Speech Recognition Model Training Method, Apparatus, and Computer Device

Abstract

The present application relates to a speech recognition model training method, device and computer equipment. The method includes: obtaining multiple first recognition results for each audio file to be annotated in a set of audio files to be annotated; the multiple first recognition results are obtained by multiple speech recognition models performing speech recognition on the audio files to be annotated; through the multiple first recognition results of each audio file to be annotated, the corresponding audio file to be annotated is automatically annotated to determine the annotation result and result confidence of each audio file to be annotated; the automatic annotation process includes voting verification and fragment verification; integrating each audio file to be annotated and the corresponding annotation result and result confidence to generate an annotation data set; based on the annotation data set, training a preset speech recognition algorithm to obtain a target speech recognition model. The use of this method can reduce training costs and improve model accuracy.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a speech recognition model training method, apparatus, and computer equipment. Background Art

[0002] As speech recognition technology becomes increasingly widely used in the field of artificial intelligence, the demand for its accuracy is increasing. Speech recognition models are typically trained using training data, which requires the collection of large amounts of labeled audio data. To ensure accuracy, training data is often annotated manually, but this method is time-consuming and labor-intensive, resulting in high labor costs.

[0003] Traditional technology also uses artificial intelligence for automatic labeling. However, due to the unstable quality of labeled audio data, in the case of unsupervised learning, the training data itself contains errors, which greatly reduces the accuracy of the speech recognition model.

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

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

[0006] In a first aspect, the present application provides a speech recognition model training method, the speech recognition model training method comprising:

[0007] Obtaining multiple first recognition results for each audio file to be annotated in the set of audio files to be annotated; the multiple first recognition results are obtained by performing speech recognition on the audio files to be annotated by multiple speech recognition models;

[0008] Performing automatic tagging processing on the corresponding audio files to be tagged based on the plurality of first recognition results of each audio file to be tagged, and determining a tagging result and a confidence level of each audio file to be tagged; the automatic tagging processing includes voting verification and segment verification;

[0009] Integrate each of the audio files to be annotated, the corresponding annotation results, and the result confidence levels to generate an annotation dataset;

[0010] The preset speech recognition algorithm is trained based on the labeled data set to obtain a target speech recognition model.

[0011] In one embodiment, automatically labeling the corresponding audio file to be labeled based on the plurality of first recognition results of each audio file to be labeled, and determining the labeling result and result confidence of each audio file to be labeled includes:

[0012] For each of the audio files to be annotated, calculating the number of common characters between a target recognition result and other recognition results; the target recognition result is any first recognition result among the multiple first recognition results; the other recognition results are first recognition results other than the target recognition result;

[0013] The target recognition result with the largest number of common characters is used as the labeling result of the corresponding audio file to be labeled;

[0014] The matching segments between the annotation result and other recognition results are determined, and the result confidence of the annotation result is determined according to the number of the matching segments.

[0015] In one embodiment, determining the matching segments between the annotation result and other recognition results, and determining the result confidence of the annotation result according to the number of the matching segments includes:

[0016] Segmenting the annotation result using a preset text processing algorithm to obtain a plurality of first segments;

[0017] Segmenting the other recognition results using a preset text processing algorithm to obtain a plurality of second segments;

[0018] Matching each of the first segments with each of the second segments to determine matching segments;

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

[0020] In one embodiment, integrating each audio file to be annotated and the corresponding annotation results and result confidences to generate an annotation dataset includes:

[0021] The audio files to be annotated and the annotation results whose result confidence meets the preset confidence level are used as annotation data;

[0022] Get multiple preset text and speech generation models;

[0023] Inputting the preset text into the speech generation model to obtain a plurality of generated audio files;

[0024] The generated audio file and the corresponding preset text are used as generated data;

[0025] The generated data and the labeled data are integrated to generate a labeled data set.

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

[0027] Obtaining multiple second recognition results for 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;

[0028] Automatically labeling the corresponding generated audio files based on the multiple second recognition results of each generated audio file, and determining the labeling result and result confidence of each generated audio file;

[0029] The plurality of generated data are screened according to the annotation result and the result confidence of each generated audio file to obtain screened generated data.

[0030] In one embodiment, the annotated data set is generated using multiple data generation threads, and the training of a preset speech recognition algorithm based on the annotated data set to obtain a target speech recognition model includes:

[0031] Writing the labeled data set into a data queue through multiple data generation threads;

[0032] The labeled data sets in the data queue are read one by one through multiple training threads, and are input into the preset speech recognition algorithm for training to obtain a target speech recognition model.

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

[0034] Obtaining the amount of data in the data queue;

[0035] According to the amount of data, the number of the data generation threads and / or the number of the training threads are adjusted.

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

[0037] An acquisition module is configured to acquire a plurality of first recognition results for each audio file to be annotated in the set of audio files to be annotated; the plurality of first recognition results are obtained by performing speech recognition on the audio files to be annotated by a plurality of speech recognition models;

[0038] An automatic annotation module, configured to automatically annotate the corresponding audio file to be annotated based on the plurality of first recognition results of each audio file to be annotated, and determine the annotation result and result confidence of each audio file to be annotated; the automatic annotation process includes voting verification and segment verification;

[0039] The dataset generation module is used to integrate each audio file to be annotated, the corresponding annotation results and the result confidence to generate an annotation dataset;

[0040] The training module is used to train the preset speech recognition algorithm according to the labeled data set to obtain a target speech recognition model.

[0041] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the above-described method when executing the computer program.

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

[0043] The above-mentioned speech recognition model training method, device and computer equipment obtain multiple first recognition results of each audio file to be labeled in the audio file set to be labeled; the multiple first recognition results are obtained by multiple speech recognition models performing speech recognition on the audio files to be labeled; the corresponding audio files to be labeled are automatically labeled based on the multiple first recognition results of each audio file to be labeled, and the labeling result and result confidence of each audio file to be labeled are determined; the automatic labeling process includes voting verification and fragment verification; each audio file to be labeled and the corresponding labeling result and result confidence are integrated to generate a labeling data set; a preset speech recognition algorithm is performed based on the labeling data set. Training is performed to obtain a target speech recognition model, and multiple speech recognition models are used to perform speech recognition separately to obtain the first recognition result, which can reduce the deviation of a single model, and automatic labeling including voting verification and fragment verification is performed based on multiple first recognition results to obtain labeling results and result confidence, which can reduce erroneous labeling and improve the consistency and reliability of the labeling results. Based on the labeling results and result confidence, a labeling dataset is further obtained, and a more accurate labeling dataset can be obtained for model training, realizing multi-model joint recognition and automatic verification, thereby reducing dependence on manual labeling, and at the same time improving model performance through high-quality labeling datasets, thereby achieving the effect of reducing training costs and improving model accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A diagram illustrating an application environment of a speech recognition model training method according to an embodiment;

[0045] Figure 2 1 is a flow chart of a method for training a speech recognition model in one embodiment;

[0046] Figure 3 Schematic diagram of a flow chart of a speech recognition model training method in another embodiment;

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

[0048] Figure 5 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0050] The speech recognition model training method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. The terminal 102 communicates with the server 104 through the 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 on the cloud or other network servers. The terminal 102 obtains multiple first recognition results for each audio file to be annotated in the audio file set to be annotated by communicating with the server 104; the multiple first recognition results are obtained by multiple speech recognition models performing speech recognition on the audio files to be annotated; through the multiple first recognition results of each audio file to be annotated, the corresponding audio files to be annotated are automatically annotated to determine the annotation results and result confidence of each audio file to be annotated; the automatic annotation processing includes voting verification and fragment verification; each audio file to be annotated and the corresponding annotation results and result confidence are integrated to generate an annotation data set; according to the annotation data set, the preset speech recognition algorithm is trained to obtain the target speech recognition model. The terminal 102 can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers and portable wearable devices. The portable wearable device may be a smart watch, a smart bracelet, a head-mounted device, etc. For example, the head-mounted device may also include AR glasses, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers.

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

[0052] Step S100: obtaining a plurality of first recognition results for each audio file to be annotated in the set of audio files to be annotated.

[0053] The audio file set to be annotated can be a collection of unannotated audio files, and the audio data can be audio data used for speech recognition model training, which can be collected from actual application scenarios or obtained from public datasets. For example, the audio file set to be annotated can include, but is not limited to, phone recordings, meeting minutes, and voice commands.

[0054] The multiple first recognition results are obtained by performing speech recognition on the audio file to be annotated using multiple speech recognition models. For example, each audio file to be annotated can be input into each selected speech recognition model, and the first recognition results output by each speech recognition model can be collected, thereby obtaining multiple first recognition results for a single audio file to be annotated. It will be appreciated that by combining the recognition results of multiple speech recognition models for annotation, the potential bias of a single model can be reduced, thereby improving the accuracy of the annotation results.

[0055] Step S200 : Automatically label the corresponding audio file to be labeled based on the plurality of first recognition results of each audio file to be labeled, and determine a labeling result and a result confidence level of each audio file to be labeled.

[0056] The annotation result may be the text content of the finalized audio file, and the result confidence may indicate the credibility of the annotation result. For example, the result confidence may include but is not limited to a voting verification score, a fragment verification score, and the like.

[0057] The automatic annotation process includes voting verification and fragment verification. Voting verification can be a verification rule based on the majority principle, and fragment verification can be a verification rule based on text segmentation. In this embodiment, voting verification can be used only to determine the annotation result or the result confidence, or it can be used to determine the annotation result and the result confidence at the same time; fragment verification can also be used only to determine the annotation result or the result confidence, or it can be used to determine the annotation result and the result confidence at the same time.

[0058] In an exemplary embodiment, voting verification can be used to annotate results and fragment verification can be used to calculate result confidence. Alternatively, fragment verification can be used to annotate results and voting verification can be used to calculate result confidence.

[0059] In another exemplary embodiment, voting verification can be used to determine the first annotation result and the first result confidence respectively, and fragment verification can be used to determine the second annotation result and the second result confidence; if the first annotation result is the same as the second annotation result, the annotation result is determined by merging, and the result confidence is determined comprehensively based on the first result confidence and the second result confidence; 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 based on the size relationship between the first result confidence and the second result confidence; accordingly, the result confidence can be determined by using the result confidence 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 result confidence value can be increased according to a preset ratio, and when the first annotation result and the second annotation result are different, the result confidence value can be reduced according to a preset ratio.

[0060] When voting verification is used to determine the labeling result, the character overlap 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 can be used as the labeling result; in another exemplary embodiment, for the same audio file to be labeled, the characters whose total occurrence times exceed a preset threshold in multiple first recognition results can also be counted. The preset threshold can be determined by prior knowledge, and then the first recognition result containing the largest number of the above-mentioned characters in the recognition results can be used as the labeling result; other methods can also be used to determine the first recognition result with the highest accuracy based on the majority principle, which is not limited in this embodiment.

[0061] When voting verification is used to determine the confidence of the result, the character overlap 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, the characters whose total number of occurrences exceeds a preset threshold in multiple first recognition results can also be counted. The preset threshold can be determined through prior knowledge, and then the result confidence is determined based on the proportion of the above-mentioned number of characters in each first recognition result in the total number of words; other methods of determining the result confidence based on the majority principle can also be used, which is not limited in this embodiment.

[0062] When using fragment verification to determine the labeling results and / or result confidence, an audio segmentation method based on sound wave amplitude or deep learning can be used to obtain an audio segment for each first recognition result, and then the text segment corresponding to each audio segment is determined with the first recognition result. Alternatively, a text segmentation method based on sliding windows or semantics can be used to directly obtain multiple text segments for each first recognition result.

[0063] When using fragment verification to determine the annotation results, based on the obtained text fragments, the text fragment overlap between each first recognition result and other first recognition results can be compared to determine the frequency of occurrence of the text fragments of each first recognition result, and the first recognition result with the highest frequency of occurrence of the text fragments can be used as the annotation result; in another exemplary embodiment, for the same audio file to be annotated, the text fragments whose total number of occurrences exceeds a preset threshold in multiple first recognition results can also be counted, and the preset threshold can be determined based on prior knowledge, and then the first recognition result containing the largest number of the above-mentioned text fragments in the recognition results can be used as the annotation result; other methods can also be used to determine the first recognition result with the highest accuracy based on text segmentation, which is not limited in this embodiment.

[0064] When using fragment verification to determine the result confidence, the text fragment overlap 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, the text fragments whose total number of occurrences exceeds a preset threshold in multiple first recognition results can also be counted. The preset threshold can be determined based on prior knowledge, and then the result confidence is determined based on the proportion of the above-mentioned text fragments in each first recognition result to the total number of fragments; other methods of determining the result confidence based on the text segmentation principle can also be used, which is not limited in this embodiment.

[0065] This embodiment obtains annotation results and result confidence levels through automatic annotation processing including voting verification and fragment verification, thereby reducing erroneous annotations and improving the consistency and reliability of annotation results.

[0066] Step S300 : Integrate each of the audio files to be annotated, the corresponding annotation results, and the result confidence levels to generate an annotation dataset.

[0067] The annotated dataset can be a data set containing audio files and their corresponding annotation results and confidence levels, which is used for subsequent model training. For example, the annotated dataset 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 level.

[0068] The result confidence can be used as an indicator to screen high-quality annotated data, and data with higher result confidence can be used to generate annotated datasets. For example, audio files with result confidence higher than a preset confidence threshold and the corresponding annotation results can be used to form an annotated dataset.

[0069] Step S400: training a preset speech recognition algorithm based on the labeled data set to obtain a target speech recognition model.

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

[0071] To train a preset speech recognition algorithm, you can divide the labeled dataset into a training set and a validation set, train the preset speech recognition algorithm using the training set, adjust model parameters to minimize the loss function, evaluate model performance using the validation set, adjust hyperparameters (such as learning rate and batch size) to optimize the model, and save the trained model as the target speech recognition model.

[0072] The present embodiment provides a speech recognition model training method, which obtains multiple first recognition results for each audio file to be annotated in a set of audio files to be annotated; the multiple first recognition results are obtained by performing speech recognition on the audio files to be annotated by multiple speech recognition models; the corresponding audio files to be annotated are automatically annotated based on the multiple first recognition results of each audio file to be annotated, and the annotation result and result confidence of each audio file to be annotated are determined; the automatic annotation process includes voting verification and segment verification; each audio file to be annotated and the corresponding annotation result and result confidence are integrated to generate an annotation data set; a preset speech recognition algorithm is tested based on the annotation data set. Training is performed to obtain a target speech recognition model. Multiple speech recognition models are used to perform speech recognition respectively to obtain the first recognition result, which can reduce the deviation of a single model. Automatic labeling including voting verification and fragment verification is performed based on multiple first recognition results to obtain labeling results and result confidence. This can reduce incorrect labeling and improve the consistency and reliability of the labeling results. Based on the labeling results and result confidence, a labeling dataset is further obtained, and a more accurate labeling dataset can be obtained for model training, realizing multi-model joint recognition and automatic verification, thereby reducing dependence on manual labeling. At the same time, the model performance is improved through high-quality labeling datasets, thereby achieving the effect of reducing training costs and improving model accuracy.

[0073] In one embodiment, voting verification can be achieved by counting the number of common characters between each first recognition result and other first recognition results, and fragment verification can be achieved by verifying the overlap of the marked result and the fragments of other first recognition results.

[0074] The step of automatically labeling the corresponding audio file to be labeled based on the plurality of first recognition results of each audio file to be labeled, and determining the labeling result and result confidence of each audio file to be labeled includes:

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

[0076] The target recognition result with the largest number of common characters is used as the labeling result of the corresponding audio file to be labeled;

[0077] The matching segments between the annotation result and other recognition results are determined, and the result confidence of the annotation result is determined according to the number of the matching segments.

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

[0079] To determine the matching segments between the annotation results and other recognition results, the method may be to divide the text into multiple segments and then compare the recurrence of these segments in other recognition results. For example, after obtaining multiple text segments through an audio segmentation method based on sound wave amplitude or deep learning, or a text segmentation method based on semantics, the ratio of the number of segments in the annotation results that appear in other recognition results to the total number of segments is calculated, thereby comprehensively calculating the confidence level of the result. Furthermore, a weighted average method may be used to assign a higher weight to segments with a higher degree of segment overlap, thereby achieving a more detailed assessment of the reliability of the annotation results and further improving the reliability of the results 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 labeling by counting the number of common characters and taking the target recognition result with the largest number of common characters as the labeling result; by calculating the result confidence by the matching results of the fragments, the verification ability of local details can be further improved, ensuring that the labeling results are highly consistent at both the global and local levels, thereby achieving the effect of improving the accuracy of the labeling results and the reliability of the result confidence.

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

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

[0083] Segmenting the other recognition results using a preset text processing algorithm to obtain a plurality of second segments;

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

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

[0086] 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 (e.g., segmentation by punctuation, spaces, etc.) and deep learning-based segmentation (e.g., using a Transformer model to predict text boundaries). In one specific embodiment, the segmentation granularity can be determined based on the structural features of the input text, and the preset text processing algorithm is used for segmentation based on this segmentation granularity.

[0087] The first segments may be multiple small text segments extracted from the annotation result by a preset text processing algorithm. In a specific embodiment, the annotation result "open chat" may be divided into two first segments: "open" and "chat".

[0088] The second segment can be multiple small text segments extracted from other recognition results using the same preset text processing algorithm. For example, other recognition results can be "Open also need ingredients", "Open friends", "OK also friends", etc. Accordingly, "Open also need ingredients" can be divided into three second segments: "Open", "Also", and "Ingredients".

[0089] In order to improve the comparability between the first segment and the second segment, the preset text processing algorithm may use the same granularity when segmenting the annotation results and other first speech recognition results.

[0090] Matching the first segment and the second segment may be calculating the similarity or consistency between the first segment and the second segment. In a specific embodiment, when the matching requirement is strict, the matching condition may be determining whether the first segment and the second segment completely overlap.

[0091] A matching segment may be a segment with a high consistency in multiple first recognition results. For example, if a first segment appears in a preset number of second segments, the segment may be considered a matching segment.

[0092] Exemplarily, to determine the matching fragments, each first fragment can be traversed and compared with all second fragments one by one, the number of successful matches of each first fragment in the second fragment set can be recorded, and the first fragment whose number of matches exceeds a preset threshold can be determined as the matching fragment, so that matching fragments with higher consistency can be screened out, reflecting the consistency of the annotation results at the fragment level.

[0093] The ratio of the number of matching segments to the total number of the first segments is used as the confidence level of the annotation result, which can more accurately reflect the overall reliability of the annotation result. It can be understood that the higher the ratio, the higher the reliability of the annotation result at the segment level.

[0094] This embodiment provides a speech recognition model training method, which obtains multiple first segments by segmenting the annotation results using a preset text processing algorithm; obtains multiple second segments by segmenting the other recognition results using a preset text processing algorithm; matches each of the first segments with each of the second segments to determine matching segments; and uses the ratio of the number of matching segments to the total number of the first segments as the result confidence of the annotation result, thereby achieving the technical effects of improving the accuracy of the result confidence, enhancing the reliability of the annotation results, and optimizing the quality of training data.

[0095] In one embodiment, integrating each audio file to be annotated and the corresponding annotation results and result confidences to generate an annotation dataset includes:

[0096] The audio files to be annotated and the annotation results whose result confidence meets the preset confidence level are used as annotation data;

[0097] Get multiple preset text and speech generation models;

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

[0099] The generated audio file and the corresponding preset text are used as generated data;

[0100] The generated data and the labeled data are integrated to generate a labeled data set.

[0101] The annotated data may be the audio files and their corresponding annotated results retained after screening for confidence. The preset confidence level may be a pre-set confidence threshold used to screen for high-quality annotated data, which may be derived from prior knowledge. For example, the preset confidence level may include, but is not limited to, specific values such as 0.8 and 0.9.

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

[0103] The preset text may be a set of pre-prepared text data used to generate speech data, which may be manually written or obtained from a public data set. For example, the preset text may include but is not limited to news reports, literary excerpts, daily conversations, etc.

[0104] A text-to-speech (TTS) model can be a model that converts text into speech signals, and may include, but is not limited to, a deep learning-based Tacotron model or a WaveNet model. For example, the preset TTS model can be a multilingual or multi-language TTS model.

[0105] In this embodiment, by collecting preset texts and using a speech synthesis model to convert each preset text into a corresponding speech segment, an audio file can be generated, thereby generating additional training data through the speech synthesis model and increasing the diversity of the training data.

[0106] In this embodiment, the annotated data set may be an integrated result including the annotated data and the generated audio file, and is used for model training of a preset speech recognition algorithm.

[0107] This embodiment provides a speech recognition model training method, which uses the audio files to be annotated and their annotation results that meet the preset confidence level as annotation data to screen out high-quality annotation data to ensure the accuracy of the training data; at the same time, multiple generated audio files are generated through the speech synthesis model and preset text, which can make up for the lack of diversity in the training data. Finally, the annotated data and the generated audio files are integrated to generate an annotated data set, which can construct a more comprehensive training data set, thereby improving the quality and diversity of the annotated data, optimizing the model training effect, and achieving the effect of improving the accuracy and generalization ability of the target speech recognition model.

[0108] In one embodiment, after generating the audio file and the corresponding preset text as generated data, the method further includes:

[0109] Obtaining a plurality of second recognition results for each of the generated audio files;

[0110] Automatically labeling the corresponding generated audio files based on the multiple second recognition results of each generated audio file, and determining the labeling result and result confidence of each generated audio file;

[0111] The plurality of generated data are screened according to the annotation result and the result confidence of each generated audio file to obtain screened 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, all generated audio files may be traversed, each generated audio file may be input into the selected multiple speech recognition models one by one, and the second recognition result output by each speech recognition model may be collected.

[0114] The corresponding generated audio files are automatically annotated based on the multiple second recognition results of each generated audio file. This may be verified using a method such as multiple speech recognition results, which is not described in detail in this embodiment.

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

[0116] The present embodiment provides a speech recognition model training method, which can reduce the deviation that may exist in a single model by obtaining multiple second recognition results for each generated audio file and using multiple speech recognition models to perform speech recognition on each generated audio file separately; automatically annotate the corresponding generated audio file through the multiple second recognition results of each generated audio file, determine the annotation results and result confidence of each generated audio file, thereby improving the accuracy and consistency of the annotation results; screen the multiple generated audio files through the annotation results and result confidence to obtain multiple screened 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 embodiment, the annotated data set is generated using multiple data generation threads, and the training of a preset speech recognition algorithm based on the annotated data set to obtain a target speech recognition model includes:

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

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

[0120] The data queue may be a queue for storing data waiting to be used for training. For example, the data queue may include audio files and their corresponding annotation results.

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

[0122] The data generation thread may be a thread for generating training data. Exemplarily, the data generation thread may include a thread for automatically annotating audio files to be annotated, or a thread for generating audio files by using preset text.

[0123] The training thread may be a thread used to train a preset speech recognition algorithm. The number of training threads may be determined based on prior knowledge, or may be adjusted based on the real-time status of the data queue.

[0124] This embodiment provides a speech recognition model training method, which 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. Data can be generated synchronously during the model training process, thereby avoiding the need to generate all training data and reducing the time waste caused by step-by-step processing, thereby achieving the effect of improving model training efficiency.

[0125] In one embodiment, the method further comprises:

[0126] Obtaining the amount of data in the data queue;

[0127] According to the amount of data, the number of the data generation threads and / or the number of the training threads are adjusted.

[0128] The data quantity may be the number or size of real-time data stored in the current data queue, which may be obtained by reading the state information of the queue and calculating the total number of bytes or entries.

[0129] According to the amount of data, the number of data generation threads and / or the number of training threads are adjusted, and it can be determined whether the number of current data generation threads and training threads needs to be adjusted based on the amount of data. For example, when the amount of data in the data queue exceeds the preset data amount condition for a long time, it can be considered that the current data generation thread has a strong ability to generate annotation data or generate audio files, and the processing ability of the training thread is relatively weak, and the speed matching of the generated data and the training data can be achieved by increasing the number of training threads or reducing the number of data generation threads; conversely, when the amount of data in the data queue is lower than the preset data amount condition for a long time, it can be considered that the current data generation thread has a weak ability to generate annotation data or generate audio files, and the processing ability of the training thread is relatively strong, and the speed matching of the generated data and the training data can be achieved by reducing the number of training threads or increasing the number of data generation threads.

[0130] In this embodiment, by adjusting the number of data generation threads and / or the number of training threads according to the amount of data, the speech generation speed can be dynamically adjusted to avoid the problem of training interruption or insufficient data caused by too little data in the data queue, while ensuring that the data queue is always maintained within a reasonable data volume range, thereby improving data utilization efficiency and the continuity of model training.

[0131] A speech recognition model training method provided in this embodiment monitors the amount of data in the data queue and adjusts the number of data generation threads and / or the number of training threads based on the data amount. It can dynamically adjust the timing of generating audio files and adding annotation data, thereby avoiding unnecessary waste of resources and further ensuring the stability and continuity of the model training process, thereby achieving the effect of improving resource utilization and model training efficiency.

[0132] In order to more clearly illustrate the technical solution of this application, this application also provides a detailed embodiment.

[0133] In one embodiment, a method for training a speech recognition model is provided, which is applied to the speech interaction scenario of smart glasses, and is aimed at enhancing the whisper basic model in specific areas, such as Figure 3 As shown, the training methods include:

[0134] Step S1, data preparation.

[0135] Step S11: Obtain files and tools for generating data. The files may include multiple timbre files, and the tools may include a Text-to-Speech (TTS) model (e.g., the ChatTTS model) and multiple Automatic Speech Recognition (ASR) models. This embodiment uses four speech recognition models (e.g., the whisper-large-v3 model, the paraformer model, the paddle-speech model, and the qwen-audio model) as an example. The TTS model generates audio files, using random or specified timbres to convert domain text into speech. The sampling rate is converted to the sampling rate set in the target speech recognition model. In one specific embodiment, the sampling rate is 16k.

[0136] Step S12: Labeling data. This involves collecting ASR corpus (e.g., voice data from the CommonVoice platform) in the target language (e.g., Chinese, English) for voice interaction scenarios, including audio files and corresponding text.

[0137] Step S13: Obtain unlabeled data. In this embodiment, audio files collected by the MR glasses may be obtained.

[0138] Step S2: data labeling.

[0139] For unlabeled audio data, we use a multi-model voting and information segment compensation correction method to automatically label it, where:

[0140] Four ASR models are used to recognize an audio file, and four speech recognition results are obtained. These four speech recognition results are then fused to obtain the most reliable result. For example, the four results include "I close the 3D island", "Close the 3D navigation", "Three closes and three low falls", and "Close the 3D navigation".

[0141] Find the result that has the most characters in common with the other three results, which is "Turn off 3D navigation". Then use other speech recognition results to vote and verify the fragments in this speech recognition result. For each fragment with a length of at least 2 characters, as long as there are other results that have the same fragment, it can be considered correct. The proportion of unconfirmed fragments and correct fragments is returned.

[0142] Highlight the unconfirmed segments in the most reliable result in the spreadsheet document for easy review. If the same segments can be found in other results for the three segments "off", "3d", and "navigation", then the result can be considered correct.

[0143] In another specific embodiment, the four voice recognition results include "open also", "open chat", "ok also Pengtian" and "open ingredients", and the most reliable result is "open chat", then the return information is [{"diff_span":[2,4], "gt":"chat"}], where diff_span is the unconfirmed fragments and the total number of fragments, and gt is the unconfirmed fragment content. At this time, the proportion of correct fragments, that is, the result confidence is 0.5.

[0144] After performing the above processing on a batch of audio files, the results are saved in a spreadsheet document (such as an Excel file). Samples with empty segments that cannot be confirmed and samples with a correct segment ratio greater than a threshold, such as samples with a ratio exceeding a threshold of 0.9, are filtered out as training samples. In this way, a large amount of high-quality labeled data can be quickly obtained in the absence of labeling resources.

[0145] Step S3: model training.

[0146] The model training process uses data generators and data loaders, where:

[0147] The data generator prepares a data queue and launches multiple background threads to generate audio samples and write them to the data queue. The data in the data queue comes from generated audio files and annotated data. The ratio of generated audio files to annotated data can be determined based on prior knowledge. For example, generated audio files can account for 10% and annotated data can account for 90%.

[0148] Each background thread runs files and tools (multiple timbre files, TTS model, ASR model) for generating audio files to generate audio samples.

[0149] To ensure the quality of the generated audio, the generated data is recognized using multiple speech recognition models, and voting and information segment compensation correction are performed based on the recognition results. The quality of the generated audio file can be determined by examining the recognition results of the generated audio. Generated audio files that can be corrected through voting to obtain the original text results are considered high-quality audio, while audio data that cannot be corrected is discarded. Therefore, samples with empty segments that cannot be confirmed in the voting results, or samples with a correct segment ratio greater than a threshold (such as 0.9), can be considered high-quality audio.

[0150] The data queue is protected from multi-threaded access. When the amount of data in the data queue exceeds a threshold of 300, writing to the queue is suspended.

[0151] The data loader is used to load dynamically generated data in the data queue one by one, rather than all at once. This allows training data to be generated online and can use data composed of richer tones and text to train the model. At the same time, it reduces the time for data generation and avoids the waiting time caused by the slow generation of high-quality audio data using the TTS model.

[0152] The data loader reads data from the data queue and uses it for model training.

[0153] Furthermore, after a piece of 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 file and the annotated data.

[0154] When there is no data in the queue, the number of queues is tracked in the reading and generating threads to adjust the number of generating threads. If the number is insufficient or the waiting time is long, more threads are added to meet the demand 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 to manage the process;

[0156] Furthermore, the main training thread and the data generation thread can use different graphics cards. Each thread generates samples in batches. The test obtains the configuration that can generate the maximum number of samples per second, for example, 16 samples in a single batch.

[0157] This embodiment provides a speech recognition model training method that integrates dynamic data generation into the training process. This allows the model to be trained using data composed of a richer combination of timbre and text, while reducing the time required for data generation. Furthermore, the data annotation process allows for rapid acquisition of a large amount of high-quality annotated data in the absence of annotation resources, thereby reducing training costs and improving model accuracy.

[0158] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0159] Based on the same inventive concept, the present application also provides a speech recognition model training device for implementing the speech recognition model training method described above. The solution to the problem provided by this device is similar to the solution described in the method described above. Therefore, the specific limitations of one or more speech recognition model training device embodiments provided below can be found in the above-mentioned limitations of the speech recognition model training method and will not be repeated here.

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

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

[0162] The automatic tagging module 200 is configured to automatically tag the corresponding audio file to be tagged based on the multiple first recognition results of each audio file to be tagged, and determine the tagging result and result confidence of each audio file to be tagged. The automatic tagging process includes voting verification and segment verification.

[0163] The data set generation module 300 is used to integrate each audio file to be annotated and the corresponding annotation results and result confidence to generate an annotated data set.

[0164] The training module 400 is used to train a preset speech recognition algorithm based on the annotated data set to obtain a target speech recognition model.

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

[0166] For each of the audio files to be annotated, calculating the number of common characters between a target recognition result and other recognition results; the target recognition result is any first recognition result among the multiple first recognition results; the other recognition results are first recognition results other than the target recognition result;

[0167] The target recognition result with the largest number of common characters is used as the labeling result of the corresponding audio file to be labeled;

[0168] The matching segments between the annotation result and other recognition results are determined, and the result confidence of the annotation result is determined according to the number of the matching segments.

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

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

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

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

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

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

[0175] The audio files to be annotated and the annotation results whose result confidence meets the preset confidence level are used as annotation data;

[0176] Get multiple preset text and speech generation models;

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

[0178] The generated audio file and the corresponding preset text are used as generated data;

[0179] The generated data and the labeled data are integrated to generate a labeled data set.

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

[0181] Obtaining multiple second recognition results for 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] Automatically labeling the corresponding generated audio files based on the multiple second recognition results of each generated audio file, and determining the labeling result and result confidence of each generated audio file;

[0183] The plurality of generated data are screened according to the annotation result and the result confidence of each generated audio file to obtain screened generated data.

[0184] In one embodiment, multiple data generation threads are used to generate the labeled data set, and the training module 400 is further configured to:

[0185] Writing the labeled data set into a data queue through multiple data generation threads;

[0186] The labeled data sets in the data queue are read one by one through multiple training threads, and are input 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] Obtaining the amount of data in the data queue;

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

[0190] Each module in the speech recognition model training device described above can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0191] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, memory, a communication interface, a display screen, and an input device connected via a system bus. 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 computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal via wired or wireless communication. The wireless communication can be achieved via Wi-Fi, a mobile cellular network, NFC (near-field communication), or other technologies. When executed by the processor, the computer program implements a method for training a speech recognition model. The display screen of the computer device can be a liquid crystal display or an electronic ink display. The input device of the computer device can be a touch layer covering the display screen, or keys, a trackball, or a touchpad provided on the computer device housing, or an external keyboard, touchpad, or mouse.

[0192] Those skilled in the art will understand that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

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

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

[0195] Performing automatic tagging processing on the corresponding audio files to be tagged based on the plurality of first recognition results of each audio file to be tagged, and determining a tagging result and a confidence level of each audio file to be tagged; the automatic tagging processing includes voting verification and segment verification;

[0196] Integrate each of the audio files to be annotated, the corresponding annotation results, and the result confidence levels to generate an annotation dataset;

[0197] The preset speech recognition algorithm is trained based on the labeled data set 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] Obtaining multiple first recognition results for each audio file to be annotated in the set of audio files to be annotated; the multiple first recognition results are obtained by performing speech recognition on the audio files to be annotated by multiple speech recognition models;

[0200] Performing automatic tagging processing on the corresponding audio files to be tagged based on the plurality of first recognition results of each audio file to be tagged, and determining a tagging result and a confidence level of each audio file to be tagged; the automatic tagging processing includes voting verification and segment verification;

[0201] Integrate each of the audio files to be annotated, the corresponding annotation results, and the result confidence levels to generate an annotation dataset;

[0202] The preset speech recognition algorithm is trained based on the labeled data set 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 used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0204] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the 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 above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. 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), magnetic 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 take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), data processing logic devices based on quantum computing, and the like.

[0205] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, 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, they should be considered to be within the scope of this specification.

[0206] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A speech recognition model training method, characterized in that: The speech recognition model training method comprises: Obtaining multiple first recognition results for each audio file to be annotated in the set of audio files to be annotated; the multiple first recognition results are obtained by performing speech recognition on the audio files to be annotated by multiple speech recognition models; Automatically annotating the corresponding audio file to be annotated based on the plurality of first recognition results for each audio file to be annotated, and determining an annotation result and a confidence level of each audio file to be annotated; the automatic annotation process includes voting verification and segment verification; the annotation result is a target recognition result with the largest number of common characters; the target recognition result is any one of the plurality of first recognition results; Integrate each of the audio files to be annotated, the corresponding annotation results, and the result confidence levels to generate an annotation dataset; The preset speech recognition algorithm is trained based on the labeled data set to obtain a target speech recognition model.

2. The speech recognition model training method according to claim 1, characterized in that The step of automatically labeling the corresponding audio file to be labeled based on the plurality of first recognition results of each audio file to be labeled, and determining the labeling result and result confidence of each audio file to be labeled includes: For each of the audio files to be annotated, calculating the number of common characters between the target recognition result and other recognition results; the other recognition results are first recognition results other than the target recognition result; The target recognition result with the largest number of common characters is used as the labeling result of the corresponding audio file to be labeled; The matching segments between the annotation result and other recognition results are determined, and the result confidence of the annotation result is determined according to the number of the matching segments.

3. The speech recognition model training method according to claim 2, characterized in that: Determining the matching segments between the annotation result and other recognition results, and determining the result confidence of the annotation result according to the number of the matching segments includes: Segmenting the annotation result using a preset text processing algorithm to obtain a plurality of first segments; Segmenting the other recognition results using a preset text processing algorithm to obtain a plurality of second segments; Matching each of the first segments with each of the second segments to determine matching segments; The ratio of the number of the matching segments to the total number of the first segments is used as the result confidence of the annotation result.

4. The speech recognition model training method according to claim 1, characterized in that The step of integrating each audio file to be annotated and the corresponding annotation results and result confidence to generate an annotation dataset includes: The audio files to be annotated and the annotation results whose result confidence meets the preset confidence level are used as annotation data; Get multiple preset text and speech generation models; Inputting the preset text into the speech generation model to obtain a plurality of generated audio files; The generated audio file and the corresponding preset text are used as generated data; The generated data and the labeled data are integrated to generate a labeled data set.

5. The speech recognition model training method according to claim 4, characterized in that: After generating the audio file and the corresponding preset text as generated data, the method further includes: Obtaining multiple second recognition results for 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; Automatically annotating the corresponding generated audio files based on the plurality of second recognition results of each generated audio file, and determining an annotation result and a result confidence level of each generated audio file; The plurality of generated data are screened according to the annotation result and the result confidence of each generated audio file to obtain screened generated data.

6. The speech recognition model training method according to claim 1, characterized in that: The method of generating the annotated data set by using a plurality of data generation threads and training a preset speech recognition algorithm based on the annotated data set to obtain a target speech recognition model includes: Writing the labeled data set into a data queue through multiple data generation threads; The labeled data sets in the data queue are read one by one through multiple training threads, and are input into the preset speech recognition algorithm for training to obtain a target speech recognition model.

7. The speech recognition model training method according to claim 6, characterized in that: The method further comprises: Obtaining the amount of data in the data queue; According to the amount of data, the number of the data generation threads and / or the number of the training threads are adjusted.

8. A speech recognition model training device, characterized in that: The device comprises: An acquisition module is configured to acquire a plurality of first recognition results for each audio file to be annotated in the set of audio files to be annotated; the plurality of first recognition results are obtained by performing speech recognition on the audio files to be annotated by a plurality of speech recognition models; An automatic tagging module is configured to automatically tag the corresponding audio files to be tagged based on the multiple first recognition results of each audio file to be tagged, and determine the tagging result and result confidence of each audio file to be tagged; the automatic tagging process includes voting verification and segment verification; the tagging result is the target recognition result with the largest number of common characters; the target recognition result is any one of the multiple first recognition results; The dataset generation module is used to integrate each audio file to be annotated, the corresponding annotation results and the result confidence to generate an annotation dataset; The training module is used to train the preset speech recognition algorithm according to the labeled data set to obtain a target speech recognition model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: 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