Speech recognition training data selection method and apparatus, terminal device, and storage medium

Abstract

The application discloses a speech recognition training data selection method and device, a terminal equipment and a storage medium, wherein the method comprises the following steps: discretizing a general corpus and a target corpus to generate a plurality of discrete label sequences, and calculating the probability distribution of the discrete label sequences through an n-gram language model; then, the difference degrees between different discrete corpora are calculated according to the probability distribution under the n-gram language model; and finally, a subset of the general corpus with the minimum difference degree from the target corpus is selected as training data. The quality of the training data selected by the application is more suitable for a target scene and is better, and the application is helpful to improve the performance of speech recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of speech recognition, and in particular to a speech recognition training data selection method and device, a terminal device and a storage medium. BACKGROUND

[0002] In an automatic speech recognition system, selecting appropriate training data is crucial to system performance. Since speech annotation work is both expensive and time-consuming, a method is needed to select a subset of data that matches the target scenario for annotation and training. Existing data selection methods mainly rely on manual selection or transcription-based statistical methods, which often cannot accurately measure the matching degree of training corpus and target corpus, resulting in low quality of selected training data. Therefore, there is an urgent need to research new training data selection methods. SUMMARY

[0003] The present application provides a speech recognition training data selection method, device, terminal device and storage medium to solve the technical problem of low quality of speech recognition training data selected by the prior art.

[0004] To solve the above technical problems, the present application provides a speech recognition training data selection method, comprising:

[0005] obtaining a general corpus and a target corpus corresponding to a target scenario; wherein the target corpus comprises a plurality of first speech data; the general corpus comprises a plurality of second speech data;

[0006] determining all subsets of the general corpus with a preset value, and constructing a candidate corpus set by taking each subset as an element;

[0007] discretizing each first speech data in the target corpus to generate a plurality of first discrete label sequences;

[0008] constructing a first n-gram language model according to all first discrete label sequences corresponding to the target corpus;

[0009] discretizing each second speech data in each element in the candidate corpus set to generate a plurality of second discrete label sequences;

[0010] constructing a second n-gram language model according to all second discrete label sequences corresponding to the element;

[0011] comparing all first discrete label sequences and all second discrete label sequences corresponding to the element, and randomly selecting one from all identical first discrete label sequences and second discrete label sequences as a common discrete label sequence;

[0012] computing a first probability distribution of the common discrete label sequence under the first n-gram language model; computing a second probability distribution of the common discrete label sequence under the second n-gram language model;

[0013] according to the first probability distribution and the second probability distribution, computing a corpus difference degree corresponding to the element;

[0014] determining a target element with the minimum corpus difference degree, and taking all second speech data in the target element as training data required by the target scene.

[0015] As a preferred solution, the discretization processing on each first speech data in the target corpus to generate a plurality of first discrete label sequences comprises:

[0016] For each first speech data, a first frame-level acoustic feature sequence is generated by performing feature extraction on the first speech data through a convolutional neural network;

[0017] all first feature vectors in the first frame-level acoustic feature sequence are clustered;

[0018] For each first feature vector, a first discrete label of the first feature vector is calculated according to the first feature vector and the corresponding cluster center;

[0019] The first discrete label sequence is constructed according to the first discrete labels of all first feature vectors.

[0020] As a preferred solution, the discretization processing on each second speech data in the element to generate a plurality of second discrete label sequences comprises:

[0021] For each second speech data, a second frame-level acoustic feature sequence is generated by performing feature extraction on the second speech data through a convolutional neural network;

[0022] all second feature vectors in the second frame-level acoustic feature sequence are clustered;

[0023] For each second feature vector, a second discrete label of the second feature vector is calculated according to the second feature vector and the corresponding cluster center;

[0024] The second discrete label sequence is constructed according to the second feature vector and the second discrete label.

[0025] As a preferred solution, the computing of the first probability distribution of the common discrete label sequence under the first n-gram language model comprises:

[0026] counting occurrence times of a first unit sequence of length n-1 in all the first discrete label sequences;

[0027] counting occurrence times of a second unit sequence of length n in all the first discrete label sequences for each common discrete acoustic unit in the common discrete label sequence, wherein the last discrete acoustic unit of the second unit sequence is the common discrete acoustic unit;

[0028] calculating a first n-gram conditional probability of the common discrete acoustic unit by an n-gram conditional probability calculation formula according to the occurrence times of the first unit sequence and the occurrence times of the second unit sequence;

[0029] aggregating the first n-gram conditional probabilities of all the common discrete acoustic units to obtain the first probability distribution.

[0030] As a preferred solution, the calculating the second probability distribution of the common discrete label sequence under the second n-gram language model comprises:

[0031] counting occurrence times of a third unit sequence of length n-1 in all the second discrete label sequences in the element;

[0032] counting occurrence times of a fourth unit sequence of length n in all the second discrete label sequences in the element for each common discrete acoustic unit in the common discrete label sequence, wherein the last discrete acoustic unit of the fourth unit sequence is the common discrete acoustic unit;

[0033] calculating a second n-gram conditional probability of the common discrete acoustic unit by the n-gram conditional probability calculation formula according to the occurrence times of the third unit sequence and the occurrence times of the fourth unit sequence;

[0034] aggregating the second n-gram conditional probabilities of all the common discrete acoustic units to obtain the second probability distribution.

[0035] As a preferred solution, the n-gram conditional probability calculation formula is:

[0036] ;

[0037] In the formula, denotes an n-gram conditional probability; denotes the occurrence times of ; denotes the occurrence times of ; is a common discrete acoustic unit.

[0038] As a preferred solution, the formula for calculating the corpus difference degree is:

[0039] ;

[0040] In the formula, represents the corpus difference degree; represents the first probability distribution; represents the second probability distribution.

[0041] On the basis of the above embodiment, another embodiment of the present application provides a speech recognition training data selection device, comprising a data acquisition module, a discretization module, a difference degree calculation module and a training data selection module.

[0042] The data acquisition module is used to intentionally acquire a general corpus and a target corpus corresponding to a target scene; wherein the target corpus comprises a plurality of first speech data; the general corpus comprises a plurality of second speech data; a subset of the general corpus with a preset value is determined, and each subset is used as an element to construct a candidate corpus set;

[0043] The discretization module is used to discretize each first speech data in the target corpus to generate a plurality of first discrete label sequences;

[0044] The difference degree calculation module is used to construct a first n-gram language model according to all first discrete label sequences corresponding to the target corpus;

[0045] The discretization module is also used to discretize each second speech data in each element in the candidate corpus set to generate a plurality of second discrete label sequences;

[0046] The difference degree calculation module is also used to construct a second n-gram language model according to all second discrete label sequences corresponding to each element in the candidate corpus set; compare all first discrete label sequences and all second discrete label sequences corresponding to the element, randomly select one from all identical first discrete label sequences and second discrete label sequences as a common discrete label sequence; calculate the first probability distribution of the common discrete label sequence under the first n-gram language model; calculate the second probability distribution of the common discrete label sequence under the second n-gram language model; calculate the corpus difference degree corresponding to the element according to the first probability distribution and the second probability distribution;

[0047] The training data selection module is configured to determine a target element with the minimum corpus difference, and use all second voice data in the target element as the required training data of the target scene.

[0048] On the basis of the above-mentioned embodiments, a terminal device is provided in another embodiment of the application, which comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, and the processor implements the voice recognition training data selection method described in the above-mentioned embodiments of the application when executing the computer program.

[0049] On the basis of the above-mentioned embodiments, a storage medium is provided in another embodiment of the application, which comprises a stored computer program, wherein the storage medium controls a device where the storage medium is located to execute the voice recognition training data selection method described in the above-mentioned embodiments of the application when the computer program is running.

[0050] Compared with the prior art, the embodiments of the application have the following beneficial effects:

[0051] The application first discretizes the corpus to generate a plurality of discrete label sequences, then calculates the probability distribution of the discrete label sequences through an n-gram language model, and then calculates the difference between different discrete corpora according to the probability distribution under the n-gram language model. The universal corpus subset with the minimum difference from the target corpus is selected as the training data. The application measures the similarity between the universal corpus and the target corpus based on the corpus difference, realizes the screening of the training data, and the quality of the screened training data is more suitable for the target scene and better, which helps to improve the performance of the voice recognition. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a flowchart of a voice recognition training data selection method provided by an embodiment of the application;

[0053] Figure 2 is a structural diagram of a voice recognition training data selection device provided by an embodiment of the application. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0055] Embodiment One

[0056] Please refer toFigure 1 The above is a flowchart illustrating a speech recognition training data selection method according to an embodiment of the present invention, comprising:

[0057] S1. Obtain a general corpus and a target corpus corresponding to the target scene; wherein, the target corpus includes several first speech data; the general corpus includes several second speech data.

[0058] It should be noted that a general corpus refers to a large-scale speech dataset containing a wide range of speech features and diverse speech characteristics, used to train basic speech models; while a target corpus refers to a speech dataset in a specific domain or application scenario, reflecting the specific speech features and language patterns of the target task.

[0059] The target scenario refers to the scenario where speech recognition needs to be applied.

[0060] S2. Determine all subsets of the general corpus with a preset size, and construct a candidate corpus set using each subset as an element.

[0061] S3. Discretize each first speech data in the target corpus to generate several first discrete label sequences.

[0062] In a preferred embodiment, the discretization of each first speech data in the target corpus to generate a plurality of first discrete label sequences includes:

[0063] For each of the first speech data, features are extracted from the first speech data using a convolutional neural network to generate a first frame-level acoustic feature sequence;

[0064] Cluster all first feature vectors in the first frame-level acoustic feature sequence;

[0065] For each of the first feature vectors, a first discrete label of the first feature vector is calculated based on the first feature vector and the corresponding cluster center.

[0066] Construct the first discrete label sequence based on the first discrete labels of all first feature vectors.

[0067] In this embodiment, for each piece of first voice data, the first voice data is first processed... ,express The speech data sampled at each time step represents the total number of time steps. Feature extraction is performed using a convolutional neural network. Generate frame-level acoustic features ,in ,express Moment 3D feature vectors. The feature extraction process can be represented as: Subsequently, the K-means clustering algorithm was used to cluster all the first feature vectors in the first frame-level acoustic feature sequence to obtain discrete labels. : ,in, Let the i-th cluster center be denoted as . , The number of clusters is a predefined number.

[0068] S4. Construct a first n-gram language model based on all the first discrete label sequences corresponding to the target corpus.

[0069] It should be noted that the n-gram language model constructed in this invention is a Markov probability model. The n-gram model is a statistical model that is based on the Markov assumption and models the probability of the nth element appearing given the first n-1 elements.

[0070] S5. For each element in the candidate corpus set, discretize each second speech data in the element to generate several second discrete label sequences.

[0071] In a preferred embodiment, the discretization of each second speech data in the elements to generate a plurality of second discrete label sequences includes:

[0072] For each of the second speech data, features are extracted from the second speech data using a convolutional neural network to generate a second frame-level acoustic feature sequence;

[0073] Cluster all second feature vectors in the second frame-level acoustic feature sequence;

[0074] For each of the second feature vectors, a second discrete label is calculated based on the second feature vector and the corresponding cluster center.

[0075] Based on the second feature vector and the second discrete label, construct the second discrete label sequence.

[0076] S6. Construct a second n-gram language model based on all the second discrete label sequences corresponding to the elements.

[0077] S7. Compare all the first discrete label sequences and all the second discrete label sequences corresponding to the element, and randomly select one from all the same first discrete label sequences and second discrete label sequences as the common discrete label sequence.

[0078] In step S7, by comparing all the first discrete label sequences and the second discrete label sequences, the first discrete label sequences and the second discrete label sequences that are the same are first identified. There may be multiple such sequences, and one of them is randomly selected as the common discrete label sequence. The common discrete label sequence is both a first discrete label sequence and a second discrete label sequence.

[0079] S8. Calculate the first probability distribution of the shared discrete label sequence under the first n-gram language model; calculate the second probability distribution of the shared discrete label sequence under the second n-gram language model.

[0080] In a preferred embodiment, calculating the first probability distribution of the shared discrete label sequence under the first n-gram language model includes:

[0081] Count the number of occurrences of the first unit sequence of length n-1 in all first discrete label sequences;

[0082] For each shared discrete acoustic unit in the shared discrete tag sequence, count the number of occurrences of the second unit sequence of length n in all first discrete tag sequences; wherein, the last discrete acoustic unit of the second unit sequence is the shared discrete acoustic unit;

[0083] Based on the occurrence counts of the first unit sequence and the second unit sequence, the first n-gram conditional probability of the shared discrete acoustic unit is calculated using the n-gram conditional probability calculation formula.

[0084] The first probability distribution is obtained by aggregating the first n-gram conditional probabilities of all shared discrete acoustic units.

[0085] In a preferred embodiment, calculating the second probability distribution of the shared discrete label sequence under the second n-gram language model includes:

[0086] Count the number of occurrences of the third unit sequence of length n-1 in all the second discrete label sequences of the element;

[0087] For each shared discrete acoustic unit in the shared discrete tag sequence, count the number of occurrences of the fourth unit sequence of length n in all second discrete tag sequences of the element; wherein, the last discrete acoustic unit of the fourth unit sequence is the shared discrete acoustic unit.

[0088] Based on the occurrence counts of the third unit sequence and the fourth unit sequence, the second n-gram conditional probability of the shared discrete acoustic unit is calculated using the n-gram conditional probability calculation formula.

[0089] The second probability distribution is obtained by aggregating the second n-gram conditional probabilities of all shared discrete acoustic units.

[0090] In a preferred embodiment, the formula for calculating the n-gram conditional probability is:

[0091] ;

[0092] In the formula, Represents the conditional probability of an n-gram; express of Number of occurrences; Indicated Number of occurrences; It consists of a shared discrete acoustic unit.

[0093] In this embodiment, the n-gram model is used to perform probability modeling on each speech segment in the discrete corpus, obtaining the probability distribution of the discrete corpus: First, all speech signals in the corpus are converted into discrete unit sequences, resulting in the form of... The discrete representation of, where each This corresponds to a discrete acoustic unit; subsequently, an n-gram language model is constructed based on these discrete sequences by statistically analyzing all unit sequences of length n in the entire corpus. Calculate the conditional probability distribution based on the number of occurrences of . Finally, by aggregating all n-gram probabilities, the probability distribution of a discrete unit sequence Y under the n-gram language model obtained using this discrete corpus can be obtained. This probability distribution can characterize the statistical properties and contextual dependencies of discrete speech units in the corpus, and can represent the characteristics of the corpus to a certain extent.

[0094] S9. Calculate the corpus difference degree corresponding to the element based on the first probability distribution and the second probability distribution.

[0095] In a preferred embodiment, the formula for calculating the corpus dissimilarity is:

[0096] ;

[0097] In the formula, Indicates the degree of corpus difference; Represents the first probability distribution; This represents the second probability distribution.

[0098] In this embodiment, the common discrete label sequence is calculated. The SCD can be obtained by finding the Kullback-Leibler divergence between the first probability distribution corresponding to the general corpus and the second probability distribution corresponding to the target corpus.

[0099] S10. Determine the target element with the smallest difference in the corpus, and use all the second speech data in the target element as the training data required for the target scene.

[0100] It should be noted that determining the target element with the minimum corpus difference is essentially: given a general corpus U and a target corpus Q, sample a subset S of size C from U such that SCD(Q, S) is minimized.

[0101] In this embodiment, the target corpus Q may be very small and cannot fully represent the target scenario. Therefore, directly calculating the SCD may lead to overfitting S to Q. To improve the generalization ability of the selected set S, an interpolation method is used on the probability distributions of U and Q to obtain a new probability distribution Q', with an interpolation factor of λ. Then, Q' is used to replace Q as the distribution of the target discrete corpus.

[0102] Example 2

[0103] Please refer to Figure 2 This is a schematic diagram of the structure of a speech recognition training data selection device according to an embodiment of the present invention, including: a data acquisition module, a discretization module, a difference calculation module and a training data selection module;

[0104] The data acquisition module intentionally acquires a general corpus and a target corpus corresponding to the target scene; wherein, the target corpus includes several first speech data; the general corpus includes several second speech data; determines all subsets of the general corpus with a preset size, and constructs a candidate corpus set using each subset as an element;

[0105] The discretization module is used to discretize each first speech data in the target corpus to generate several first discrete label sequences.

[0106] The difference calculation module is used to construct a first n-gram language model based on all the first discrete label sequences corresponding to the target corpus.

[0107] The discretization module is further configured to discretize each second speech data in each element of the candidate corpus set to generate several second discrete label sequences.

[0108] The difference calculation module is further configured to, for each element in the candidate corpus set, construct a second n-gram language model based on all second discrete label sequences corresponding to the element; compare all first discrete label sequences with all second discrete label sequences corresponding to the element, and randomly select one from all identical first and second discrete label sequences as a common discrete label sequence; calculate a first probability distribution of the common discrete label sequence under the first n-gram language model; calculate a second probability distribution of the common discrete label sequence under the second n-gram language model; and calculate the corpus difference degree corresponding to the element based on the first probability distribution and the second probability distribution.

[0109] The training data selection module is used to determine the target element with the smallest difference in the corpus, and to use all the second speech data in the target element as the training data required for the target scene.

[0110] Example 3

[0111] Accordingly, embodiments of the present invention provide a terminal device, the terminal device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the speech recognition training data selection method described in the above embodiments of the invention.

[0112] Example 4

[0113] Accordingly, embodiments of the present invention provide a storage medium, the storage medium including a stored computer program, wherein, when the computer program is running, it controls the device where the storage medium is located to execute the speech recognition training data selection method described in the above embodiments of the invention.

[0114] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0115] Those skilled in the art will clearly understand that, for convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0116] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0117] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the device, connecting various parts of the device via various interfaces and lines.

[0118] The memory can be used to store the computer program. The processor implements various functions of the device by running or executing the computer program stored in the memory and calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0119] The storage medium is a storage medium in which the computer program is stored. When executed by a processor, the computer program can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0120] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for selecting training data for speech recognition, characterized in that, include: Obtain a general corpus and a target corpus corresponding to the target scene; wherein, the target corpus includes several first speech data; the general corpus includes several second speech data; Identify all subsets of the general corpus that are of a preset size, and construct a candidate corpus set using each subset as an element; Each first speech data in the target corpus is discretized to generate several first discrete tag sequences; Construct a first n-gram language model based on all the first discrete label sequences corresponding to the target corpus; For each element in the candidate corpus set, each second speech data in the element is discretized to generate several second discrete label sequences. Construct a second n-gram language model based on all the second discrete label sequences corresponding to the elements; Compare all the first discrete label sequences and all the second discrete label sequences corresponding to the element, and randomly select one from all the same first discrete label sequences and second discrete label sequences as the common discrete label sequence; Calculate the first probability distribution of the shared discrete label sequence under the first n-gram language model; calculate the second probability distribution of the shared discrete label sequence under the second n-gram language model; Calculate the corpus difference degree corresponding to the element based on the first probability distribution and the second probability distribution; Identify the target element with the smallest difference in the corpus, and use all the second speech data in the target element as the training data required for the target scene.

2. The speech recognition training data selection method as described in claim 1, characterized in that, The discretization process for each first speech data in the target corpus to generate several first discrete label sequences includes: For each of the first speech data, features are extracted from the first speech data using a convolutional neural network to generate a first frame-level acoustic feature sequence; Cluster all first feature vectors in the first frame-level acoustic feature sequence; For each of the first feature vectors, a first discrete label of the first feature vector is calculated based on the first feature vector and the corresponding cluster center. Construct the first discrete label sequence based on the first discrete labels of all first feature vectors.

3. The speech recognition training data selection method as described in claim 2, characterized in that, The discretization process for each second speech data in the element to generate several second discrete label sequences includes: For each of the second speech data, features are extracted from the second speech data using a convolutional neural network to generate a second frame-level acoustic feature sequence; Cluster all second feature vectors in the second frame-level acoustic feature sequence; For each of the second feature vectors, a second discrete label is calculated based on the second feature vector and the corresponding cluster center. Based on the second feature vector and the second discrete label, construct the second discrete label sequence.

4. The speech recognition training data selection method as described in claim 1, characterized in that, The calculation of the first probability distribution of the shared discrete label sequence under the first n-gram language model includes: Count the number of occurrences of the first unit sequence of length n-1 in all first discrete label sequences; For each shared discrete acoustic unit in the shared discrete tag sequence, count the number of occurrences of the second unit sequence of length n in all first discrete tag sequences; wherein, the last discrete acoustic unit of the second unit sequence is the shared discrete acoustic unit; Based on the occurrence counts of the first unit sequence and the second unit sequence, the first n-gram conditional probability of the shared discrete acoustic unit is calculated using the n-gram conditional probability calculation formula. The first probability distribution is obtained by aggregating the first n-gram conditional probabilities of all shared discrete acoustic units.

5. The speech recognition training data selection method as described in claim 4, characterized in that, The calculation of the second probability distribution of the shared discrete label sequence under the second n-gram language model includes: Count the number of occurrences of the third unit sequence of length n-1 in all the second discrete label sequences of the element; For each shared discrete acoustic unit in the shared discrete tag sequence, count the number of occurrences of the fourth unit sequence of length n in all second discrete tag sequences of the element; wherein, the last discrete acoustic unit of the fourth unit sequence is the shared discrete acoustic unit. Based on the occurrence counts of the third unit sequence and the fourth unit sequence, the second n-gram conditional probability of the shared discrete acoustic unit is calculated using the n-gram conditional probability calculation formula. The second probability distribution is obtained by aggregating the second n-gram conditional probabilities of all shared discrete acoustic units.

6. The speech recognition training data selection method as described in claim 5, characterized in that, The formula for calculating the conditional probability of the n-gram is: ； In the formula, Represents the conditional probability of an n-gram; express of Number of occurrences; Indicated Number of occurrences; It consists of a shared discrete acoustic unit.

7. The speech recognition training data selection method as described in claim 1, characterized in that, The formula for calculating the corpus difference is: ； In the formula, Indicates the degree of corpus difference; Represents the first probability distribution; This represents the second probability distribution.

8. A speech recognition training data selection device, characterized in that, include: The module includes a data acquisition module, a discretization module, a difference calculation module, and a training data selection module. The data acquisition module intentionally acquires a general corpus and a target corpus corresponding to the target scene; wherein, the target corpus includes several first speech data; the general corpus includes several second speech data; determines all subsets of the general corpus with a preset size, and constructs a candidate corpus set using each subset as an element; The discretization module is used to discretize each first speech data in the target corpus to generate several first discrete label sequences. The difference calculation module is used to construct a first n-gram language model based on all the first discrete label sequences corresponding to the target corpus. The discretization module is further configured to discretize each second speech data in each element of the candidate corpus set to generate several second discrete label sequences. The difference calculation module is further configured to, for each element in the candidate corpus set, construct a second n-gram language model based on all second discrete label sequences corresponding to the element; compare all first discrete label sequences with all second discrete label sequences corresponding to the element, and randomly select one from all identical first and second discrete label sequences as a common discrete label sequence; calculate a first probability distribution of the common discrete label sequence under the first n-gram language model; calculate a second probability distribution of the common discrete label sequence under the second n-gram language model; and calculate the corpus difference degree corresponding to the element based on the first probability distribution and the second probability distribution. The training data selection module is used to determine the target element with the smallest difference in the corpus, and to use all the second speech data in the target element as the training data required for the target scene.

9. A terminal device, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the speech recognition training data selection method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the storage medium is located to perform the speech recognition training data selection method as described in any one of claims 1 to 7.

Images