining a new acoustic model, but retraining a new acoustic model is not only highly complex, but the inventors found it in the study. In some application scenarios, it is not desirable to change the state definition model. It is only necessary to change the state description model. For example, if the state definition model is changed, the decoding map used in the recognition needs to be reconstructed, but it is not desirable to reconstruct the decoding map and use The original training data for training the original model is not available for a variety of reasons (eg, the original training data is confidential). If a new acoustic model is retrained, the recognition accuracy of the new acoustic model obtained by retraining may be lower than the recognition accuracy of the original acoustic model.

Therefore, the present invention proposes a technical solution that does not destroy the structure of the original acoustic model under the premise of ensuring the accuracy of speech recognition by updating the original acoustic model.

In the embodiment of the present invention, the training sample and the original acoustic model are acquired, and the acoustic state corresponding to the training text in the training sample is determined by using the original acoustic model, and each acoustic can be determined according to the acoustic state and the acoustic characteristics corresponding to each training text. The acoustic characteristics corresponding to the state. Therefore, the state description model of the acoustic state can be retrained directly by using the acoustic features corresponding to the acoustic state, and the state description model obtained by retraining is used to update the original state description model in the original acoustic model, so that after the update can be obtained Acoustic model. The updated acoustic model can continue to be used for speech recognition. With the embodiment of the present invention, only the state description model in the original acoustic model can be retrained without training a new acoustic model, which reduces the training complexity. And by updating the original acoustic model, the structure of the original acoustic model is not destroyed, and the accuracy of speech recognition is ensured.

The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

Example 1

FIG. 1 is a flowchart of an embodiment of an acoustic model training method for speech recognition according to an embodiment of the present invention, which may include the following steps:

101: Obtain a training sample.

Wherein the training sample includes an acoustic feature and a training text corresponding to the acoustic feature.

Wherein, the acoustic features in the training samples may be obtained by extracting from training speech.

That is, specifically, by acquiring the training speech and the training text corresponding to the training speech, and extracting the acoustic features of the training speech, the acoustic features and their corresponding training texts are used as training samples.

In the embodiment of the present invention, since it is not necessary to retrain a new acoustic model, the number of training samples can be greatly reduced.

The training sample can be determined according to the application scenario in which the original acoustic model is no longer applicable. Since the application scenario is no longer applicable, the acoustic model cannot recognize the exact text of the corresponding application scenario, and therefore needs to be retrained.

102: Acquire an original acoustic model, and use the original acoustic model to determine an acoustic state in the original acoustic model corresponding to each training text.

Specifically, the state of the original acoustic model may be used to define a model to determine an acoustic state corresponding to each training text. The acoustic state corresponding to each training text includes a plurality, that is, it corresponds to a sequence of acoustic states.

The state definition model is used to determine a word or a phoneme in a specific context, and its corresponding acoustic state training.

The state description model is used to determine the probability of the acoustic feature in an acoustic state when the acoustic feature is given, thereby determining the sequence of acoustic states that best match the acoustic feature.

Wherein, the acoustic state is a basic unit constituting a utterance of a text, and may refer to a unit smaller than a phoneme obtained by further dividing a phoneme. Combine acoustic states to obtain phonemes, and combine phonemes to get text.

The training text is composed of words, so the original acoustic model corresponding to each training text can be obtained. The sequence of acoustic states in the model.

103: Determine an acoustic feature corresponding to each acoustic state according to an acoustic state and an acoustic feature corresponding to each training text.

According to the correspondence between the training text and the acoustic features, the acoustic state corresponding to the training text can be obtained, that is, the acoustic state corresponding to the acoustic features can be obtained.

As another embodiment, specifically, the acoustic features may be segmented according to an acoustic state and an acoustic feature corresponding to each training text to obtain segmentation data;

The segmentation data corresponding to each acoustic state is determined.

That is, each acoustic state specifically corresponds to segmentation data in the acoustic features.

Specifically, the acoustic feature may be forcibly aligned by the original acoustic model, specifically, the training text is used to construct a linear decoding network, and the Viterbi algorithm is used to perform the acoustic features corresponding to the training file. Segmentation is performed to obtain segmentation data so that segmentation data corresponding to each acoustic state can be determined.

104: Re-training the state description model of the acoustic state using the acoustic features corresponding to each acoustic state.

Specifically, the state description model of the acoustic state is retrained by using segmentation data corresponding to each acoustic state.

105: Using the state description model obtained by retraining, updating the original state description model in the original acoustic model to obtain an updated acoustic model.

The state description model obtained by retraining can replace the original state description model in the original acoustic model, and the other structures of the original acoustic model are unchanged, thereby obtaining the updated acoustic model. The updated acoustic model is to re-train the acoustic model according to the training sample, and the training sample is a sample of the applicable application scene, so that the acoustic model obtained by the retraining is applicable to the speech recognition of the application scene.

In this embodiment, only the state description model in the original acoustic model can be retrained without training a new acoustic model, which reduces the training complexity and passes the original acoustic model. The type is updated without destroying the structure of the original acoustic model, and the structure of the original acoustic model can be used to define the model and other structures to ensure the accuracy of speech recognition.

In the embodiment of the present invention, in the case that the original training data is lost and the state definition model in the original acoustic model is not desired to be changed, the state model can be retrained by retraining the acoustic model in the original acoustic model. Training complexity can be reduced while the acoustic model can be adapted to the current application scenario.

The state description model can be obtained by training deep neural networks (DNN), and can be implemented by a Back Propgation algorithm. Of course, other mathematical models, such as a mixed Gaussian model (GMM), can also be used. , Gaussian Mixture Model), using the EM (Expectation Maximization Algorithm) algorithm implementation, etc., can be selected according to the actual situation, the present invention does not limit this.

Example 2

FIG. 2 is a schematic structural diagram of an embodiment of an acoustic model training apparatus for voice recognition according to an embodiment of the present disclosure, where the apparatus may include:

The sample obtaining module 201 is configured to acquire a training sample.

The sample acquisition module may specifically acquire training speech and training text, and extract acoustic features of the training speech.

The training sample can be determined according to an application scenario in which the original acoustic model is no longer applicable. Since the application scenario is no longer applicable, the acoustic model cannot recognize the exact text of the corresponding application scenario, so Retraining.

The first determining module 202 is configured to acquire an original acoustic model, and use the original acoustic model to determine an acoustic state in the original acoustic model corresponding to each training text.

Specifically, the first determining module may be configured to determine an acoustic state corresponding to each training text by using a state definition model of the original acoustic model. The acoustic state corresponding to each training text includes a plurality, that is, it corresponds to a sequence of acoustic states.

The training text consists of text, so a sequence of acoustic states in the original acoustic model corresponding to each training text can be obtained.

The second determining module 203 is configured to determine an acoustic feature corresponding to each acoustic state according to an acoustic state and an acoustic feature corresponding to each training text.

In another embodiment, the second determining module may be specifically configured to segment the acoustic features according to an acoustic state and an acoustic feature corresponding to each training text, obtain segmentation data, and determine each acoustic state correspondingly. Segmentation data.

Specifically, the acoustic feature may be forcibly aligned by the original acoustic model, the linear decoding network may be constructed by using the training text, and the acoustic features corresponding to the training file may be segmented by using the Viterbi algorithm. Obtain segmentation data to determine Segmentation data corresponding to each acoustic state.

Therefore, as a further embodiment, the second determining module establishes a linear decoding network by using each training text, and uses a Viterbi algorithm to segment the acoustic features corresponding to the training file to obtain segment data, and determine Segmentation data corresponding to each acoustic state.

The training module 204 is configured to retrain the state description model of the acoustic state by using the acoustic features corresponding to each acoustic state.

The updating module 205 is configured to update the original state description model in the original acoustic model by using a state description model obtained by retraining, and obtain an updated acoustic model.

In this embodiment, only the state description model in the original acoustic model can be retrained without training a new acoustic model, which reduces the training complexity and does not destroy the original acoustics by updating the original acoustic model. The structure of the model can continue to use the structure of the original acoustic model to define the model and other structures to ensure the accuracy of speech recognition.

Example 3

As shown in FIG. 3, an embodiment of the present invention further discloses an electronic device including at least one processor 810; and a memory 800 communicably connected to the at least one processor 810; wherein the memory 800 is stored An instruction executed by the at least one processor 810, the instructions being executed by the at least one processor 810 to enable the at least one processor 810 to acquire training samples; the training samples including acoustic features and the acoustics Training text corresponding to the feature; An original acoustic model, and using the original acoustic model, determining an acoustic state in the original acoustic model corresponding to each training text; determining an acoustic corresponding to each acoustic state according to an acoustic state and an acoustic feature corresponding to each training text Feature; re-training the state description model of the acoustic state by using the acoustic features corresponding to each acoustic state; updating the original state description model in the original acoustic model by using the state description model obtained by retraining, and obtaining the updated state Acoustic model. The electronic device also includes an input device 830 and an output device 840 that are electrically coupled to the memory 800 and the processor, the electrical connections preferably being connected by a bus.

In the electronic device of this embodiment, preferably, determining the acoustic characteristics corresponding to each acoustic state according to the acoustic state and the acoustic characteristics corresponding to each training text comprises: according to the acoustic state and acoustic characteristics corresponding to each training text, Separating the acoustic features to obtain segmentation data; determining segmentation data corresponding to each acoustic state; and using the acoustic features corresponding to each acoustic state, retraining to obtain a state description model for each acoustic state includes: utilizing The segmentation data corresponding to each acoustic state is retrained to obtain a state description model of the acoustic state.

In the electronic device of this embodiment, preferably, the acquiring the training sample comprises: acquiring training speech and training text, and extracting an acoustic feature of the training speech.

In the electronic device of this embodiment, preferably, determining, by using the original acoustic model, determining an acoustic state in the original acoustic model corresponding to each training text comprises: determining, by using a state definition model in the original acoustic model, The acoustic state corresponding to each training text.

In the electronic device of the embodiment, preferably, the acoustic feature is segmented according to an acoustic state and an acoustic feature corresponding to each training text to obtain segmentation data; and determining segment data corresponding to each acoustic state includes : Using each training text to establish a linear decoding network, using the Viterbi algorithm, segmenting the acoustic features corresponding to the training file, obtaining segmentation data, and determining segmentation data corresponding to each acoustic state.

Example 4

Embodiments of the present invention also disclose a non-volatile computer storage medium, wherein the storage The medium stores computer executable instructions that, when executed by the electronic device, enable the electronic device to: acquire training samples; the training samples include acoustic features and training text corresponding to the acoustic features; and acquire an original acoustic model And using the original acoustic model to determine an acoustic state in the original acoustic model corresponding to each training text; determining an acoustic feature corresponding to each acoustic state according to an acoustic state and an acoustic feature corresponding to each training text; Acoustic features corresponding to each acoustic state are retrained to obtain a state description model of the acoustic state; using a state description model obtained by retraining, the original state description model in the original acoustic model is updated to obtain an updated acoustic model.

In the storage medium of this embodiment, preferably, determining the acoustic characteristics corresponding to each acoustic state according to the acoustic state and the acoustic characteristics corresponding to each training text comprises: according to the acoustic state and acoustic characteristics corresponding to each training text, Separating the acoustic features to obtain segmentation data; determining segmentation data corresponding to each acoustic state; and using the acoustic features corresponding to each acoustic state, retraining to obtain a state description model for each acoustic state includes: utilizing The segmentation data corresponding to each acoustic state is retrained to obtain a state description model of the acoustic state.

In the storage medium of this embodiment, preferably, the acquiring the training samples comprises: acquiring training speech and training text, and extracting acoustic features of the training speech.

In the storage medium of the embodiment, preferably, determining, by using the original acoustic model, determining an acoustic state in the original acoustic model corresponding to each training text comprises: determining, by using a state definition model in the original acoustic model, The acoustic state corresponding to each training text.

The storage medium of the embodiment, preferably, the acoustic feature is segmented according to an acoustic state and an acoustic feature corresponding to each training text to obtain segmentation data; and determining segment data corresponding to each acoustic state includes : Using each training text to establish a linear decoding network, using the Viterbi algorithm, segmenting the acoustic features corresponding to the training file, obtaining segmentation data, and determining segmentation data corresponding to each acoustic state.

Example 5

The embodiment of the invention further provides a computer program product, the computer program product comprising A computer program stored on a non-transitory computer readable storage medium, the computer program comprising program instructions that, when executed by a computer, cause the computer to perform the method described in the above embodiments.

The state description model can be obtained by training deep neural networks (DNN), and can be implemented by a Back Propgation algorithm. Of course, other mathematical models, such as a mixed Gaussian model (GMM), can also be used. , Gaussian Mixture Model), using the EM (Expectation Maximization Algorithm) algorithm to achieve, etc., can be selected according to the actual situation.

Those skilled in the art will appreciate that embodiments of the present invention can be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or a combination of software and hardware. Moreover, the invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) including computer usable program code.

The present invention has been described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (system), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flowchart illustrations and/or FIG. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing device to produce a machine for the execution of instructions for execution by a processor of a computer or other programmable data processing device. The work specified in one or more blocks of a flow or a flow and/or a block diagram of a flowchart Able device.

The computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture comprising the instruction device. The apparatus implements the functions specified in one or more blocks of a flow or a flow and/or block diagram of the flowchart.

These computer program instructions can also be loaded onto a computer or other programmable data processing device such that a series of operational steps are performed on a computer or other programmable device to produce computer-implemented processing for execution on a computer or other programmable device. The instructions provide steps for implementing the functions specified in one or more of the flow or in a block or blocks of a flow diagram.

It is apparent that the above-described embodiments are merely illustrative of the examples, and are not intended to limit the embodiments. Other variations or modifications of the various forms may be made by those skilled in the art in light of the above description. There is no need and no way to exhaust all of the implementations. Obvious changes or variations resulting therefrom are still within the scope of the invention.

Claims

An acoustic model training method for speech recognition, comprising:

Obtaining a training sample; the training sample includes an acoustic feature and a training text corresponding to the acoustic feature;

Acquiring an original acoustic model, and using the original acoustic model, determining an acoustic state in the original acoustic model corresponding to each training text;

Determining an acoustic feature corresponding to each acoustic state according to an acoustic state and an acoustic feature corresponding to each training text;

Re-training the state description model of the acoustic state using the acoustic features corresponding to each acoustic state;

Using the state description model obtained by retraining, the original state description model in the original acoustic model is updated to obtain an updated acoustic model.
The method according to claim 1, wherein the determining the acoustic characteristics corresponding to each acoustic state according to the acoustic state and the acoustic characteristics corresponding to each training text comprises:

Separating the acoustic features according to an acoustic state and an acoustic feature corresponding to each training text to obtain segmentation data;

Determining segmentation data corresponding to each acoustic state;

The state description model for re-training each acoustic state using the acoustic features corresponding to each acoustic state includes:

The state description model of the acoustic state is retrained using segmentation data corresponding to each acoustic state.
The method of claim 1 wherein said obtaining training samples comprises:

The training speech and the training text are obtained, and the acoustic features of the training speech are extracted.
The method according to claim 1, wherein the determining, by the original acoustic model, the acoustic state in the original acoustic model corresponding to each training text comprises:

Determining the sound corresponding to each training text by using the state definition model in the original acoustic model Learning state.
The method according to claim 2, wherein the acoustic features are segmented according to an acoustic state and an acoustic feature corresponding to each training text to obtain segmentation data; and a score corresponding to each acoustic state is determined. Segment data includes:

A linear decoding network is established by using each training text, and the acoustic characteristics corresponding to the training file are segmented by using a Viterbi algorithm to obtain segment data, and segment data corresponding to each acoustic state is determined.
An acoustic model training device for speech recognition, comprising:

a sample acquisition module, configured to acquire a training sample; the training sample includes an acoustic feature and a training text corresponding to the acoustic feature;

a first determining module, configured to acquire an original acoustic model, and use the original acoustic model to determine an acoustic state in the original acoustic model corresponding to each training text;

a second determining module, configured to determine an acoustic feature corresponding to each acoustic state according to an acoustic state and an acoustic feature corresponding to each training text;

a training module, configured to re-train a state description model of the acoustic state by using an acoustic feature corresponding to each acoustic state;

And an update module, configured to update the original state description model in the original acoustic model by using a state description model obtained by retraining, and obtain an updated acoustic model.
The device according to claim 6, wherein the second determining module is specifically configured to:

Separating the acoustic features according to an acoustic state and an acoustic feature corresponding to each training text, obtaining segmentation data, and determining segmentation data corresponding to each acoustic state;

The training module is specifically configured to:

The state description model of the acoustic state is retrained using segmentation data corresponding to each acoustic state.
The apparatus according to claim 6, wherein said sample acquisition module is specifically used to:

The training speech and the training text are obtained, and the acoustic features of the training speech are extracted.
The device according to claim 6, wherein the first determining module is specifically configured to:

The acoustic state corresponding to each training text is determined using a state definition model in the original acoustic model.
The device according to claim 7, wherein the second determining module is specifically configured to:

A linear decoding network is established by using each training text, and the acoustic characteristics corresponding to the training file are segmented by using a Viterbi algorithm to obtain segment data, and segment data corresponding to each acoustic state is determined.
An electronic device, comprising: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions Executed by the at least one processor to enable the at least one processor to:

Obtaining a training sample; the training sample includes an acoustic feature and a training text corresponding to the acoustic feature;

Acquiring an original acoustic model, and using the original acoustic model, determining an acoustic state in the original acoustic model corresponding to each training text;

Determining an acoustic feature corresponding to each acoustic state according to an acoustic state and an acoustic feature corresponding to each training text;

Re-training the state description model of the acoustic state using the acoustic features corresponding to each acoustic state;

Using the state description model obtained by retraining, the original state description model in the original acoustic model is updated to obtain an updated acoustic model.
A non-volatile computer storage medium characterized in that the storage medium stores computer executable instructions for causing the computer to perform the method of claims 1-5.
A computer program product comprising a computer program stored on a non-transitory computer readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, The computer performs the method of claims 1-5.