Speech recognition model training method and device and computer equipment

Abstract

The invention relates to a speech recognition model training method and device and computer equipment. The method comprises the following steps: acquiring a plurality of first identification results of each to-be-labeled audio file in a to-be-labeled audio file set; the plurality of first recognition results are obtained by performing voice recognition on the to-be-labeled audio file through a plurality of voice recognition models; according to the plurality of first identification results of each to-be-labeled audio file, performing automatic labeling processing on the corresponding to-be-labeled audio file, and determining a labeling result and a result confidence coefficient of each to-be-labeled audio file; the automatic labeling processing comprises voting verification and fragment verification; integrating each to-be-labeled audio file and the corresponding labeling result and result confidence to generate a labeling data set; and training a preset speech recognition algorithm according to the annotation data set to obtain a target speech recognition model. By adopting the method, the training cost can be reduced and the model accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and particularly to a method, apparatus, and computer device for training a speech recognition model. Background Art

[0002] With the increasingly widespread application of speech recognition technology in the field of artificial intelligence, people have put forward higher and higher requirements for the accuracy of speech recognition technology. A speech recognition model is usually trained with training data, and the training data requires collecting a large amount of labeled audio data. To ensure accuracy, the labeling of training data is often obtained by manually labeling audio data. However, this method is time-consuming and laborious, and the labor cost is relatively high.

[0003] Traditional technologies also use artificial intelligence for automatic labeling. However, due to the unstable quality of labeled audio data and the existence of errors in the training data itself in the case of unsupervised learning, the accuracy of the speech recognition model is greatly reduced.

[0004] It can be seen that the existing speech recognition model training technologies still have problems of high training cost and low model accuracy. Summary of the Invention

[0005] Based on this, it is necessary to provide a method, apparatus, and computer device for training a speech recognition model that can reduce the training cost and improve the model accuracy for the above technical problems.

[0006] In a first aspect, this application provides a method for training a speech recognition model, and the method for training a speech recognition model includes:

[0007] Obtain multiple first recognition results of each to-be-labeled audio file in a to-be-labeled audio file set; the multiple first recognition results are obtained by performing speech recognition on the to-be-labeled audio file by multiple speech recognition models;

[0008] Perform automatic labeling processing on the corresponding to-be-labeled audio file through the multiple first recognition results of each to-be-labeled audio file, and determine the labeling result and result confidence of each to-be-labeled audio file; the automatic labeling processing includes vote verification and segment verification;

[0009] Integrate each to-be-labeled audio file and the corresponding labeling result and result confidence to generate a labeled data set;

[0010] Train a preset speech recognition algorithm according to the labeled data set to obtain a target speech recognition model.

[0011] In one embodiment, the automatic annotation processing of each to-be-annotated audio file by using the multiple first recognition results of each to-be-annotated audio file, and determining the annotation result and result confidence of each to-be-annotated audio file include:

[0012] For each to-be-annotated audio file, calculate the number of common characters between the target recognition result and other recognition results; the target recognition result is any one of the multiple first recognition results; the other recognition results are the first recognition results except the target recognition result;

[0013] Take the target recognition result with the largest number of common characters as the annotation result of the corresponding to-be-annotated audio file;

[0014] Determine the matching segments between the annotation result and other recognition results respectively, and determine the result confidence of the annotation result through the number of the matching segments.

[0015] In one embodiment, the determining the matching segments between the annotation result and other recognition results respectively, and determining the result confidence of the annotation result through the number of the matching segments include:

[0016] Segment the annotation result by using a preset text processing algorithm to obtain multiple first segments;

[0017] Segment the other recognition results by using a preset text processing algorithm to obtain multiple second segments;

[0018] Match each first segment with each second segment respectively to determine the matching segments;

[0019] Take the ratio of the number of the matching segments to the total number of the first segments as the result confidence of the annotation result.

[0020] In one embodiment, the integrating each to-be-annotated audio file and the corresponding annotation result and result confidence to generate an annotation data set includes:

[0021] Take the to-be-annotated audio file and the annotation result whose result confidence meets the preset confidence as annotation data;

[0022] Obtain multiple preset texts and a speech generation model;

[0023] Input the preset text into the speech generation model to obtain multiple generated audio files;

[0024] Take the generated audio files and the corresponding preset texts as generated data;

[0025] Integrate the generated data and the labeled data to generate a labeled data set.

[0026] In one embodiment, after using the generated audio file and the corresponding preset text as the generated data, the method further includes:

[0027] Obtain multiple second recognition results for each of the generated audio files; the multiple second recognition results are obtained by multiple speech recognition models performing speech recognition on the generated audio files;

[0028] Perform automatic annotation processing on the corresponding generated audio file through the multiple second recognition results of each generated audio file to determine the annotation result and result confidence of each generated audio file;

[0029] Filter the multiple generated data through the annotation result and result confidence of each generated audio file to obtain the filtered generated data.

[0030] In one embodiment, multiple generated data threads are used to generate the labeled data set, and the method of training a preset speech recognition algorithm according to the labeled data set to obtain a target speech recognition model includes:

[0031] Write the labeled data set into a data queue through multiple generated data threads;

[0032] Read the labeled data set in the data queue item by item through multiple training threads and input it into the preset speech recognition algorithm for training to obtain a target speech recognition model.

[0033] In one embodiment, the method further includes:

[0034] Obtain the number of data in the data queue;

[0035] Adjust the number of generated data threads and / or adjust the number of training threads according to the number of data.

[0036] In a second aspect, the present application provides a speech recognition model training device, and the device includes:

[0037] An acquisition module, configured to obtain multiple first recognition results for each to-be-labeled audio file in a to-be-labeled audio file set; the multiple first recognition results are obtained by multiple speech recognition models performing speech recognition on the to-be-labeled audio files;

[0038] An automatic annotation module, configured to perform automatic annotation processing on the corresponding to-be-labeled audio file through the multiple first recognition results of each to-be-labeled audio file to determine the annotation result and result confidence of each to-be-labeled audio file; the automatic annotation processing includes vote verification and segment verification;

[0039] A dataset generation module, configured to integrate each audio file to be annotated, as well as the corresponding annotation result and result confidence, and generate an annotated dataset.

[0040] A training module, configured to train a preset speech recognition algorithm according to the annotated dataset to obtain a target speech recognition model.

[0041] In a third aspect, the present application provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the method described above is implemented.

[0042] In a fourth aspect, the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method described above is implemented.

[0043] The above speech recognition model training method, device and computer device obtain multiple first recognition results of each audio file to be annotated in the audio file set to be annotated; the multiple first recognition results are obtained by multiple speech recognition models performing speech recognition on the audio file to be annotated; through the multiple first recognition results of each audio file to be annotated, automatic annotation processing is performed on the corresponding audio file to be annotated to determine the annotation result and result confidence of each audio file to be annotated; the automatic annotation processing includes voting verification and segment verification; integrating each audio file to be annotated, as well as the corresponding annotation result and result confidence, to generate an annotated dataset; training a preset speech recognition algorithm according to the annotated dataset to obtain a target speech recognition model. Using multiple speech recognition models to perform speech recognition respectively to obtain first recognition results can reduce the deviation of a single model, and perform automatic annotation including voting verification and segment verification according to multiple first recognition results to obtain the annotation result and result confidence, which can reduce mislabeling and improve the consistency and reliability of the annotation result. Further obtaining an annotated dataset according to the annotation result and result confidence can obtain a more accurate annotated dataset for model training, realizing joint recognition of multiple models and automatic verification, thereby reducing the dependence on manual annotation, and at the same time improving the model performance through a high-quality annotated dataset, so as to achieve the effects of reducing training costs and improving the model accuracy. Description of the Drawings

[0044] Figure 1 It is an application environment diagram of the speech recognition model training method in an embodiment;

[0045] Figure 2 It is a flowchart of the speech recognition model training method in an embodiment;

[0046] Figure 3 It is a schematic flowchart of a method for training a speech recognition model in another embodiment;

[0047] Figure 4 It is a structural block diagram of a speech recognition model training device in an embodiment;

[0048] Figure 5 It is an internal structure diagram of a computer device in an embodiment. Detailed implementation manners

[0049] In order to make the objectives, technical solutions and advantages of the present application clearer and more understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0050] The speech recognition model training method provided by the embodiments of the present application can be applied to an application environment as Figure 1 shown. Among them, the terminal 102 communicates with the server 104 through a network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or placed in the cloud or other network servers. The terminal 102 communicates with the server 104 to obtain multiple first recognition results of each to-be-annotated audio file in the to-be-annotated audio file set; the multiple first recognition results are obtained by multiple speech recognition models performing speech recognition on the to-be-annotated audio file; through the multiple first recognition results of each to-be-annotated audio file, automatic annotation processing is performed on the corresponding to-be-annotated audio file to determine the annotation result and result confidence of each to-be-annotated audio file; the automatic annotation processing includes vote verification and segment verification; integrate each to-be-annotated audio file and the corresponding annotation result and result confidence to generate an annotation data set; according to the annotation data set, train a preset speech recognition algorithm to obtain a target speech recognition model. Among them, the terminal 102 can be but is not limited to various personal computers, laptop computers, smart phones, tablet computers, and portable wearable devices. The portable wearable device can be a smart watch, a smart bracelet, a head-mounted device, etc. Exemplarily, the head-mounted device can also include an AR glasses, etc. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers.

[0051] In one embodiment, as Figure 2 shown, a speech recognition model training method is provided. Taking the method applied to the Figure 1 terminal 102 as an example, the method includes the following steps:

[0052] Step S100, obtain multiple first recognition results of each to-be-annotated audio file in the to-be-annotated audio file set.

[0053] Among them, the set of audio files to be labeled can be a set of audio files that have not been labeled yet. The audio data can be audio data for training a speech recognition model, which can be obtained by collecting from actual application scenarios or from public datasets. Exemplarily, the set of audio files to be labeled can include, but is not limited to, telephone recordings, meeting records, voice commands, etc.

[0054] The multiple first recognition results are obtained by multiple speech recognition models performing speech recognition on the audio files to be labeled. Exemplarily, each audio file to be labeled can be input into each selected speech recognition model, and the first recognition results output by each speech recognition model are collected, so as to obtain multiple first recognition results of a single audio file to be labeled. It can be understood that by combining the recognition results of multiple speech recognition models for labeling, the bias that may exist in a single model can be reduced, and the accuracy of the labeling results can be improved.

[0055] Step S200, perform automatic labeling processing on each of the audio files to be labeled through the multiple first recognition results of each of the audio files to be labeled, and determine the labeling result and result confidence of each of the audio files to be labeled.

[0056] Among them, the labeling result can be the text content of the finally determined audio file, and the result confidence can represent the credibility of the labeling result. Exemplarily, the result confidence can include, but is not limited to, voting verification scores, segment verification scores, etc.

[0057] The automatic labeling processing includes voting verification and segment verification. Voting verification can be a verification rule based on the majority principle, and segment verification can be a verification rule based on text segmentation. In this embodiment, voting verification can be used only to determine the labeling result or the result confidence, or can be used simultaneously to determine the labeling result and the result confidence; segment verification can also be used only to determine the labeling result or the result confidence, or can be used simultaneously to determine the labeling result and the result confidence.

[0058] In an exemplary embodiment, voting verification can be used for the labeling result, and segment verification can be used for calculating the result confidence. Alternatively, segment verification can be used for the labeling result, and voting verification can be used for calculating the result confidence.

[0059] In another exemplary embodiment, voting verification can be used to separately determine the first annotation result and the first result confidence level, and segment verification can be used to determine the second annotation result and the second result confidence level. If the first annotation result is the same as the second annotation result, the annotation result is determined by merging, and the result confidence level is comprehensively determined based on the first result confidence level and the second result confidence level. If the first annotation result is different from the second annotation result, the first annotation result or the second annotation result can be determined as the annotation result according to the magnitude relationship between the first result confidence level and the second result confidence level. Correspondingly, the determination of the result confidence level can be the result confidence level corresponding to the selected first annotation result or second annotation result. Further, when the first annotation result and the second annotation result are the same, the value of the result confidence level can also be increased according to a preset ratio. When the first annotation result and the second annotation result are different, the value of the result confidence level can be decreased according to a preset ratio.

[0060] When using voting verification to determine the annotation result, the character overlap degree between each first recognition result and other first recognition results can be compared to determine the text occurrence frequency of each first recognition result, and the first recognition result with the highest text occurrence frequency is used as the annotation result. In another exemplary embodiment, for the same audio file to be annotated, the characters whose total occurrence times in multiple first recognition results exceed a preset threshold can also be counted. The preset threshold can be determined through prior knowledge, and then the first recognition result containing the largest number of the above characters in the recognition result is used as the annotation result. It can also be other ways to determine the first recognition result with the highest accuracy according to the majority principle, which is not limited in this embodiment.

[0061] When using voting verification to determine the result confidence level, the character overlap degree between each first recognition result and other first recognition results can be compared to comprehensively determine the result confidence level of each first recognition result. In another exemplary embodiment, for the same audio file to be annotated, the characters whose total occurrence times in multiple first recognition results exceed a preset threshold can also be counted. The preset threshold can be determined through prior knowledge, and then the result confidence level is determined according to the proportion of the number of the above characters in each first recognition result in the total number of characters. It can also be other ways to determine the result confidence level according to the majority principle, which is not limited in this embodiment.

[0062] When using segment verification to determine the annotation result and / or the result confidence level, an audio segmentation method based on sound wave amplitude or deep learning can be used to obtain the audio segments of each first recognition result, and then the text segments corresponding to each audio segment are determined in combination with the first recognition result. Or a text segmentation method based on a sliding window or semantics can be used to directly obtain multiple text segments of each first recognition result.

[0063] When using segment verification to determine the annotation result, according to the obtained text segments, the text segment overlap degree between each first recognition result and other first recognition results can be compared to determine the text segment occurrence frequency of each first recognition result, and the first recognition result with the highest text segment occurrence frequency is used as the annotation result; in another exemplary embodiment, for the same audio file to be annotated, text segments whose total occurrence times in multiple first recognition results exceed a preset threshold can also be counted. The preset threshold can be determined according to prior knowledge, and then the first recognition result containing the largest number of the above text segments in the recognition results is used as the annotation result; it can also be other ways to determine the first recognition result with the highest accuracy according to text segmentation. This embodiment does not limit this here.

[0064] When using segment verification to determine the result confidence, the text segment overlap degree between each first recognition result and other first recognition results can be compared to comprehensively determine the result confidence of each first recognition result; in another exemplary embodiment, for the same audio file to be annotated, text segments whose total occurrence times in multiple first recognition results exceed a preset threshold can also be counted. The preset threshold can be determined according to prior knowledge, and then the result confidence is determined according to the proportion of the number of the above text segments in each first recognition result in the total number of segments; it can also be other ways to determine the result confidence according to text segmentation principles. This embodiment does not limit this here.

[0065] Through the automatic annotation process including voting verification and segment verification, this embodiment obtains the annotation result and the result confidence, which can achieve the effects of reducing incorrect annotations and improving the consistency and reliability of the annotation result.

[0066] Step S300, integrate each audio file to be annotated and the corresponding annotation result and result confidence to generate an annotation data set.

[0067] Among them, the annotation data set can be a data set containing audio files and their corresponding annotation results and confidences, and is used for subsequent model training. Exemplarily, the annotation data set can include, but is not limited to, fields such as the audio file path, the content of the annotated text, and the numerical value of the result confidence.

[0068] The result confidence can be used as an indicator for screening high-quality annotation data, and the data with a higher result confidence is used to generate the annotation data set. Exemplarily, audio files with a result confidence higher than a preset confidence threshold and the corresponding annotation results can be used to form the annotation data set.

[0069] Step S400, train a preset speech recognition algorithm according to the annotation data set to obtain a target speech recognition model.

[0070] Among them, the preset speech recognition algorithm may refer to the initial speech recognition model algorithm for training. Exemplarily, it may be an algorithm based on deep learning, including but not limited to the CTC model based on RNN, the Attention model based on Transformer, etc.

[0071] Training the preset speech recognition algorithm may involve dividing the labeled data set into a training set and a validation set, training the preset speech recognition algorithm with the training set, adjusting the model parameters to minimize the loss function, evaluating the model performance with the validation set, adjusting the hyperparameters (such as the learning rate, batch size, etc.) to optimize the model, and saving the trained model as the target speech recognition model.

[0072] A speech recognition model training method provided in this embodiment includes obtaining multiple first recognition results of each to-be-labeled audio file in the to-be-labeled audio file set; the multiple first recognition results are obtained by performing speech recognition on the to-be-labeled audio file by multiple speech recognition models; through the multiple first recognition results of each to-be-labeled audio file, performing automatic annotation processing on the corresponding to-be-labeled audio file to determine the annotation result and result confidence of each to-be-labeled audio file; the automatic annotation processing includes vote verification and segment verification; integrating each to-be-labeled audio file and the corresponding annotation result and result confidence to generate a labeled data set; training a preset speech recognition algorithm according to the labeled data set to obtain a target speech recognition model. Using multiple speech recognition models to perform speech recognition respectively to obtain the first recognition results can reduce the deviation of a single model, and performing automatic annotation including vote verification and segment verification based on the multiple first recognition results to obtain the annotation result and result confidence can reduce mislabeling and improve the consistency and reliability of the annotation result. Further obtaining a labeled data set based on the annotation result and result confidence can obtain a more accurate labeled data set for model training, realizing joint recognition of multiple models and automatic verification, thereby reducing the dependence on manual annotation, and at the same time improving the model performance through a high-quality labeled data set, so as to achieve the effects of reducing the training cost and improving the model accuracy.

[0073] In one embodiment, vote verification may be achieved by calculating the number of common characters between each first recognition result and other first recognition results, and segment verification may be verified by the segment overlap situation between the annotation result and other first recognition results.

[0074] The step of performing automatic annotation processing on the corresponding to-be-labeled audio file through the multiple first recognition results of each to-be-labeled audio file to determine the annotation result and result confidence of each to-be-labeled audio file includes:

[0075] For each of the to-be-annotated audio files, calculate the number of common characters between the target recognition result and other recognition results;

[0076] Take the target recognition result with the largest number of common characters as the annotation result for the corresponding to-be-annotated audio file;

[0077] Determine the matching segments between the annotation result and other recognition results respectively, and determine the result confidence of the annotation result based on the number of the matching segments.

[0078] Among them, the number of common characters can be the number of identical characters in two first recognition results. The target recognition result is any one of the first recognition results; the other recognition results are the first recognition results other than the target recognition result. Exemplarily, each first recognition result can be taken as the target recognition result in turn, and the first recognition results other than the target recognition result are taken as other recognition results, and the number of common characters between the target recognition result and all other recognition results is calculated; the number of common characters of each target recognition result is counted; the first recognition result with the largest number of common characters is selected as the annotation result for the corresponding to-be-annotated audio file, so as to effectively quantify the consistency degree between the output results of different models and reduce the possible deviation of a single model.

[0079] Determining the matching segments between the annotation result and other recognition results respectively can be to divide the text into multiple segments and then compare the repeated occurrences of these segments in other recognition results. Exemplarily, after obtaining multiple text segments through an audio segmentation method based on acoustic wave amplitude or deep learning, or a text segmentation method based on semantics, calculate the ratio of the number of segments in the annotation result that appear in other recognition results to the total number of segments, so as to comprehensively calculate the result confidence. Further, through a weighted average method, segments with a higher segment overlap degree can be given a higher weight, so as to more finely evaluate the reliability of the annotation result and further improve the result reliability in long texts or complex scenarios.

[0080] A speech recognition model training method provided in this embodiment can effectively capture the consistency between the output results of different models and reduce the possibility of incorrect annotation by counting the number of common characters and taking the target recognition result with the largest number of common characters as the annotation result; by calculating the result confidence based on the matching results of segments, the verification ability for local details can be further improved, ensuring that the annotation result has high consistency at both the global and local levels, so as to achieve the effect of improving the accuracy of the annotation result and the reliability of the result confidence.

[0081] In one embodiment, determining the matching segments between the annotation result and other recognition results respectively, and determining the result confidence of the annotation result based on the number of the matching segments includes:

[0082] Segment the annotation result with a preset text processing algorithm to obtain a plurality of first segments;

[0083] Segment the other recognition results with the same preset text processing algorithm to obtain a plurality of second segments;

[0084] Match each of the first segments with each of the second segments respectively to determine the matching segments;

[0085] Take the ratio of the number of the matching segments to the total number of the first segments as the result confidence of the annotation result.

[0086] Among them, the preset text processing algorithm can be an algorithm for dividing text into multiple segments, which can be obtained through rule definition or deep learning model training. Exemplarily, the preset text processing algorithm includes but is not limited to rule-based segmentation (such as dividing by punctuation marks, spaces, etc.) and deep learning-based segmentation (such as using a Transformer model to predict text boundaries). In a specific embodiment, the segmentation granularity can be determined according to the structural characteristics of the input text, and based on this segmentation granularity, the preset text processing algorithm is used for segmentation.

[0087] The first segment can be multiple small segments of text extracted from the annotation result through a preset text processing algorithm. In a specific embodiment, for the annotation result "Open chat", it can be segmented into two first segments: "Open" and "chat".

[0088] The second segment can be multiple small segments of text extracted from other recognition results through the same preset text processing algorithm. Exemplarily, the other recognition results can be results such as "Open also requires ingredients", "Open friend", "OK also friend", etc. Correspondingly, "Open also requires ingredients" can be segmented into three second segments: "Open", "also", and "ingredients".

[0089] To improve the comparability between the first segment and the second segment, the preset text processing algorithm can use the same granularity when segmenting the annotation result and other first speech recognition results.

[0090] Matching the first segment and the second segment can be calculating the similarity or consistency between the first segment and the second segment. In a specific embodiment, in the case where the matching requirements are relatively strict, the matching situation can be determining whether the first segment and the second segment completely coincide.

[0091] The matching segments can be segments with relatively high consistency among multiple first recognition results. Exemplarily, if a first segment appears in a preset number of second segments, then this segment can be regarded as a matching segment.

[0092] Exemplarily, to determine the matching segments, it can be to traverse each first segment, compare it with all the second segments one by one, record the number of successful matches of each first segment in the set of second segments, and determine the first segments with the number of matches exceeding the preset threshold as the matching segments, so that the matching segments with relatively high consistency can be screened out, reflecting the consistency degree of the annotation results at the segment level.

[0093] Take the ratio of the number of the matching segments to the total number of the first segments as the result confidence of the annotation result, so that the overall reliability of the annotation result can be more accurately reflected. It can be understood that the higher the proportion, the higher the reliability of the annotation result at the segment level.

[0094] A method for training a speech recognition model provided in this embodiment, by segmenting the annotation result with a preset text processing algorithm to obtain multiple first segments; segmenting the other recognition results with a preset text processing algorithm to obtain multiple second segments; matching each of the first segments with each of the second segments respectively to determine the matching segments; taking the ratio of the number of the matching segments to the total number of the first segments as the result confidence of the annotation result, can achieve the technical effects of improving the accuracy of the result confidence, enhancing the reliability of the annotation result, and optimizing the quality of the training data.

[0095] In one of the embodiments, the integrating each audio file to be annotated and the corresponding annotation result and result confidence to generate an annotation dataset includes:

[0096] Taking the audio file to be annotated and the annotation result whose result confidence meets the preset confidence as the annotation data;

[0097] Obtaining multiple preset texts and a speech generation model;

[0098] Inputting the preset text into the speech generation model to obtain multiple generated audio files;

[0099] Taking the generated audio files and the corresponding preset texts as the generated data;

[0100] Integrating the generated data and the annotation data to generate an annotation dataset.

[0101] Among them, the labeled data can be the audio files retained after screening by the result confidence, and their corresponding labeling results. The preset confidence can be a preset confidence threshold for screening high-quality labeled data, which can be obtained through prior knowledge. Exemplarily, the preset confidence can include but is not limited to specific values such as 0.8, 0.9, etc.

[0102] In a specific embodiment, it can be to compare the result confidence of each audio file one by one, and save the audio files and their labeling results with the result confidence greater than or equal to the preset confidence as labeled data, so as to screen out high-quality labeled data to ensure the accuracy of the training data.

[0103] The preset text can be a set of text data prepared in advance for generating speech data, which can be obtained by manual writing or from public datasets. Exemplarily, the preset text can include but is not limited to news reports, excerpts of literary works, daily conversations, etc.

[0104] The text-to-speech (TTS) model can be a model that can convert text into speech signals, and can include but is not limited to deep learning-based Tacotron models, WaveNet models, etc. Exemplarily, the preset text-to-speech model can adopt a multi-language and multi-lingual speech synthesis model.

[0105] In this embodiment, by collecting the preset text and using the text-to-speech model to convert each piece of preset text into a corresponding speech segment, the generated audio files can be obtained, so as to realize generating additional training data through the text-to-speech model and increase the diversity of the training data.

[0106] In this embodiment, the labeled data set can be the integrated result of the labeled data and the generated audio files, and is used for the model training of the preset speech recognition algorithm.

[0107] A method for training a speech recognition model provided in this embodiment screens out high-quality labeled data by using the to-be-labeled audio files whose result confidence meets the preset confidence and their labeling results to ensure the accuracy of the training data; at the same time, multiple generated audio files are generated through the text-to-speech model and the preset text, which can make up for the lack of diversity in the training data. Finally, the labeled data and the generated audio files are integrated to generate a labeled data set, and a more comprehensive training data set can be constructed, so as to improve the quality and diversity of the labeled data, optimize the model training effect, and achieve the effect of improving the accuracy and generalization ability of the target speech recognition model.

[0108] In one of the embodiments, after using the generated audio files and the corresponding preset text as generated data, it further includes:

[0109] Obtain multiple second recognition results for each of the generated audio files;

[0110] Based on the multiple second recognition results of each generated audio file, perform automatic annotation processing on the corresponding generated audio file to determine the annotation result and result confidence of each generated audio file;

[0111] Based on the annotation result and result confidence of each generated audio file, filter the multiple generated data to obtain the filtered generated data.

[0112] Among them, the multiple second recognition results are obtained by performing speech recognition on the generated audio file by multiple speech recognition models.

[0113] In a specific embodiment, it is possible to traverse all the generated audio files, input each generated audio file one by one into the selected multiple speech recognition models, and collect the second recognition results output by each speech recognition model.

[0114] Based on the multiple second recognition results of each generated audio file, performing automatic annotation processing on the corresponding generated audio file can be, for example, to verify using methods for multiple speech recognition results, which will not be elaborated in this embodiment.

[0115] Furthermore, since the correct text corresponding to the generated audio file is known, therefore, the generated audio file itself can also be corrected through this correct text, for example, re - input into a preset speech generation model to obtain a more accurate generated audio file.

[0116] A speech recognition model training method provided in this embodiment, by obtaining multiple second recognition results of each generated audio file and using multiple speech recognition models to perform speech recognition on each generated audio file respectively, can reduce the possible bias of a single model; based on the multiple second recognition results of each generated audio file, perform automatic annotation processing on the corresponding generated audio file to determine the annotation result and result confidence of each generated audio file, thereby improving the accuracy and consistency of the annotation results; filter multiple generated audio files through the annotation results and result confidence to obtain the filtered multiple generated audio files, thereby significantly improving the quality and reliability of the generated audio files, optimizing the quality of the training data, and achieving the effect of improving the accuracy and generalization ability of the target speech recognition model.

[0117] In one of the embodiments, multiple generated data threads are used to generate the annotation dataset, and training a preset speech recognition algorithm according to the annotation dataset to obtain a target speech recognition model includes:

[0118] Write the labeled data set into a data queue through multiple data generation threads;

[0119] Read the labeled data set in the data queue one by one through multiple training threads, and input it into the preset speech recognition algorithm for training to obtain a target speech recognition model.

[0120] Among them, the data queue can be a queue for storing data waiting to be used for training. Exemplarily, the data queue can include audio files and their corresponding annotation results, etc.

[0121] In some examples, the data queue can be, but is not limited to, a data buffer in memory, a message queue in a distributed system, etc.

[0122] The data generation thread can be a thread for generating training data. Exemplarily, the data generation thread can include a thread for automatically annotating an audio file to be annotated, or can also include a thread for obtaining a generated audio file through a preset text.

[0123] The training thread can be a thread for training a preset speech recognition algorithm. The number of training threads can be determined according to prior knowledge or can also be adjusted according to the real-time situation of the data queue.

[0124] A speech recognition model training method provided in this embodiment writes the labeled data set into a data queue through multiple data generation threads; reads the labeled data set in the data queue one by one through multiple training threads, and inputs it into the preset speech recognition algorithm for training to obtain a target speech recognition model. It can generate data synchronously during the process of training the model, thereby avoiding the waste of time caused by step-by-step processing after all training data needs to be generated, and can achieve the effect of improving the model training efficiency.

[0125] In one of the embodiments, the method further includes:

[0126] Obtain the number of data in the data queue;

[0127] Adjust the number of data generation threads and / or adjust the number of training threads according to the number of data.

[0128] Among them, the number of data can be the real-time data quantity or size stored in the current data queue, and can be obtained by reading the status information of the queue and calculating its total number of bytes or entries.

[0129] According to the quantity of the data, adjusting the quantity of the data generation threads and / or adjusting the quantity of the training threads may be to determine whether the quantities of the current data generation threads and training threads need to be adjusted according to the quantity of the data. Exemplarily, when the quantity of the data in the data queue exceeds the preset data quantity condition for a long time, it can be considered that the current data generation threads have a strong ability to generate labeled data or generate audio files, and the processing ability of the training threads is relatively weak. Then, by increasing the quantity of the training threads or decreasing the quantity of the data generation threads, the speed matching of data generation and data training can be achieved; conversely, when the quantity of the data in the data queue is lower than the preset data quantity condition for a long time, it can be considered that the current data generation threads have a weak ability to generate labeled data or generate audio files, and the processing ability of the training threads is relatively strong. Then, by decreasing the quantity of the training threads or increasing the quantity of the data generation threads, the speed matching of data generation and data training can be achieved.

[0130] In this embodiment, by adjusting the quantity of the data generation threads and / or adjusting the quantity of the training threads according to the quantity of the data, the voice generation speed can be dynamically adjusted, the problem that the training is interrupted or the data is insufficient due to too low a data quantity in the data queue can be avoided, and at the same time, it is ensured that the data queue always maintains within a reasonable data quantity range, thereby improving the data utilization efficiency and the continuity of model training.

[0131] A voice recognition model training method provided in this embodiment monitors the quantity of the data in the data queue, and adjusts the quantity of the data generation threads and / or adjusts the quantity of the training threads according to the quantity of the data, so as to dynamically adjust the timing of generating audio files and adding labeled data, avoid unnecessary resource waste, and further ensure the stability and continuity of the model training process, thereby achieving the effect of improving the resource utilization rate and the model training efficiency.

[0132] To more clearly elaborate the technical solution of the present application, the present application also provides a detailed embodiment.

[0133] In one embodiment, a training method of a voice recognition model is provided, which is applied to the voice interaction scenario of smart glasses for enhancing the whisper base model in a specific field, such as Figure 3 As shown, the training method includes:

[0134] Step S1, data preparation.

[0135] Step S11: Obtain the files and tools for generating data. The files may include multiple voice files, and the tools may include a TTS model for text-to-speech (such as the ChatTTS model), multiple automatic speech recognition (ASR) models. In this embodiment, four ASR models are taken as an example (such as the whisper-large-v3 model, paraformer model, paddle-speech model, and qwen-audio model); generate an audio file through the TTS model, using a random voice or a specified voice, convert the domain text into speech, and convert the sampling rate to the sampling rate set in the target speech recognition model. In a specific embodiment, the sampling rate is 16k.

[0136] Step S12: Annotate the data. By collecting ASR corpora (such as obtaining voice data from the CommonVoice platform) in the voice interaction scenario of the target language (such as Chinese, English), including audio files and corresponding texts.

[0137] Step S13: Obtain unannotated data. In this embodiment, the audio files collected on the MR glasses can be obtained.

[0138] Step S2: Data annotation.

[0139] For the unannotated audio data, an automatic annotation method using multi-model voting and information fragment compensation and correction is adopted, where:

[0140] An audio file is recognized by four ASR models to obtain four speech recognition results, and these four speech recognition results are fused to obtain the most reliable result. For example, the four results include "I turn off the 3D island", "Turn off the 3D navigation", "Three close three low down", and "Turn off the 3D navigation".

[0141] Find the result that has the most common characters with the other 3 results. This result is "Turn off the 3D navigation". Then, use the other speech recognition results to vote and verify the fragments in this speech recognition result. For each fragment with a character length of not less than 2, as long as the same fragment also exists in other results, it can be considered that this fragment is correct, and the proportion of the unconfirmed fragments and the correct fragments is returned.

[0142] Highlight the unconfirmed fragments in the most reliable result in the table document for easy viewing. If for the three fragments "Turn off", "3D", and "navigation", the same fragments can be found in other results, it can be considered that this result is correct.

[0143] In another specific embodiment, the four speech recognition results include "open also want", "open chat", "ok also friend day", and "open ingredients". The most reliable result is "open chat", and the return information is [{"diff_span": [2, 4], "gt": "chat"}]. Here, diff_span is the ratio of the unconfirmed segment to the total segment, and gt is the content of the unconfirmed segment. At this time, the proportion of the correct segment, that is, the result confidence level, is 0.5.

[0144] After processing a batch of audio files as described above, save the results to a tabular document (such as an excel file), and filter out samples with an empty unconfirmed segment and samples with a proportion of correct segments greater than a threshold, such as samples with a proportion exceeding the threshold of 0.9, as training samples. Thus, in the case of a lack of annotation resources, a large amount of high-quality annotation data can be quickly obtained.

[0145] Step S3, model training.

[0146] The process of model training uses a data generator and a data loader, where:

[0147] The data generator is used to prepare a data queue, and start multiple background threads to generate audio samples and write them into the data queue. The data in the data queue comes from the generated audio files and annotation data. Among them, the ratio of the generated audio files to the annotation data can be determined according to prior knowledge. Exemplarily, the proportion of the generated audio files can be 10%, and the proportion of the annotation data can be 90%.

[0148] On each background thread, run files and tools for generating audio files (multiple tone files, TTS models, ASR models) to generate audio samples.

[0149] To ensure the quality of the generated audio, use multiple speech recognition models to recognize the generated data, and perform voting and information segment compensation and correction based on the obtained recognition results. Among them, the quality of the generated audio file can be judged by checking the recognition result of the generated audio. The generated audio file that can be corrected by the voting method to obtain the original text result can be considered as high-quality audio, and the audio data that cannot be corrected is discarded; thus, samples with an empty unconfirmed segment in the voting result or samples with a proportion of correct segments greater than a threshold (such as 0.9) can be used as high-quality audio.

[0150] Perform multi-threaded access protection on the data queue. When the amount of data in the data queue exceeds the threshold of 300, pause writing to the queue.

[0151] The data loader is used to load the dynamically generated data in the data queue one by one instead of loading it all at once, enabling the training data to be generated online. Moreover, it can use data composed of richer tones and texts to train the model, while reducing the data generation time and avoiding the waiting time caused by the slow generation of high-quality audio data by the TTS model.

[0152] The data loader reads data from the data queue and inputs it into model training.

[0153] Furthermore, after a data flows out of the data queue, the data flowing into the data queue can be determined according to a preset ratio of the generated audio files and the annotated data.

[0154] When waiting when there is no data in the queue, track the quantity status of the queue in the reading and generating threads to adjust the number of generating threads. For example, if the quantity is insufficient and the waiting time is relatively long, add more threads to meet the need for loading training data and reduce the waiting time for loading data.

[0155] Furthermore, when there are multiple graphics cards or multiple computing resources, a distributed computing framework such as ray can be used for process management;

[0156] Furthermore, the training main thread and the thread for generating data can use different graphics cards. Each thread generates samples in batches, and the configuration that can generate the largest number of samples per second is tested. For example, 16 samples are generated in a single batch.

[0157] A method for training a speech recognition model provided in this embodiment can use data composed of richer tones and texts to train the model by integrating dynamic data generation into the training process, while reducing the data generation time. In the data annotation process, a large amount of high-quality annotated data can be quickly obtained in the case of a lack of annotation resources, thereby achieving the effects of reducing the training cost and improving the model accuracy.

[0158] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are sequentially shown according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same moment, but can be executed at different moments. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0159] Based on the same inventive concept, an embodiment of this application further provides a speech recognition model training device for implementing the speech recognition model training method involved above. The solution provided by this device for solving problems is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the speech recognition model training device provided below can refer to the limitations on the speech recognition model training method in the above text and will not be elaborated here.

[0160] In one embodiment, as Figure 4 shown, a speech recognition model training device is provided, including an acquisition module 100, an automatic annotation module 200, a dataset generation module 300, and a training module 400, where:

[0161] The acquisition module 100 is configured to acquire multiple first recognition results of each to-be-annotated audio file in the to-be-annotated audio file set. The multiple first recognition results are obtained by multiple speech recognition models performing speech recognition on the to-be-annotated audio file.

[0162] The automatic annotation module 200 is configured to perform automatic annotation processing on the corresponding to-be-annotated audio file through the multiple first recognition results of each to-be-annotated audio file, and determine the annotation result and result confidence of each to-be-annotated audio file. The automatic annotation processing includes voting verification and segment verification.

[0163] The dataset generation module 300 is configured to integrate each to-be-annotated audio file and the corresponding annotation result and result confidence to generate an annotated dataset.

[0164] The training module 400 is configured to train a preset speech recognition algorithm according to the annotated dataset to obtain a target speech recognition model.

[0165] In one of the embodiments, the automatic annotation module 200 is further configured to:

[0166] For each to-be-annotated audio file, calculate the number of common characters between the target recognition result and other recognition results; the target recognition result is any one of the multiple first recognition results; the other recognition results are the first recognition results other than the target recognition result;

[0167] Use the target recognition result with the largest number of common characters as the annotation result of the corresponding to-be-annotated audio file;

[0168] Determine the matching segments between the annotation result and other recognition results respectively, and determine the result confidence of the annotation result through the number of the matching segments.

[0169] In one of the embodiments, the automatic annotation module 200 is further configured to:

[0170] Segment the annotation result with a preset text processing algorithm to obtain a plurality of first segments;

[0171] Segment the other recognition results with a preset text processing algorithm to obtain a plurality of second segments;

[0172] Match each of the first segments with each of the second segments to determine the matching segments;

[0173] Take the ratio of the number of the matching segments to the total number of the first segments as the result confidence of the annotation result.

[0174] In one embodiment, the dataset generation module 300 is further configured to:

[0175] Use the audio file to be annotated and the annotation result whose result confidence meets the preset confidence as annotation data;

[0176] Obtain a plurality of preset texts and speech generation models;

[0177] Input the preset text into the speech generation model to obtain a plurality of generated audio files;

[0178] Use the generated audio files and the corresponding preset texts as generated data;

[0179] Integrate the generated data and the annotation data to generate an annotation dataset.

[0180] In one embodiment, the dataset generation module 300 is further configured to:

[0181] Obtain multiple second recognition results of each of the generated audio files; the multiple second recognition results are obtained by performing speech recognition on the generated audio files by multiple speech recognition models;

[0182] Perform automatic annotation processing on the corresponding generated audio files through the multiple second recognition results of each generated audio file to determine the annotation result and the result confidence of each generated audio file;

[0183] Filter the multiple generated data through the annotation result and the result confidence of each generated audio file to obtain the filtered generated data.

[0184] In one embodiment, multiple generated data threads are used to generate the annotation dataset, and the training module 400 is further configured to:

[0185] Write the annotation dataset into a data queue through multiple generated data threads;

[0186] Read the labeled data set in the data queue item by item through multiple training threads, and input it into the preset speech recognition algorithm for training to obtain a target speech recognition model.

[0187] In one embodiment, the training module 400 is further configured to:

[0188] Obtain the number of data in the data queue;

[0189] According to the number of data, adjust the number of data generation threads and / or adjust the number of training threads.

[0190] Each module in the above speech recognition model training device can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor of the computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above respective modules.

[0191] In one embodiment, a computer device is provided. The computer device can be a terminal, and its internal structure diagram can be as Figure 5 shown. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a mobile cellular network, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a speech recognition model training method. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, a trackball, or a touchpad provided on the shell of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0192] Those skilled in the art can understand that Figure 5 the structure shown in

[0193] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the speech recognition model training method of any of the above embodiments is implemented:

[0194] Obtain multiple first recognition results of each to-be-annotated audio file in the to-be-annotated audio file set; the multiple first recognition results are obtained by performing speech recognition on the to-be-annotated audio file by multiple speech recognition models;

[0195] Through the multiple first recognition results of each to-be-annotated audio file, perform automatic annotation processing on the corresponding to-be-annotated audio file to determine the annotation result and result confidence of each to-be-annotated audio file; the automatic annotation processing includes vote verification and segment verification;

[0196] Integrate each to-be-annotated audio file and the corresponding annotation result and result confidence to generate an annotation data set;

[0197] According to the annotation data set, train a preset speech recognition algorithm to obtain a target speech recognition model.

[0198] In one embodiment, a computer storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the speech recognition model training method of any of the above embodiments is implemented:

[0199] Obtain multiple first recognition results of each to-be-annotated audio file in the to-be-annotated audio file set; the multiple first recognition results are obtained by performing speech recognition on the to-be-annotated audio file by multiple speech recognition models;

[0200] Through the multiple first recognition results of each to-be-annotated audio file, perform automatic annotation processing on the corresponding to-be-annotated audio file to determine the annotation result and result confidence of each to-be-annotated audio file; the automatic annotation processing includes vote verification and segment verification;

[0201] Integrate each to-be-annotated audio file and the corresponding annotation result and result confidence to generate an annotation data set;

[0202] According to the annotation data set, train a preset speech recognition algorithm to obtain a target speech recognition model.

[0203] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data that have been authorized by the user or fully authorized by all parties.

[0204] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.

[0205] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0206] The above-described embodiments only represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for training a speech recognition model, characterized in that, The method for training the speech recognition model includes: Obtaining a plurality of first recognition results for each to-be-annotated audio file in the to-be-annotated audio file set; the plurality of first recognition results are obtained by performing speech recognition on the to-be-annotated audio file by a plurality of speech recognition models; Performing automatic annotation processing on the corresponding to-be-annotated audio file through the plurality of first recognition results of each to-be-annotated audio file, and determining the annotation result and result confidence of each to-be-annotated audio file; the automatic annotation processing includes vote verification and segment verification; Integrating each to-be-annotated audio file and the corresponding annotation result and result confidence to generate an annotation data set; Training a preset speech recognition algorithm according to the annotation data set to obtain a target speech recognition model.

2. The method for training a speech recognition model according to claim 1, wherein The performing automatic annotation processing on the corresponding to-be-annotated audio file through the plurality of first recognition results of each to-be-annotated audio file, and determining the annotation result and result confidence of each to-be-annotated audio file includes: For each to-be-annotated audio file, calculating the number of common characters between the target recognition result and other recognition results; the target recognition result is any one of the plurality of first recognition results; the other recognition results are the first recognition results other than the target recognition result; Taking the target recognition result with the largest number of common characters as the annotation result of the corresponding to-be-annotated audio file; Determining the matching segments between the annotation result and other recognition results respectively, and determining the result confidence of the annotation result through the number of the matching segments.

3. The method for training a speech recognition model according to claim 2, wherein The determining the matching segments between the annotation result and other recognition results respectively, and determining the result confidence of the annotation result through the number of the matching segments includes: Segmenting the annotation result by a preset text processing algorithm to obtain a plurality of first segments; Segmenting the other recognition results by a preset text processing algorithm to obtain a plurality of second segments; Matching each first segment with each second segment respectively to determine the matching segments; Taking the ratio of the number of the matching segments to the total number of the first segments as the result confidence of the annotation result.

4. The method for training a speech recognition model according to claim 1, wherein The integrating each to-be-annotated audio file and the corresponding annotation result and result confidence to generate an annotation data set includes: Taking the to-be-annotated audio file and the annotation result whose result confidence meets the preset confidence as annotation data; Obtaining a plurality of preset texts and a speech generation model; Inputting the preset text into the speech generation model to obtain a plurality of generated audio files; Taking the generated audio files and the corresponding preset texts as generated data; Integrating the generated data and the annotation data to generate an annotation data set.

5. The method for training a speech recognition model according to claim 4, wherein After taking the generated audio files and the corresponding preset texts as generated data, it further includes: Obtaining a plurality of second recognition results for each generated audio file; the plurality of second recognition results are obtained by performing speech recognition on the generated audio file by a plurality of speech recognition models; Based on multiple second recognition results of each generated audio file, perform automatic annotation processing on the corresponding generated audio file to determine the annotation result and result confidence of each generated audio file; Based on the annotation result and result confidence of each generated audio file, filter the multiple generated data to obtain the filtered generated data.

6. The method for training a speech recognition model according to claim 1, wherein Use multiple generated data threads to generate the annotation data set, and based on the annotation data set, train a preset speech recognition algorithm to obtain a target speech recognition model, including: Write the annotation data set into a data queue through multiple generated data threads; Read the annotation data set in the data queue item by item through multiple training threads and input it into the preset speech recognition algorithm for training to obtain a target speech recognition model.

7. The method for training a speech recognition model according to claim 6, wherein The method further includes: Obtain the number of data in the data queue; Based on the number of data, adjust the number of generated data threads and / or adjust the number of training threads.

8. A voice recognition model training device, characterized in that The device includes: An acquisition module, configured to acquire multiple first recognition results of each to-be-annotated audio file in a to-be-annotated audio file set; the multiple first recognition results are obtained by performing speech recognition on the to-be-annotated audio file by multiple speech recognition models; An automatic annotation module, configured to perform automatic annotation processing on the corresponding to-be-annotated audio file through multiple first recognition results of each to-be-annotated audio file to determine the annotation result and result confidence of each to-be-annotated audio file; the automatic annotation processing includes vote verification and segment verification; A data set generation module, configured to integrate each to-be-annotated audio file and the corresponding annotation result and result confidence to generate an annotation data set; A training module, configured to train a preset speech recognition algorithm based on the annotation data set to obtain a target speech recognition model.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images