Method, device, electronic device and storage medium for screening model training data

Abstract

The embodiment of the present invention provides a method, device, electronic device and storage medium for screening model training data. The method includes: obtaining a first training audio and a second training audio, aligning the first training audio and the second training audio, obtaining a first target audio and a second target audio at the same amplitude, respectively obtaining a first sliding window similarity of the first target audio, a second sliding window similarity of the second target audio, and a zero crossing rate ratio and an average energy ratio between the first target audio and the second target audio; if the first sliding window similarity, the second sliding window similarity, the zero crossing rate ratio and the average energy ratio all meet preset conditions, the first target audio and the second target audio are used as target model training data, thereby avoiding manual screening of audios one by one, and automatically screening out model training data with high similarity, thereby improving screening efficiency and screening quality, and thus improving the recognition ability of a speech recognition model.

Description

Technical Field

[0001] The present invention relates to the field of speech processing technology, and in particular to a method for screening model training data, a device for screening model training data, an electronic device and a computer-readable storage medium. Background Art

[0002] Audio refers to sound waves that the human ear can hear with an amplitude between 20Hz and 20kHz. It can represent files used to store sound content or filter vibrations, etc. In the digital field, audio signals are discrete signals composed of digital symbols (0 and 1 in binary), rather than using physical means to save signals, such as recording with ordinary magnetic tapes.

[0003] In a speech training scenario, if the speech recognition model is directly trained using audio with a different sampling rate corresponding to the speech recognition model, the training effect of the speech recognition model will be poor, and the number of audios used to make the speech recognition model is limited. The more audios that match the sampling rate corresponding to the speech recognition model are used to train the speech recognition model, the better the effect and higher the accuracy of the speech recognition model.

[0004] Therefore, in order to obtain more audio that matches the speech recognition model, there are many methods for converting audio with different sampling rates in the prior art. One is to directly digitally convert audio with different sampling rates, for example, directly increase or decrease the sampling points on the audio. This method of directly increasing or decreasing the sampling points is likely to cause audio distortion or errors in the audio data. Another method is to use a real channel for simulation. During the simulation process, the original audio is first played out, and then the played audio is resampled. This method of resampling the original audio after simulation is greatly affected by environmental factors such as environmental noise and channel disturbances. The audio is prone to frame loss or reduced signal-to-noise ratio. However, in the actual audio screening process, users are often required to manually screen the audio, and use the multiple similar audios after screening as model training data for the speech recognition model. However, when faced with a large amount of audio, the manual screening method not only wastes a lot of human resources and has low efficiency in screening audio, but also has problems such as low similarity of the screened audio. Summary of the invention

[0005] The embodiments of the present invention provide a method, device, electronic device and computer-readable storage medium for screening model training data to solve the problems of poor screened audio quality, low similarity and low screening efficiency in the process of screening model training data.

[0006] The embodiment of the present invention discloses a method for screening model training data, the method comprising:

[0007] Acquire a first training audio and a second training audio from each initial audio;

[0008] Acquire a sequence of mutual correlation coefficients between the first training audio and the second training audio, and use the mutual correlation coefficient with the largest value as an audio alignment point between the first training audio and the second training audio;

[0009] Aligning the first training audio with the second training audio according to the audio alignment point to obtain a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio;

[0010] Obtaining a first sliding window similarity variance of the first target audio, a second sliding window similarity variance of the second target audio, and a zero crossing rate ratio and an average energy ratio between the first target audio and the second target audio;

[0011] If the first sliding window similarity variance and the second sliding window similarity variance are less than or equal to the preset sliding window similarity variance, and the zero-crossing rate ratio and the average energy ratio are within the preset audio parameter range, the first target audio and the second target audio are used as target model training data for training a speech recognition model.

[0012] Optionally, obtaining the first training audio and the second training audio from each initial audio includes:

[0013] Two initial audios are selected from the initial model training data for training, and amplitude normalization processing is performed on the frame sequences of the two initial audios to obtain the first training audio and the second training audio whose audio amplitudes are in the same amplitude range.

[0014] Optionally, the two initial audios include a first initial audio and a second initial audio, and obtaining the first training audio and the second training audio whose audio amplitudes are in the same amplitude range includes:

[0015] Performing frame processing on the first initial audio and the second initial audio respectively to obtain a first frame sequence of the first initial audio and a second frame sequence of the second initial audio;

[0016] Acquire a first sampling rate of the first initial audio and a second sampling rate of the second initial audio;

[0017] If the first sampling rate is less than the second sampling rate, inserting at least one zero-value point at equal intervals in the first frame sequence to generate the first training audio whose audio amplitudes are in the same amplitude range;

[0018] If the second sampling rate is less than the first sampling rate, at least one zero-value point is inserted at equal intervals in the second frame sequence to generate the second training audio whose audio amplitude is in the same amplitude range.

[0019] Optionally, the aligning the first training audio and the second training audio according to the audio alignment point to obtain a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio includes:

[0020] Performing an alignment process on the first training audio and the second training audio according to the audio alignment point, and determining an alignment area and a non-alignment area corresponding to the alignment process;

[0021] In response to the deletion process for the non-aligned area, a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio are obtained.

[0022] Optionally, the first target audio and the second target audio have the same window function corresponding to the same amplitude range, and the obtaining of a first sliding window similarity variance of the first training audio and a second sliding window similarity variance of the second training audio includes:

[0023] Performing windowing processing on the first target audio and the second target audio by using the window function to generate first sliding windows corresponding to the first target audio and second sliding windows corresponding to the second target audio;

[0024] Acquire a first autocorrelation coefficient corresponding to each of the first sliding windows, and use the mutual correlation coefficient and the first autocorrelation coefficient to determine each first sliding window similarity corresponding to each of the first sliding windows and a first sliding window similarity variance;

[0025] The second autocorrelation coefficient corresponding to each second sliding window is obtained, and the mutual correlation coefficient and the second autocorrelation coefficient are used to determine each second sliding window similarity corresponding to each second sliding window and a second sliding window similarity variance.

[0026] Optionally, the obtaining a zero-crossing rate ratio and an average energy ratio between the first target audio and the second target audio includes:

[0027] Acquire a first zero-crossing rate and a first average energy corresponding to the first target audio, and a second zero-crossing rate and a second average energy corresponding to the second target audio;

[0028] Calculate a zero-crossing rate ratio between the first target audio and the second target audio using the first zero-crossing rate and the second zero-crossing rate;

[0029] The first average energy and the second average energy are used to calculate and determine an average energy ratio between the first target audio and the second target audio.

[0030] Optionally, the preset audio parameter range includes a preset zero crossing rate ratio range and a preset average energy ratio range, and if the first sliding window similarity variance and the second sliding window similarity variance are less than or equal to the preset sliding window similarity variance, and the zero crossing rate ratio and the average energy ratio are within the preset audio parameter range, then using the first target audio and the second target audio as target model training data for training a speech recognition model, comprising:

[0031] If the first sliding window similarity variance of the first target audio and the second sliding window similarity variance of the second target audio are less than or equal to a preset sliding window similarity variance, and three consecutive first sliding window similarities and second sliding window similarities are greater than or equal to a preset sliding window similarity;

[0032] And, if the zero-crossing rate ratio is within a preset zero-crossing rate ratio range, and the average energy ratio is within a preset average energy ratio range, the first target audio and the second target audio are used as target model training data for training a speech recognition model.

[0033] The embodiment of the present invention discloses a device for screening model training data, the device comprising:

[0034] A training audio acquisition module, used to acquire a first training audio and a second training audio from each initial audio;

[0035] an audio alignment point determination module, configured to obtain a sequence of mutual correlation coefficients between the first training audio and the second training audio, and use the mutual correlation coefficient with the largest value as an audio alignment point between the first training audio and the second training audio;

[0036] a target audio acquisition module, configured to align the first training audio and the second training audio according to the audio alignment point, to obtain a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio;

[0037] An audio parameter acquisition module, used to acquire a first sliding window similarity variance of the first target audio, a second sliding window similarity variance of the second target audio, and a zero crossing rate ratio and an average energy ratio between the first training audio and the second training audio;

[0038] A model training data screening module is used to use the first target audio and the second target audio as target model training data for training a speech recognition model if the first sliding window similarity variance and the second sliding window similarity variance are less than or equal to a preset sliding window similarity variance, and the zero crossing rate ratio and the average energy ratio are within a preset audio parameter range.

[0039] Optionally, the training audio acquisition module includes:

[0040] The training audio acquisition submodule is used to select two initial audios from the initial model training data used for training, and perform amplitude normalization processing on the frame sequences of the two initial audios to obtain the first training audio and the second training audio whose audio amplitudes are in the same amplitude range.

[0041] Optionally, the two initial audios include a first initial audio and a second initial audio, and the training audio acquisition submodule is specifically used to:

[0042] Performing frame processing on the first initial audio and the second initial audio respectively to obtain a first frame sequence of the first initial audio and a second frame sequence of the second initial audio;

[0043] Acquire a first sampling rate of the first initial audio and a second sampling rate of the second initial audio;

[0044] If the first sampling rate is less than the second sampling rate, inserting at least one zero-value point at equal intervals in the first frame sequence to generate the first training audio whose audio amplitudes are in the same amplitude range;

[0045] If the second sampling rate is less than the first sampling rate, at least one zero-value point is inserted at equal intervals in the second frame sequence to generate the second training audio whose audio amplitude is in the same amplitude range.

[0046] Optionally, the target audio acquisition module is specifically used for:

[0047] Performing an alignment process on the first training audio and the second training audio according to the audio alignment point, and determining an alignment area and a non-alignment area corresponding to the alignment process;

[0048] In response to the deletion process for the non-aligned area, a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio are obtained.

[0049] Optionally, the first target audio and the second target audio have the same window function corresponding to them within the same amplitude range, and the audio parameter acquisition module is specifically used to:

[0050] Performing windowing processing on the first target audio and the second target audio by using the window function to generate first sliding windows corresponding to the first target audio and second sliding windows corresponding to the second target audio;

[0051] Acquire a first autocorrelation coefficient corresponding to each of the first sliding windows, and use the mutual correlation coefficient and the first autocorrelation coefficient to determine each first sliding window similarity corresponding to each of the first sliding windows and a first sliding window similarity variance;

[0052] The second autocorrelation coefficient corresponding to each second sliding window is obtained, and the mutual correlation coefficient and the second autocorrelation coefficient are used to determine each second sliding window similarity corresponding to each second sliding window and a second sliding window similarity variance.

[0053] Optionally, the audio parameter acquisition module is specifically used to:

[0054] Acquire a first zero-crossing rate and a first average energy corresponding to the first target audio, and a second zero-crossing rate and a second average energy corresponding to the second target audio;

[0055] Calculate a zero-crossing rate ratio between the first target audio and the second target audio using the first zero-crossing rate and the second zero-crossing rate;

[0056] The first average energy and the second average energy are used to calculate and determine an average energy ratio between the first target audio and the second target audio.

[0057] Optionally, the preset audio parameter range includes a preset zero-crossing rate ratio range and a preset average energy ratio range, and the model training data screening module is specifically used for:

[0058] If the first sliding window similarity variance of the first target audio and the second sliding window similarity variance of the second target audio are less than or equal to a preset sliding window similarity variance, and three consecutive first sliding window similarities and second sliding window similarities are greater than or equal to a preset sliding window similarity;

[0059] And, if the zero-crossing rate ratio is within a preset zero-crossing rate ratio range, and the average energy ratio is within a preset average energy ratio range, the first target audio and the second target audio are used as target model training data for training a speech recognition model.

[0060] The embodiment of the present invention further discloses an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0061] The memory is used to store computer programs;

[0062] The processor is used to implement the method described in the embodiment of the present invention when executing the program stored in the memory.

[0063] The embodiment of the present invention further discloses a computer-readable storage medium having instructions stored thereon, which, when executed by one or more processors, enables the processors to execute the method described in the embodiment of the present invention.

[0064] The embodiments of the present invention include the following advantages:

[0065] In an embodiment of the present invention, a first training audio and a second training audio can be obtained from each initial audio, and then a sequence of mutual correlation coefficients between the first training audio and the second training audio can be obtained, the mutual correlation coefficient with the largest value can be used as an audio alignment point between the first training audio and the second training audio, and the first training audio and the second training audio can be aligned according to the audio alignment point to obtain a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio, and then a first sliding window similarity variance of the first target audio, a second sliding window similarity variance of the second target audio, and a zero crossing rate ratio and an average energy ratio between the first target audio and the second target audio can be obtained. If the first sliding window similarity variance and the second sliding window similarity variance are less than or equal to a preset sliding window similarity variance, and the zero crossing rate ratio and the average energy ratio are within a preset audio parameter range, the first target audio is aligned. The first target audio and the second target audio are used as model training data for training a speech recognition model, so that the first target audio and the second target audio at the same amplitude are obtained by aligning the training audios, and then the first sliding window similarity of the first target audio, the second sliding window similarity of the second target audio, and the zero crossing rate ratio and the average energy ratio between the first target audio and the second target audio are respectively obtained. If the first sliding window similarity, the second sliding window similarity, the zero crossing rate ratio and the average energy ratio all meet the preset conditions, it is characterized that the similarity between the first target audio and the second target audio is high. Therefore, the process of manually screening the audios one by one can be avoided. By automatically screening out model training data with high similarity, the screening efficiency and screening quality are improved, and the first target audio and the second target audio with high similarity are used as model training data for training a speech recognition model to improve the accuracy of the speech recognition model. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is a flowchart of the steps of a method for screening model training data provided in an embodiment of the present invention;

[0067] Figure 2 is a structural block diagram of a device for screening model training data provided in an embodiment of the present invention;

[0068] Figure 3 It is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0069] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0070] Speech signal processing can include speech recognition, speech synthesis, speech coding, etc. In speech recognition technology, speech endpoint detection (EPD) is an important step in processing speech signals. It is generally used to determine the endpoint position of the beginning of the audio signal and the endpoint position of the end of the speech signal, so as to collect the voice data that is actually used, reduce the amount of data and calculations processed by the speech recognition system, and improve the processing efficiency of speech recognition.

[0071] One of the core invention points of the embodiment of the present invention is to obtain the first training audio and the second training audio from each initial audio, and then obtain the sequence of mutual correlation coefficients between the first training audio and the second training audio, use the mutual correlation coefficient with the largest value as the audio alignment point of the first training audio and the second training audio, and align the first training audio and the second training audio according to the audio alignment point to obtain the first target audio corresponding to the first training audio and the second target audio corresponding to the second training audio, and then obtain the first sliding window similarity variance of the first target audio, the second sliding window similarity variance of the second target audio, and the zero crossing rate ratio and the average energy ratio between the first target audio and the second target audio. If the first sliding window similarity variance and the second sliding window similarity variance are less than or equal to the preset sliding window similarity variance, and the zero crossing rate ratio and the average energy ratio are within the preset audio parameter range, the first target audio is aligned. The first target audio and the second target audio are used as model training data for training the speech recognition model, so that the first target audio and the second target audio at the same amplitude are obtained by aligning the training audio, and then the first sliding window similarity of the first target audio, the second sliding window similarity of the second target audio, and the zero crossing rate ratio and the average energy ratio between the first target audio and the second target audio are respectively obtained. If the first sliding window similarity, the second sliding window similarity, the zero crossing rate ratio and the average energy ratio all meet the preset conditions, it is characterized that the similarity between the first target audio and the second target audio is high. Therefore, the process of manually screening the audios one by one can be avoided. By automatically screening the first target audio and the second target audio with high similarity, the screening efficiency and screening quality are improved, and the first target audio and the second target audio with high similarity are used as model training data for training the speech recognition model to improve the accuracy of the speech recognition model.

[0072] It should be noted that various data are involved in the embodiments of the present invention, and each technical term is explained below:

[0073] The initial model training data may be a number of unscreened model training data used for training, which may include various types of data such as audio, video, picture and text, such as audio A, audio B, audio C, audio D, picture E, picture F, video G, text H and other data. The user may randomly select or the speech recognition system may automatically select the first initial audio and the second initial audio of the initial model training data. For example, in the initial model training data such as audio A, audio B, audio C, audio D, picture E, picture F, video G, text H, the user or the speech recognition system may select the first initial audio as audio A and the second initial audio as audio B.

[0074] A frame sequence may be a sequence corresponding to each frame in the initial audio after the initial audio is subjected to amplitude normalization and frame processing, and may include a first frame sequence and a second frame sequence, wherein the first frame sequence may be a sequence corresponding to each frame in the first initial audio after the first initial audio is subjected to amplitude normalization and frame processing, and the second frame sequence may be a sequence corresponding to each frame in the second initial audio after the second initial audio is subjected to amplitude normalization and frame processing.

[0075] The training audio may be audio data with the same amplitude and sampling rate generated after amplitude normalization and frame division processing are performed on the selected initial audio. The training audio may include a first training audio corresponding to the first initial audio and a second training audio corresponding to the second initial audio. The amplitudes of the first training audio and the second training audio are in the same amplitude range and have the same sampling rate.

[0076] The target audio may be target model training data for training a speech recognition model generated after aligning the training audio, and may include a first target audio and a second target audio, such as audio A and audio B.

[0077] The target model training data may be audio data that meets preset conditions, has been screened, and is used to train a speech recognition model. It may include a first target audio and a second target audio. For example, when the first sliding window similarity of the first target audio and the second sliding window similarity of the second target audio are less than or equal to the preset sliding window audio similarity, and the zero crossing rate ratio and the average energy ratio between the first target audio and the second target audio are within a preset audio parameter range, the first target audio and the second target audio may be used as target model training data.

[0078] Reference Figure 1, shows a flowchart of a method for screening model training data provided in an embodiment of the present invention, which may specifically include the following steps:

[0079] Step 101, obtaining a first training audio and a second training audio from each initial audio;

[0080] In an embodiment of the present invention, two initial audios are first selected from the initial model training data used for training as the first initial audio and the second initial audio, and then the first initial audio and the second initial audio are respectively subjected to amplitude normalization processing to obtain the first training audio and the second training audio whose audio amplitudes are in the same amplitude range, thereby determining the audio amplitudes of different audios to be in the same amplitude range, so as to facilitate subsequent data processing of different audios.

[0081] Optionally, the amplitude normalization processing can be to reduce or enlarge the amplitude of the first training audio and the second training audio to the amplitude range of [-1, 1] respectively. For example, the initial audio includes audio A, audio B, audio C and audio D. When the user selects the first training audio as audio A and the second training audio as audio B, the amplitude of audio A and audio B are reduced or enlarged to the amplitude range of [-1, 1] respectively.

[0082] In a specific implementation, the first initial audio and the second initial audio can be framed separately to obtain a first frame sequence of the first initial audio and a second frame sequence of the second initial audio, and then the first sampling rate of the first initial audio and the second sampling rate of the second initial audio are obtained. If the first sampling rate is less than the second sampling rate, at least one zero-value point is inserted at equal intervals in the first frame sequence to generate a first training audio that is in the same amplitude range and has the same sampling rate as the second training audio. If the second sampling rate is less than the first sampling rate, at least one zero-value point is inserted at equal intervals in the second frame sequence to generate a second training audio that is in the same amplitude range and has the same sampling rate as the first training audio.

[0083] Specifically, the audio data collected by the voice acquisition device is often in segments. Therefore, in order to process the audio data in batches, the audio data can be segmented according to a specified length, such as a time period or a sampling number, during the frame processing process. Each segment of audio can be called a frame. The sampling rate (sampling frequency) can be the number of times the sound is sampled per unit time, for example, 16khz, 8khz, 1khz. For example, 1khz represents the decomposition of the audio signal into 10,000 parts per second. The higher the sampling frequency and the shorter the sampling interval, the more sound sample data the computer obtains per unit time, and the more accurate the sound waveform. According to the sampling theorem, when the sampling frequency is higher than twice the highest frequency of the sound signal, aliasing can be avoided and the discrete audio signal can be converted into a continuous signal. Interpolation and zero padding in a frame sequence with a low sampling rate is performed by increasing the number of sampling points of the target audio with a low sampling rate so that the sampling rates of different audio data remain consistent. Interpolation and zero padding can increase the spectral sampling points of the audio and reduce the spectral sampling interval, but will not change the spectral profile and frequency resolution of each audio. It can weaken the fence effect and improve the spectral resolution.

[0084] In one example, after amplitude normalization and frame division processing are performed on the first initial audio A and the second initial audio B, a first frame sequence of the first initial audio A and a second frame sequence of the second initial audio B can be obtained, and the first sampling rate a of the first initial audio and the second sampling rate b of the second initial audio can be obtained. If the first sampling rate a is 8 khz and the second sampling rate b is 16 khz, then the first sampling rate a is less than the second sampling rate b, indicating that the first sampling rate of the first initial audio A is low, and it is necessary to insert zero value points at equal intervals in the first frame sequence of the first initial audio, and to increase the number of sampling points of the spectrum to ensure the first initial audio A and the second initial audio B. The sampling rates are both 16khz. Similarly, if the first sampling rate a is 16khz and the second sampling rate b is 8khz, the second sampling rate b is less than the first sampling rate a, indicating that the second sampling rate of the second initial audio A is lower, and it is necessary to insert zero-value points at equal intervals in the second frame sequence of the second initial audio B, and increase the number of sampling points of the spectrum to ensure that the sampling rates of the first initial audio and the second initial audio are both 16khz, so as to generate the first training audio and the second training audio with audio amplitudes in the same amplitude range of [-1, 1] and sampling rates of 16khz by performing amplitude normalization and frame processing on the first initial audio and the second initial audio.

[0085] Step 102, obtaining a sequence of mutual correlation coefficients between the first training audio and the second training audio, and taking the mutual correlation coefficient with the largest value as an audio alignment point between the first training audio and the second training audio;

[0086] In an embodiment of the present invention, because the lengths of two different training audios may be different, for example, during the simulation process, the first training audio may have an extra segment or a missing segment at the front, the correlation degree of the two different audios can be determined by calculating the maximum correlation coefficient of the two audios, wherein the maximum correlation coefficient is the position where the correlation degree between the first training audio and the second training audio is the largest. By calculating the mutual correlation sequence between the first training audio and the second training audio, the maximum correlation coefficient is selected from the mutual correlation coefficient sequence as the audio alignment point of the first training audio and the second training audio.

[0087] Optionally, a mutual correlation coefficient sequence can be obtained and a mutual correlation curve can be drawn based on the mutual correlation function between the first training audio and the second training audio, so as to intuitively obtain the mutual correlation coefficient with the largest value between the first training audio and the second training audio. The normalized mutual correlation coefficient is a floating point number between 0 and 1. The larger the mutual correlation coefficient, the greater the correlation between the first training audio and the second training audio. For example, there are multiple mutual correlation coefficients in the mutual correlation coefficient sequence, namely: 0.3, 0.4, 0.5, 0.6, etc., then the largest mutual correlation coefficient 0.6 can be used as the audio alignment point.

[0088] Step 103, aligning the first training audio and the second training audio according to the audio alignment point to obtain a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio;

[0089] In an embodiment of the present invention, the position point with the greatest correlation between the first training audio and the second training audio is used as the audio alignment point, and then the first training audio and the second training audio are aligned according to the audio alignment point, so that the aligned first training audio can be used as the first target audio, and the aligned second training audio can be used as the second target audio.

[0090] In a specific implementation, during the process of aligning the first training audio and the second training audio through audio alignment points, the alignment area and the non-alignment area between the first training audio and the second training audio are determined, and the non-alignment area is deleted to obtain the first target audio corresponding to the first training audio and the second target audio corresponding to the second training audio in the alignment area.

[0091] In one example, the first training audio A and the second training audio B are aligned according to the audio alignment point, and the alignment area between the first training audio A and the second training audio B is determined to be area ①, and the non-alignment area between the first training audio A and the second training audio B is area ②. Then, area ② can be deleted, and the first training audio in area ① is used as the first target audio, and the second training audio is used as the second target audio.

[0092] Step 104: Obtain a first sliding window similarity variance of the first target audio, a second sliding window similarity variance of the second target audio, and a zero crossing rate ratio and an average energy ratio between the first target audio and the second target audio;

[0093] In an embodiment of the present invention, in the process of screening model training data, it is necessary to obtain the first sliding window similarity variance of the first target audio, the second sliding window similarity variance of the second target audio, and the zero crossing rate ratio and average energy ratio between the first target audio and the second target audio respectively.

[0094] Optionally, the sliding window similarity variance may be the variance between the sliding window similarities corresponding to each frame generated after the audio data is subjected to framing and windowing processing, which may include a first sliding window similarity variance corresponding to the first target audio and a second sliding window similarity variance corresponding to the second target audio. The zero crossing rate (ZeroCrossingRate, referred to as ZCR) refers to the number of times the audio signal passes through the zero point (from positive to negative or from negative to positive) in each frame. The zero crossing rate of the audio data can be obtained by dividing the number of opposite positive and negative between two sampling points in the audio data by the total number of sampling points. The zero crossing rate ratio is the ratio between the first zero crossing rate of the first target audio and the second zero crossing rate of the second target audio. The average energy ratio may be the energy contained in a unit frame. Usually, after the audio data is divided into frames, the data in each frame will not be listed due to the large amount of data. Instead, the average energy is used to evaluate the data after the frame. The average energy of the audio data can be obtained by dividing the sum of the squares of each sampling value in the audio data by the total number of sampling points. The average energy ratio is the ratio of the first average energy of each frame in the first target audio to the second average energy of each frame in the second target audio.

[0095] In an optional example, the first target audio and the second target audio correspond to the same window function within the same amplitude range, and the first target audio and the second target audio are windowed by the window function to generate first sliding windows corresponding to the first target audio and second sliding windows corresponding to the second target audio, and then the first autocorrelation coefficient corresponding to each first sliding window is obtained, and the mutual correlation coefficient and the first autocorrelation coefficient are used to determine the similarities of each first sliding window corresponding to each first sliding window and the variance of the first sliding window similarity, and the second autocorrelation coefficient corresponding to each second sliding window is obtained, and the mutual correlation coefficient and the second autocorrelation coefficient are used to determine the similarities of each second sliding window corresponding to each second sliding window and the variance of the second sliding window similarity.

[0096] Specifically, for each frame signal of the first target audio and the second target audio, before performing Fourier transform, a windowing operation is performed by multiplying the frame signal with a window function, so that the windowed audio signal is more periodic when the left and right ends are connected. The framing processing of the audio signal usually adopts a sliding window method. The window can be a rectangular window, a Hamming window, etc. If a part of the audio data is directly intercepted, a rectangular window is obtained. If a window function with a high center weight and low weights on both sides is used, a Hamming window is obtained. The window length determines the amount of information in the original audio signal contained in each frame signal. When the sliding distance of the window is equal to the window length each time, there is no overlap in each frame information. When the window sliding distance is less than the window length, there is overlap between each frame information. After obtaining each first sliding window of the first target audio and each second sliding window of the second target audio, the first autocorrelation coefficient corresponding to each first sliding window and the second autocorrelation coefficient corresponding to each second sliding window can be calculated first, and then the square of the mutual correlation coefficient between the first target audio and the second target audio is divided by the product of the first autocorrelation coefficient to obtain each first sliding window similarity of each first sliding window, and the square of the mutual correlation coefficient between the first target audio and the second target audio is divided by the product of the second autocorrelation coefficient to obtain each second sliding window similarity of each second sliding window, and then the first sliding window similarity variance is calculated through the first sliding window similarities of each first sliding window, and the second sliding window similarity variance is calculated through the second sliding window similarities of each second sliding window.

[0097] In another optional example, the first zero-crossing rate and the first average energy corresponding to the first target audio, as well as the second zero-crossing rate and the second average energy corresponding to the second target audio are obtained, the first zero-crossing rate and the second zero-crossing rate are used to calculate the zero-crossing rate ratio between the first target audio and the second target audio (first zero-crossing rate: second zero-crossing rate), and the first average energy and the second average energy are used to calculate and determine the average energy ratio between the first target audio and the second target audio (first average energy: second average energy).

[0098] Step 105: If the first sliding window similarity variance and the second sliding window similarity variance are less than or equal to the preset sliding window similarity variance, and the zero-crossing rate ratio and the average energy ratio are within a preset audio parameter range, the first target audio and the second target audio are used as model training data for training a speech recognition model.

[0099] In an embodiment of the present invention, the preset sliding window similarity variance may be set according to the user's design requirements for audio similarity. For example, the preset sliding window similarity variance may be 0.5, and the smaller the preset sliding window similarity variance is, the stronger the stability between the two target audios is. The preset audio parameter range may be the ratio of audio parameters between the first target audio and the second target audio, which may include a preset zero crossing rate ratio and a preset average energy ratio. For example, the preset zero crossing rate ratio range between the first target audio and the second target audio is [0.8, 1.2], and the preset average energy ratio range between the first target audio and the second target audio is [0.8, 1.2].

[0100] In a specific implementation, if the first sliding window similarity variance of the first target audio and the second sliding window similarity variance of the second target audio are less than or equal to the preset sliding window similarity variance, and three consecutive first sliding window similarities and second sliding window similarities are greater than or equal to the preset sliding window similarities, and if the zero crossing rate ratio is within the preset zero crossing rate ratio range, and the average energy ratio is within the preset average energy ratio range, the first target audio and the second target audio are used as model training data for training the speech recognition model.

[0101] In one example, the preset sliding window similarity variance is 0.5, the preset zero crossing rate ratio range and the preset average energy ratio range are [0.8, 1.2], the first sliding window similarity variance of the first target audio and the second sliding window similarity variance of the second target audio are both 0.3, the preset sliding window similarity is 0.2, the zero crossing rate ratio between the first target audio and the second target audio is 1, and the preset average energy ratio range between the first target audio and the second target audio is 1. When the first sliding window similarity variance of the first target audio is 0.3 and the second target audio is The second sliding window similarity variance 0.3 is less than the preset sliding window similarity variance of 0.5, any three consecutive first sliding window similarities and second sliding window similarities are greater than the preset sliding window similarity 0.2, the zero crossing rate ratio 1 between the first target audio and the second target audio is in the range of [0.8, 1.2], and the average energy ratio 1 between the first target audio and the second target audio is in the range of [0.8, 1.2]. It is considered that the similarity between the first target audio and the second target audio is high, and can be used as target model training data for training a speech recognition model.

[0102] It should be noted that the embodiments of the present invention include but are not limited to the above examples. It is understandable that under the guidance of the ideas of the embodiments of the present invention, those skilled in the art can make settings according to actual conditions, and the present invention is not limited to this.

[0103] In an embodiment of the present invention, a first training audio and a second training audio can be obtained from each initial audio, and then a sequence of mutual correlation coefficients between the first training audio and the second training audio can be obtained, the mutual correlation coefficient with the largest value can be used as an audio alignment point of the first training audio and the second training audio, and the first training audio and the second training audio can be aligned according to the audio alignment point to obtain a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio, and then a first sliding window similarity variance of the first target audio, a second sliding window similarity variance of the second target audio, and a zero crossing rate ratio and an average energy ratio between the first target audio and the second target audio can be obtained. If the first sliding window similarity variance and the second sliding window similarity variance are less than or equal to a preset sliding window similarity variance, and the zero crossing rate ratio and the average energy ratio are within a preset audio parameter range, the first target audio and the second target audio are aligned. The target audio is used as model training data for training a speech recognition model, so that the first target audio and the second target audio at the same amplitude are obtained by aligning the training audio, and then the first sliding window similarity of the first target audio, the second sliding window similarity of the second target audio, and the zero crossing rate ratio and the average energy ratio between the first target audio and the second target audio are respectively obtained. If the first sliding window similarity, the second sliding window similarity, the zero crossing rate ratio and the average energy ratio all meet the preset conditions, it is characterized that the similarity between the first target audio and the second target audio is high. Therefore, the process of manually screening the audios one by one can be avoided. By automatically screening the first target audio and the second target audio with high similarity, the screening efficiency and screening quality are improved, and the first target audio and the second target audio with high similarity are used as model training data for training the speech recognition model to improve the accuracy of the speech recognition model.

[0104] It should be noted that, for the sake of simplicity, the method embodiments are described as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because according to the embodiments of the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0105] Reference Figure 2 , shows a structural block diagram of a device for screening model training data provided in an embodiment of the present invention, which may specifically include the following modules:

[0106] The training audio acquisition module 201 is used to acquire the first training audio and the second training audio from each initial audio;

[0107] An audio alignment point determination module 202 is used to obtain a sequence of mutual correlation coefficients between the first training audio and the second training audio, and use the mutual correlation coefficient with the largest value as an audio alignment point between the first training audio and the second training audio;

[0108] A target audio acquisition module 203 is used to align the first training audio and the second training audio according to the audio alignment point to obtain a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio;

[0109] An audio parameter acquisition module 204 is used to acquire a first sliding window similarity variance of the first target audio, a second sliding window similarity variance of the second target audio, and a zero crossing rate ratio and an average energy ratio between the first training audio and the second training audio;

[0110] The model training data screening module 205 is used to use the first target audio and the second target audio as target model training data for training a speech recognition model if the first sliding window similarity variance and the second sliding window similarity variance are less than or equal to a preset sliding window similarity variance, and the zero crossing rate ratio and the average energy ratio are within a preset audio parameter range.

[0111] In an optional embodiment, the training audio acquisition module 201 includes:

[0112] The training audio acquisition submodule is used to select two initial audios from the initial model training data used for training, and perform amplitude normalization processing on the frame sequences of the two initial audios to obtain the first training audio and the second training audio whose audio amplitudes are in the same amplitude range.

[0113] In an optional embodiment, the two initial audios include a first initial audio and a second initial audio, and the training audio acquisition submodule is specifically used to:

[0114] Performing frame processing on the first initial audio and the second initial audio respectively to obtain a first frame sequence of the first initial audio and a second frame sequence of the second initial audio;

[0115] Acquire a first sampling rate of the first initial audio and a second sampling rate of the second initial audio;

[0116] If the first sampling rate is less than the second sampling rate, inserting at least one zero-value point at equal intervals in the first frame sequence to generate the first training audio whose audio amplitudes are in the same amplitude range;

[0117] If the second sampling rate is less than the first sampling rate, at least one zero-value point is inserted at equal intervals in the second frame sequence to generate the second training audio whose audio amplitude is in the same amplitude range.

[0118] In an optional embodiment, the target audio acquisition module 203 is specifically used for:

[0119] Performing an alignment process on the first training audio and the second training audio according to the audio alignment point, and determining an alignment area and a non-alignment area corresponding to the alignment process;

[0120] In response to the deletion process for the non-aligned area, a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio are obtained.

[0121] In an optional embodiment, the first target audio and the second target audio have the same window function corresponding to the same amplitude range, and the audio parameter acquisition module 204 is specifically used to:

[0122] Performing windowing processing on the first target audio and the second target audio by using the window function to generate first sliding windows corresponding to the first target audio and second sliding windows corresponding to the second target audio;

[0123] Acquire a first autocorrelation coefficient corresponding to each of the first sliding windows, and use the mutual correlation coefficient and the first autocorrelation coefficient to determine each first sliding window similarity corresponding to each of the first sliding windows and a first sliding window similarity variance;

[0124] The second autocorrelation coefficient corresponding to each second sliding window is obtained, and the mutual correlation coefficient and the second autocorrelation coefficient are used to determine each second sliding window similarity corresponding to each second sliding window and a second sliding window similarity variance.

[0125] In an optional embodiment, the audio parameter acquisition module 204 is specifically used to:

[0126] Acquire a first zero-crossing rate and a first average energy corresponding to the first target audio, and a second zero-crossing rate and a second average energy corresponding to the second target audio;

[0127] Calculate a zero-crossing rate ratio between the first target audio and the second target audio using the first zero-crossing rate and the second zero-crossing rate;

[0128] The first average energy and the second average energy are used to calculate and determine an average energy ratio between the first target audio and the second target audio.

[0129] In an optional embodiment, the preset audio parameter range includes a preset zero-crossing rate ratio range and a preset average energy ratio range, and the model training data screening module 205 is specifically used for:

[0130] If the first sliding window similarity variance of the first target audio and the second sliding window similarity variance of the second target audio are less than or equal to a preset sliding window similarity variance, and three consecutive first sliding window similarities and second sliding window similarities are greater than or equal to a preset sliding window similarity;

[0131] And, if the zero-crossing rate ratio is within a preset zero-crossing rate ratio range, and the average energy ratio is within a preset average energy ratio range, the first target audio and the second target audio are used as target model training data for training a speech recognition model.

[0132] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0133] In addition, an embodiment of the present invention further provides an electronic device, including: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the various processes of the above-mentioned method for screening model training data are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be described here.

[0134] The embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, each process of the above-mentioned method for screening model training data is implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. The computer-readable storage medium is, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0135] Figure 3 A schematic diagram of the hardware structure of an electronic device for implementing various embodiments of the present invention.

[0136] The electronic device 300 includes but is not limited to: a radio frequency unit 301, a network module 302, an audio output unit 303, an input unit 304, a sensor 305, a display unit 306, a user input unit 307, an interface unit 308, a memory 309, a processor 310, and a power supply 311. Those skilled in the art will appreciate that Figure 3 The electronic device structure shown in the figure does not constitute a limitation on the electronic device, and the electronic device may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently. In the embodiments of the present invention, the electronic device includes but is not limited to a mobile phone, a tablet computer, a laptop computer, a PDA, a vehicle-mounted terminal, a wearable device, and a pedometer.

[0137] It should be understood that in the embodiment of the present invention, the radio frequency unit 301 can be used for receiving and sending signals during information transmission or communication. Specifically, after receiving downlink data from the base station, it is sent to the processor 310 for processing; in addition, uplink data is sent to the base station. Generally, the radio frequency unit 301 includes but is not limited to an antenna, at least one amplifier, a transceiver, a coupler, a low noise amplifier, a duplexer, etc. In addition, the radio frequency unit 301 can also communicate with the network and other devices through a wireless communication system.

[0138] The electronic device provides users with wireless broadband Internet access through the network module 302, such as helping users to send and receive emails, browse web pages, and access streaming media.

[0139] The audio output unit 303 can convert the audio data received by the RF unit 301 or the network module 302 or stored in the memory 309 into an audio signal and output it as sound. Moreover, the audio output unit 303 can also provide audio output related to a specific function performed by the electronic device 300 (for example, a call signal reception sound, a message reception sound, etc.). The audio output unit 303 includes a speaker, a buzzer, a receiver, etc.

[0140] The input unit 304 is used to receive audio or video signals. The input unit 304 may include a graphics processor (GPU) 3041 and a microphone 3042, and the graphics processor 3041 processes the image data of a static picture or video obtained by an image capture device (such as a camera) in a video capture mode or an image capture mode. The processed image frame can be displayed on the display unit 306. The image frame processed by the graphics processor 3041 can be stored in the memory 309 (or other storage medium) or sent via the radio frequency unit 301 or the network module 302. The microphone 3042 can receive sound and can process such sound into audio data. The processed audio data can be converted into a format output that can be sent to a mobile communication base station via the radio frequency unit 301 in the case of a telephone call mode.

[0141] The electronic device 300 also includes at least one sensor 305, such as a light sensor, a motion sensor, and other sensors. Specifically, the light sensor includes an ambient light sensor and a proximity sensor, wherein the ambient light sensor can adjust the brightness of the display panel 3061 according to the brightness of the ambient light, and the proximity sensor can turn off the display panel 3061 and / or the backlight when the electronic device 300 is moved to the ear. As a type of motion sensor, the accelerometer sensor can detect the magnitude of acceleration in each direction (generally three axes), and can detect the magnitude and direction of gravity when stationary, which can be used to identify the posture of the electronic device (such as horizontal and vertical screen switching, related games, magnetometer posture calibration), vibration recognition related functions (such as pedometer, tapping), etc.; the sensor 305 can also include a fingerprint sensor, a pressure sensor, an iris sensor, a molecular sensor, a gyroscope, a barometer, a hygrometer, a thermometer, an infrared sensor, etc., which will not be repeated here.

[0142] The display unit 306 is used to display information input by the user or information provided to the user. The display unit 306 may include a display panel 3061, which may be configured in the form of a liquid crystal display (LCD), an organic light-emitting diode (OLED), or the like.

[0143] The user input unit 307 can be used to receive input digital or character information, and to generate key signal input related to user settings and function control of the electronic device. Specifically, the user input unit 307 includes a touch panel 3071 and other input devices 3072. The touch panel 3071, also known as a touch screen, can collect the user's touch operation on or near it (such as the user's operation on the touch panel 3071 or near the touch panel 3071 using any suitable object or accessory such as a finger, stylus, etc.). The touch panel 3071 may include two parts: a touch detection device and a touch controller. Among them, the touch detection device detects the user's touch orientation, detects the signal brought by the touch operation, and transmits the signal to the touch controller; the touch controller receives the touch information from the touch detection device, converts it into the contact point coordinates, and then sends it to the processor 310, receives the command sent by the processor 310 and executes it. In addition, the touch panel 3071 can be implemented using multiple types such as resistive, capacitive, infrared and surface acoustic waves. In addition to the touch panel 3071, the user input unit 307 may also include other input devices 3072. Specifically, other input devices 3072 may include but are not limited to a physical keyboard, function keys (such as volume control keys, switch keys, etc.), a trackball, a mouse, and a joystick, which are not described in detail here.

[0144] Furthermore, the touch panel 3071 may be overlaid on the display panel 3061. When the touch panel 3071 detects a touch operation on or near it, it is transmitted to the processor 310 to determine the type of touch event. Then, the processor 310 provides corresponding visual output on the display panel 3061 according to the type of touch event. Figure 3 In the figure, the touch panel 3071 and the display panel 3061 are used as two independent components to realize the input and output functions of the electronic device. However, in some embodiments, the touch panel 3071 and the display panel 3061 can be integrated to realize the input and output functions of the electronic device, which is not limited here.

[0145] The interface unit 308 is an interface for connecting an external device to the electronic device 300. For example, the external device may include a wired or wireless headset port, an external power supply (or battery charger) port, a wired or wireless data port, a memory card port, a port for connecting a device with an identification module, an audio input / output (I / O) port, a video I / O port, a headphone port, etc. The interface unit 308 may be used to receive input (e.g., data information, power, etc.) from an external device and transmit the received input to one or more elements within the electronic device 300 or may be used to transmit data between the electronic device 300 and an external device.

[0146] The memory 309 can be used to store software programs and various data. The memory 309 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory 309 can include a high-speed access memory, and can also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0147] The processor 310 is the control center of the electronic device. It uses various interfaces and lines to connect various parts of the entire electronic device. By running or executing software programs and / or modules stored in the memory 309 and calling data stored in the memory 309, it performs various functions of the electronic device and processes data, thereby monitoring the electronic device as a whole. The processor 310 may include one or more processing units; preferably, the processor 310 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 310.

[0148] The electronic device 300 may also include a power supply 311 (such as a battery) for supplying power to each component. Preferably, the power supply 311 may be logically connected to the processor 310 through a power management system, thereby implementing functions such as charging, discharging, and power consumption management through the power management system.

[0149] In addition, the electronic device 300 includes some functional modules not shown, which will not be described in detail here.

[0150] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0151] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0152] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

[0153] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the embodiments of the present invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0154] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0155] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0156] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0157] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0158] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, ROM, RAM, magnetic disks, or optical disks.

[0159] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for screening model training data, characterized in that: include: Acquire a first training audio and a second training audio from each initial audio; Acquire a sequence of mutual correlation coefficients between the first training audio and the second training audio, and use the mutual correlation coefficient with the largest value as an audio alignment point between the first training audio and the second training audio; Aligning the first training audio with the second training audio according to the audio alignment point to obtain a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio; Obtaining a first sliding window similarity variance of the first target audio, a second sliding window similarity variance of the second target audio, and a zero crossing rate ratio and an average energy ratio between the first target audio and the second target audio; If the first sliding window similarity variance and the second sliding window similarity variance are less than or equal to a preset sliding window similarity variance, and the zero crossing rate ratio and the average energy ratio are within a preset audio parameter range, the first target audio and the second target audio are used as target model training data for training a speech recognition model; During the process of aligning the first training audio and the second training audio by the audio alignment point, an alignment area and a non-alignment area between the first training audio and the second training audio are determined, and the non-alignment area is deleted to obtain the first target audio corresponding to the first training audio and the second target audio corresponding to the second training audio in the alignment area.

2. The method according to claim 1, characterized in that The obtaining of the first training audio and the second training audio from each of the initial audios includes: Two initial audios are selected from the initial model training data for training, and amplitude normalization processing is performed on the frame sequences of the two initial audios to obtain the first training audio and the second training audio whose audio amplitudes are in the same amplitude range.

3. The method according to claim 2, characterized in that The two initial audios include a first initial audio and a second initial audio, and the step of obtaining the first training audio and the second training audio whose audio amplitudes are in the same amplitude range includes: Performing frame processing on the first initial audio and the second initial audio respectively to obtain a first frame sequence of the first initial audio and a second frame sequence of the second initial audio; Acquire a first sampling rate of the first initial audio and a second sampling rate of the second initial audio; If the first sampling rate is less than the second sampling rate, inserting at least one zero-value point at equal intervals in the first frame sequence to generate the first training audio whose audio amplitudes are in the same amplitude range; If the second sampling rate is less than the first sampling rate, at least one zero-value point is inserted at equal intervals in the second frame sequence to generate the second training audio whose audio amplitude is in the same amplitude range.

4. The method according to claim 3, characterized in that The aligning the first training audio and the second training audio according to the audio alignment point to obtain a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio includes: Performing an alignment process on the first training audio and the second training audio according to the audio alignment point, and determining an alignment area and a non-alignment area corresponding to the alignment process; In response to the deletion process for the non-aligned area, a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio are obtained.

5. The method according to claim 2, characterized in that: The first target audio and the second target audio have the same window function corresponding to the same amplitude range, and obtaining the first sliding window similarity variance of the first training audio and the second sliding window similarity variance of the second training audio includes: Performing windowing processing on the first target audio and the second target audio by using the window function to generate first sliding windows corresponding to the first target audio and second sliding windows corresponding to the second target audio; Acquire a first autocorrelation coefficient corresponding to each of the first sliding windows, and use the mutual correlation coefficient and the first autocorrelation coefficient to determine each first sliding window similarity corresponding to each of the first sliding windows and a first sliding window similarity variance; The second autocorrelation coefficient corresponding to each second sliding window is obtained, and the mutual correlation coefficient and the second autocorrelation coefficient are used to determine each second sliding window similarity corresponding to each second sliding window and a second sliding window similarity variance.

6. The method according to claim 5, characterized in that The obtaining of a zero-crossing rate ratio and an average energy ratio between the first target audio and the second target audio includes: Acquire a first zero-crossing rate and a first average energy corresponding to the first target audio, and a second zero-crossing rate and a second average energy corresponding to the second target audio; Calculate a zero-crossing rate ratio between the first target audio and the second target audio using the first zero-crossing rate and the second zero-crossing rate; The first average energy and the second average energy are used to calculate and determine an average energy ratio between the first target audio and the second target audio.

7. The method according to claim 6, characterized in that The preset audio parameter range includes a preset zero-crossing rate ratio range and a preset average energy ratio range, and if the first sliding window similarity variance and the second sliding window similarity variance are less than or equal to the preset sliding window similarity variance, and the zero-crossing rate ratio and the average energy ratio are within the preset audio parameter range, then the first target audio and the second target audio are used as target model training data for training a speech recognition model, including: If the first sliding window similarity variance of the first target audio and the second sliding window similarity variance of the second target audio are less than or equal to a preset sliding window similarity variance, and three consecutive first sliding window similarities and second sliding window similarities are greater than or equal to a preset sliding window similarity; And, if the zero-crossing rate ratio is within a preset zero-crossing rate ratio range, and the average energy ratio is within a preset average energy ratio range, the first target audio and the second target audio are used as target model training data for training a speech recognition model.

8. A device for screening model training data, characterized in that: include: A training audio acquisition module, used to acquire a first training audio and a second training audio from each initial audio; an audio alignment point determination module, configured to obtain a sequence of mutual correlation coefficients between the first training audio and the second training audio, and use the mutual correlation coefficient with the largest value as an audio alignment point between the first training audio and the second training audio; a target audio acquisition module, configured to align the first training audio and the second training audio according to the audio alignment point, to obtain a first target audio corresponding to the first training audio and a second target audio corresponding to the second training audio; An audio parameter acquisition module, used to acquire a first sliding window similarity variance of the first target audio, a second sliding window similarity variance of the second target audio, and a zero crossing rate ratio and an average energy ratio between the first training audio and the second training audio; A model training data screening module, configured to use the first target audio and the second target audio as target model training data for training a speech recognition model if the first sliding window similarity variance and the second sliding window similarity variance are less than or equal to a preset sliding window similarity variance, and the zero crossing rate ratio and the average energy ratio are within a preset audio parameter range; During the process of aligning the first training audio and the second training audio, the audio alignment point determination module determines the alignment area and the non-alignment area between the first training audio and the second training audio, deletes the non-alignment area, and obtains the first target audio corresponding to the first training audio and the second target audio corresponding to the second training audio in the alignment area.

9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; The memory is used to store computer programs; The processor is used to implement the method according to any one of claims 1 to 7 when executing the program stored in the memory.

10. A computer-readable storage medium having instructions stored thereon, which, when executed by one or more processors, cause the processors to perform the method according to any one of claims 1 to 7.

Images