A method for quickly labeling acoustic model training data in speech recognition

Abstract

The application discloses a method for quickly marking acoustic model training data in speech recognition. The method uses a trained basic acoustic model to perform speech recognition work; recorded incorrect words and sentences are cut by using an audio cutting tool to obtain incorrect audio files WrongPart; words corresponding to the WrongPart in a correct original text sentence prepared in advance are deleted to obtain remaining text ResidueScript in the original text, and a deletion position is marked; corresponding audio files ResidueWav and the WrongPart are combined according to original text positions to obtain one piece of acoustic model marking data, and finally, a plurality of different TrainFile are combined to form a training set for acoustic model fine tuning, the acoustic model is fine tuned, and the purpose of efficiently correcting error-prone parts of ASR is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of speech recognition, in particular to a method for quickly preparing training data of an acoustic model in speech recognition. BACKGROUND

[0002] Speech recognition technology has become a major technology in current artificial intelligence applications, which mainly solves the problem of converting speech into text by computer to obtain the corresponding word or character sequence from a given waveform sequence. Generally speaking, a speech recognition system mainly consists of four modules of front-end processing, acoustic model, language model and decoder. The front-end processing mainly includes three aspects of operations: endpoint detection, noise reduction and feature extraction. The acoustic model, language model and decoder belong to the back-end processing. The acoustic model is mainly used to construct the probability mapping relationship between speech and output acoustic units, and the language model is mainly used to describe the probability collocation relationship between different words, so that the recognized sentence is more smooth and natural.

[0003] The process of constructing a speech recognition system generally includes two parts of training and recognition. The training specifically refers to the training of the acoustic model and the language model, which is generally performed offline. The recognition is the actual working process of recognizing the user's speech into text, which is generally performed online.

[0004] In the training of the acoustic model, a large amount of sample labeled audio data is needed, which is mainly obtained by manual work, which is time-consuming, labor-intensive and costly. If the labeled audio data after speech recognition is directly used as the training sample of the acoustic model, the quality of the labeled audio data after speech recognition is unstable, which leads to low accuracy of the acoustic model trained based on the sample.

[0005] In addition, after obtaining the basic acoustic model, some types of words may cause recognition errors due to insufficient training data. Therefore, appropriate error-prone samples can be added as training data to continue fine-tuning the existing speech recognition model to improve the recognition effect. The preparation of error-prone samples is usually manual collection, so the process of preparing such training data is slow, which affects the rapid iteration and optimization of the model and leads to difficulty in efficiently improving the accuracy of speech recognition. SUMMARY

[0006] Therefore, in view of the above technical problems, the present application provides a method for quickly labeling acoustic model training data in speech recognition.

[0007] In a first aspect, a method for quickly labeling acoustic model training data in speech recognition includes:

[0008] using the trained basic acoustic model to perform speech recognition work, in which there are recognized errors in words and sentences;

[0009] For the recorded recognition error words and sentences, cutting is performed using an audio cutting tool to obtain an audio file of the recognition error, denoted as WrongPart, WrongPart containing one or more words recognized;

[0010] The words in the pre-prepared correct original text sentence corresponding to WrongPart are deleted, and the deletion position is marked; the remaining text in the original text obtained after deletion is denoted as ResidueScript;

[0011] ResidueScript is converted into an audio file using a speech synthesis system, denoted as ResidueWav, and ResidueWav and WrongPart are combined according to the original text position, so that WrongPart fills the deletion position, to obtain a piece of acoustic model annotation data, denoted as TrainFile;

[0012] A plurality of different TrainFiles constitute a training set TrainFiles for acoustic model fine-tuning.

[0013] In the above scheme, optionally, the pre-prepared correct original text sentence is from a text sentence set prepared in the following manner:

[0014] According to the business classification of the set industry, common business types are established, and corresponding text sentence sets are obtained for each business type for standby use;

[0015] For text sentences of different business types, cutting and recombination are performed, and expanded text sentences are obtained for each business type to collectively constitute a text sentence set under the business type.

[0016] In the above scheme, optionally, the trained basic acoustic model is obtained in the following manner:

[0017] A batch of audio data to be labeled in a set industry is obtained, and is divided according to the speech subject identity type;

[0018] For the audio data to be labeled of the determined speech subject identity type, two existing speech recognition tools are called to perform recognition respectively, and corresponding first and second preliminary recognition result data sets are obtained;

[0019] The text edit distance algorithm is used to calculate the text edit distance between the first preliminary recognition result data set and the second preliminary recognition result data set, and the recognition result and the corresponding original audio with an edit distance less than a preset threshold are reserved as main data and added to the pre-labeled set, and for the recognition result with an edit distance greater than and equal to the preset threshold, part of the recognition result and the corresponding original audio are extracted as additional data and also added to the pre-labeled set; the total audio duration of the main data is much greater than the total audio duration of the source data of the additional data.

[0020] The pre-labeled set is used as a basic training set for training an acoustic model after verification.

[0021] In the above scheme, optionally, the set industry is an industry related to a customer service field, and the voice subject identity type is divided into customer service and customer; accordingly, the division according to the voice subject identity type is specifically channel separation of the audio data according to left and right channels.

[0022] In the above scheme, optionally, before the two existing speech recognition tools are called for identification, VAD is used to cut the audio to obtain individual sentence audio.

[0023] In the above scheme, optionally, the two existing speech recognition tools adopt a Tencent speech recognition ASR interface and an Ali speech recognition ASR interface.

[0024] In the above scheme, optionally, the total audio duration with an edit distance greater than and equal to the preset threshold is 20% to 40% of the total audio duration with an edit distance less than the preset threshold.

[0025] In the above scheme, optionally, the additional data is determined in the following manner:

[0026] The data with an edit distance greater than and equal to the preset threshold is further subdivided into high, medium and low parts, and part of the data is randomly extracted from the three parts in equal proportions to obtain the additional data.

[0027] In a second aspect, a computer device includes a memory and a processor, and the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0028] In a third aspect, a computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the above method.

[0029] The present application has at least the following beneficial effects:

[0030] When using basic acoustic models for speech recognition, certain words are often misrecognized. Preparing training data for these error-prone words can improve the accuracy of speech recognition on these words. This invention uses an audio segmentation tool to segment the recorded error-recognized words and sentences, obtaining an audio file named WrongPart. Words corresponding to WrongPart in the pre-prepared correct original text sentences are deleted, and their deletion positions are marked. A speech synthesis system is used to convert the remaining text after deletion into an audio file ResidueWav, which is then combined with WrongPart according to the original text positions to obtain a labeled training file for the acoustic model. This training file is used as a training sample for fine-tuning the acoustic model, thus efficiently correcting the error-prone parts of ASR.

[0031] This invention utilizes two existing speech recognition tools to recognize the audio data to be labeled, resulting in two versions of the recognition result dataset. A text edit distance algorithm is used to calculate the text edit distance between the two datasets. Recognition results with edit distances less than a preset threshold, along with their corresponding original audio, are retained as the main data and added to the pre-labeled set. For recognition results with edit distances greater than or equal to the preset threshold, a portion of the recognition results and their corresponding original audio are extracted and added to the pre-labeled set as supplementary data. This approach maximizes the amount of data with small text edit distance differences, reduces manual verification costs, and makes the construction of the training sample set more efficient. Simultaneously, data with large text edit distances are also extracted, ensuring the diversity of training data features. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating a method for rapidly annotating acoustic model training data in speech recognition, according to an embodiment of the present invention.

[0033] Figure 2 For the present invention Figure 1 The illustrated embodiment is a flowchart of obtaining the trained basic acoustic model.

[0034] Figure 3 This is a flowchart illustrating the first stage (data acquisition of the basic training set) of an embodiment of the present invention applied to the customer service field.

[0035] Figure 4 This is a flowchart illustrating the second stage (data acquisition of supplementary training sets for frequently misspelled words) of an embodiment of the present invention applied to the customer service field. Detailed Implementation

[0036] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0037] In one embodiment, as shown in Figure 1 A method for quickly labeling acoustic model training data in speech recognition is provided, comprising the following steps:

[0038] S1: using a trained basic acoustic model to perform speech recognition, wherein words and sentences with recognition errors will occur;

[0039] S2: using an audio cutting tool to cut the recorded words and sentences with recognition errors to obtain an audio file with recognition errors, denoted as WrongPart, wherein WrongPart contains one or more words recognized;

[0040] S3: deleting the words in the pre-prepared correct original text sentences corresponding to the WrongPart and marking the deletion position; the remaining text in the original text after deletion is denoted as ResidueScript;

[0041] S4: using a speech synthesis system to convert the ResidueScript into an audio file, denoted as ResidueWav, and combining ResidueWav with WrongPart according to the original text position, so that WrongPart fills the deletion position to obtain a labeled data of an acoustic model, denoted as TrainFile;

[0042] S5: multiple different TrainFiles constitute a training set TrainFiles for acoustic model fine-tuning.

[0043] For the trained basic acoustic model mentioned in step S1 above, it can use the currently open-source acoustic model or can be reconstructed, for example, as shown in Figure 2 comprising the following steps:

[0044] Obtain a batch of audio data to be labeled in a specified industry and divide them according to the type of speech subject identity;

[0045] For the audio data to be labeled with a determined speech subject identity type, call two existing speech recognition tools to perform recognition respectively, and accordingly obtain a first preliminary recognition result data set and a second preliminary recognition result data set;

[0046] The text edit distance algorithm is used to calculate the text edit distance between the first preliminary recognition result data set and the second preliminary recognition result data set, and the recognition result and the corresponding original audio with an edit distance less than a preset threshold are reserved as main data and added to the pre-labeled set. For the recognition result with an edit distance greater than and equal to the preset threshold, part of the recognition result and the corresponding original audio are extracted as additional data and also added to the pre-labeled set; the total audio duration of the main data is much greater than the total audio duration of the source data of the additional data (selected by the edit distance, and the above method is used to construct the final text answer on this basis);

[0047] The pre-labeled set is verified and used as a basic training set for training an acoustic model.

[0048] Exemplarily, the embodiment of the application is specifically applied to the customer service field, and the training data labeling of the acoustic model in speech recognition mainly has two-stage requirements: the first stage is large-batch labeled audio data, generally not less than 1000 hours of final labeled audio files are needed to obtain the training effect of the basic acoustic model; the second stage is audio labeling of specific error types, that is, when the speech recognition model is used for recognition after the basic acoustic model, some words are always recognized incorrectly, and training data of such error-prone words needs to be prepared again to improve the recognition accuracy of the speech recognition on such error-prone words. The following is specifically described.

[0049] I. As shown in Figure 3 , the first stage of large-batch training data labeling

[0050] 1. Obtain the audio data to be labeled of an industry, with a duration of not less than 3000 hours;

[0051] Taking the education industry as an example, the audio data of the intelligent education industry is obtained, and 20,000 hours of original audio are selected.

[0052] 2. The obtained audio data is channel-separated. The audio data in the customer service field is generally right channel for customer service and left channel for customer audio. The channel-separated audio is cut by using the voice activity detection technology (Voice Activity Detection, VAD), the audio obtained after VAD cutting is called Tencent speech recognition interface and Ali speech recognition ASR interface, and AliScript and tencentScript are obtained after being recognized respectively;

[0053] For customer audio and customer audio, respectively, the same processing is described below. The difference is that when finally selecting the training data set, for example, 1:4 of the customer and customer audio data can be retained, that is, in a 1000-hour training set, 200 hours of customer audio and 800 hours of customer audio. Mainly considering that in the customer service field, the audio pronunciation of the customer service role is relatively standard Mandarin, while the audio pronunciation of the customer role often has large differences and is more likely to be misrecognized, so the proportion in the training data is higher.

[0054] Taking customer audio as an example, the audio separated by the VAD cutting channel is used to obtain a cutting audio set. The cutting audio set calls the Tencent speech recognition interface and the Ali speech recognition interface respectively, and each cutting audio obtains two corresponding recognition texts AliScript and tencentScript after interface recognition.

[0055] 3. The Leveinshtein distance (i.e. text edit distance) between AliScript and tencentScript is calculated using a text edit distance algorithm. The recognition result and the corresponding original audio with an edit distance less than a preset threshold are retained and added to the pre-labeled set. At the same time, in the part with a text edit distance greater than the preset threshold, it is divided into high, medium and low edit sentence parts, and part of the data is randomly extracted from the three parts in an equal ratio and added to the pre-labeled set. For example, the edit distance threshold can be set to 5. When the edit distance is greater than 5 and less than 10, it is considered as the low part of the edit distance; when the edit distance is between 10 and 20, it is considered as the medium part of the edit distance, and when the edit distance is greater than 20, it is considered as the high part of the edit distance. The total duration of the high, medium and low three parts of the data with an edit distance greater than the preset threshold is not less than 20% (not more than 40%) of the total duration of the audio with an edit distance less than the preset threshold. For the extraction of the audio with different AliScript and tencentScript recognition, the recognition texts of Ali and Tencent can be aligned, for example:

[0056] ali: a b c

[0057] tx: y c

[0058] Similar to beam search; 1) (a) and (y) can be selected first; 2) ab, a_, yb, and y_ can be selected, and then the smallest ppl is selected.

[0059] The advantage of this is that when the difference between the text edit distances of the Ali and Tencent ASR recognition data is small, the cost of manual verification is low; and when the text edit distance is large, part of it is also extracted to ensure the diversity of the training data features.

[0060] Such audio (pre-labeled set) can be quickly labeled by manual verification according to the reference text in subsequent labeling. At the same time, a small amount of data, such as 1 / 10, is selected from the audio with a large text edit distance to the pre-labeled set. Such audio needs to spend a certain amount of time for labeling in the subsequent manual verification process. Selecting a small amount of such audio can make the final audio training set have more rich features.

[0061] 4. The total audio duration in the pre-labeled set is counted. According to the 1000-hour requirement of the training set duration, select no less than 1000 audio to form the verification set, and the final training set of no less than 1000 hours can be obtained by manual verification of the verification set. The interface recognition error is corrected, and the corrected one is the training data set of the acoustic model through manual verification.

[0062] In particular, the selection of the verification set is to divide the pre-labeled set into different business categories using clustering. In each business category, a certain amount of audio and corresponding recognition results of the business category are selected according to the original data proportion and added to the verification set, so as to achieve the balance of the training data in different business categories.

[0063] The text corresponding to the audio in the pre-labeled set is divided into multiple different business categories using clustering and classification. The clustering and classification are both unsupervised methods, and the text data is divided into multiple categories. The classification contains specific business words and can be used as a business category, so that the pre-labeled set is divided into multiple business categories. In different business categories, a part is randomly selected according to the proportion of the audio duration in the business category and added to the verification set. All audio data in the verification set need to be checked one by one. The final verification result forms the training data set of the acoustic model in speech recognition, and the basic acoustic model can be obtained after training and can be used for speech recognition.

[0064] Through the first stage of training data preparation, the basic training set for training the acoustic model can be obtained, and the basic acoustic model can be trained. By combining the language model, pronunciation dictionary and decoder, speech recognition can be performed. The training data labeled by this method covers a wider range of businesses, and the training data reduces the labor cost. In the process of using the speech recognition model to recognize sentences, when the recognition encounters frequently wrong audio, the acoustic model training corpus needs to be expanded, which can be expanded by the following small amount of easy-to-mistake word training data labeling method.

[0065] II. As shown in Figure 4 , the second stage of small amount of easy-to-mistake word training data labeling

[0066] In the process of identifying a sentence using a speech recognition model, when an error-prone word is encountered in the identification, it is determined that the acoustic model lacks training data, and the acoustic model training corpus needs to be expanded accordingly. The selection of the expanded corpus is often limited, which affects the rapid iteration of the acoustic model. The embodiment of the present application can quickly obtain a certain amount of corpus to expand the training data by the following steps, and the acoustic model can be fine-tuned on this basis, thereby rapidly iterating and improving the effect.

[0067] 1. A business classification module pre-classifies businesses according to specific industries, establishes common business types, and obtains corresponding sentence sets under each business type for backup. For example, the corpus of the home appliance industry can be classified according to specific goods, such as color TV, refrigerator, mobile phone, etc. Sentences containing "color TV" or similar words form a business category and form a sentence set. The sentences that may be contained in this business category are as follows: "You TCL's clarity is not high enough, and it is not clear enough", "The screen of the color TV you bought is a little bit bloomed, how to contact the after-sales service", "The remote control of the TV set is lost, and it needs to be reconfigured", "Can the TV be exchanged for a new one? This one has been used for many years, and I want to buy a new one".

[0068] The text data of a specific industry such as the education industry is classified, and the pre-classified business categories of the education industry are obtained, as shown in the following table. Figure 4 As shown in the table, the "children's programming course" is contained in the programming course business category, and the "3-7 year-old English training course" is contained in the English training course business category. There are a large number of common dialogue texts under each business category.

[0069] 2. The text sentences of different business categories are cut and recombined, and the expanded text sentences of each business category are obtained. The text sentences form a text sentence set under the business type. For example, the text sentences under the color TV business category are as follows: "You TCL's clarity is not high enough, and it is not clear enough", "The screen of the color TV you bought is a little bit bloomed, how to contact the after-sales service", "The remote control of the TV set is lost, and it needs to be reconfigured", "Can the TV be exchanged for a new one? This one has been used for many years, and I want to buy a new one". Since the four words "TCL", "color TV", "TV set", and "TV" express the same business, the four words can be replaced with each other, and the original four sentences can be expanded to 16 sentences. From "You TCL's clarity is not high enough, and it is not clear enough", "You color TV's clarity is not high enough, and it is not clear enough", "You TV set's clarity is not high enough, and it is not clear enough", and "You TV's clarity is not high enough, and it is not clear enough", it is very likely that the same word will be used to replace the original sentence, thereby obtaining more expanded sentences.

[0070] As shown in the following table, Figure 4As shown, for a large number of text sentences under each business category, more combined sentences can be obtained by cutting and recombining. The sentences expanded in this way still belong to the business category and together with the original dialogue text sentences form the common business sentence set BusiSents.

[0071] 3. For sentences and words that are misrecognized in speech recognition, use an audio segmentation tool to segment them; the resulting audio file containing the misrecognized words is called the WrongPart audio file.

[0072] 4. For incorrectly identified audio words, determine their corresponding business category based on pre-defined business classifications. From this category, obtain the original text sentence for that business category. Delete words from the original text sentence that correspond to the incorrectly identified audio word and mark the deletion position as P. For example... Figure 4 As shown, a large set of business sentences, BusiSents, is obtained from this business category. Words that are the same as or similar to the incorrect audio in each sentence of the BusiSents set are deleted, and the deletion position is marked as P. If there are multiple deletion positions, they are marked as P1, P2, P3, etc. The remaining text in the original text after deletion is ResidueScript. For the remaining ResidueScript, the TTS speech synthesis system tacotron2_lpcnet (an open-source speech synthesis system in the industry) is used to synthesize ResidueWav. The synthesized audio file ResidueWav is then combined with the incorrect audio file WrongPart according to the original text positions. That is, the audio at the incorrect position occupies position P of the deleted word in the original sentence. If multiple incorrect words are deleted, they occupy positions P1, P2, P3, etc. The combination yields a piece of acoustic model annotation data, TrainFile, which is the audio file used to train the acoustic model and corresponds to the text annotation.

[0073] like Figure 4 As shown, the WrongPart audio file contains words such as "oral class" and "children". Therefore, the word "oral class" is deleted from the original text sentence "English oral class supplementary materials free" and the deletion position P is marked. The text sentence becomes "English P class supplementary materials free". The audio ResidueWav of this sentence is synthesized and combined with the audio of "oral class" according to the original text position to obtain the complete audio of "English oral class supplementary materials free" which contains the easily mispronounced "oral class".

[0074] 5. A large number of TrainFiles are obtained by the same operation to form a training set TrainFiles for fine-tuning of the acoustic model.

[0075] The synthesized audio contains the parts that are prone to errors in ASR, and the training data for fine-tuning of the model is used to correct the parts prone to errors in ASR through model fine-tuning, i.e., Fine-Tune.

[0076] It should be understood that, although each step in the above flowchart is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in the order indicated by the arrow. Moreover, at least part of the steps can include a plurality of specific steps, which can be executed at the same time or at different times, and the execution order of these specific steps is not necessarily sequential, but can also be executed in rotation or alternation with at least part of the steps or the specific steps in other steps.

[0077] In an embodiment, a computer device is provided, which can be a terminal. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The memory stores a computer program, which relates to all or part of the processes in the method of the above embodiment.

[0078] In an embodiment, a computer readable storage medium is provided, which stores a computer program, which relates to all or part of the processes in the method of the above embodiment.

[0079] The technical features of the above embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not contradict, they should be considered as within the scope of the present disclosure.

[0080] The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that, for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for fast labeling acoustic model training data in speech recognition, characterized in that, The method comprises the following steps: Carrying out speech recognition using a trained basic acoustic model, wherein words and sentences with recognition errors will occur; Using an audio cutting tool to cut the recorded words and sentences with recognition errors to obtain an audio file with recognition errors, denoted as WrongPart, wherein WrongPart contains one or more words with recognition errors; Deleting the words corresponding to WrongPart in a pre-prepared correct original text sentence and marking the deletion position; the remaining text in the original text after deletion is denoted as ResidueScript; Converting ResidueScript into an audio file using a speech synthesis system, denoted as ResidueWav, and combining ResidueWav and WrongPart according to the original text position to fill WrongPart in the deletion position, thereby obtaining an acoustic model annotation data, denoted as TrainFile; A plurality of different TrainFiles form a training set TrainFiles for acoustic model fine-tuning; The trained basic acoustic model is obtained in the following manner: Obtaining a batch of audio data to be annotated in a set industry and dividing them according to the identity type of the speech subject; For the audio data to be annotated with a determined identity type of the speech subject, calling two existing speech recognition tools to perform recognition respectively, thereby obtaining a first preliminary recognition result data set and a second preliminary recognition result data set; Using a text edit distance algorithm to calculate the text edit distance between the first preliminary recognition result data set and the second preliminary recognition result data set, retaining the recognition results with an edit distance less than a preset threshold and the corresponding original audio as main data to be added to a pre-annotation set, and extracting part of the recognition results with an edit distance greater than and equal to the preset threshold and the corresponding original audio as additional data to also be added to the pre-annotation set; the total audio duration of the main data is much longer than the total audio duration of the source data of the additional data; The pre-annotation set is verified and used as a basic training set for training an acoustic model; the basic acoustic model is trained using the basic training set. 2.The method for quickly labeling acoustic model training data in speech recognition according to claim 1, characterized in that, The pre-prepared correct original text sentence comes from a text sentence set prepared in the following manner: According to the business of a set industry, pre-classifying the business to establish common business types, and obtaining corresponding text sentence sets under each business type for standby use; For text sentences of different business types, cutting and recombining are performed, and expanded text sentences are obtained for each business type to form a text sentence set under the business type. 3.The method for quickly labeling acoustic model training data in speech recognition according to claim 1, characterized in that, The set industry is an industry related to customer service, and the identity type of the speech subject is divided into customer service and customer; accordingly, the division according to the identity type of the speech subject is specifically channel separation of the audio data according to left and right channels. 4.The method for quickly labeling acoustic model training data in speech recognition according to claim 1, characterized in that, Before calling the two existing speech recognition tools to perform recognition respectively, VAD is used to cut the audio to obtain individual sentence audio.

5. The method for fast labeling acoustic model training data in speech recognition according to claim 1, wherein, The two existing speech recognition tools use Tencent speech recognition ASR interface and Ali speech recognition ASR interface.

6. The method for fast labeling acoustic model training data in speech recognition according to claim 1, wherein, The total time length of the audio with an edit distance greater than and equal to the preset threshold value is 20% to 40% of the total time length of the audio with an edit distance less than the preset threshold value.

7. The method for fast annotation of acoustic model training data in speech recognition according to claim 6, characterized in that, The additional data is determined in the following manner: The data with an edit distance greater than and equal to the preset threshold value is further subdivided into high, medium and low parts, and part of the data is randomly extracted in equal proportions from the three parts to obtain the additional data.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor, when executing the computer program, implements the steps of the method of any one of claims 1 to 7.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by the processor, implements the steps of the method of any one of claims 1 to 7.

Images