Speech recognition data preprocessing platform

Abstract

The invention discloses a speech recognition data preprocessing platform, relates to the technical field of speech recognition, and solves the technical problem of low speech recognition data preprocessing efficiency possibly caused by neglecting and editing texts such as text segments, sentences and words which need to be recognized and scored without considering that the editing of standard texts is too tedious in the prior art. According to the invention, voice signals in the environment are collected through the audio collector; compensating a high-frequency signal of the voice signal by using a high-pass filter; dividing a voice signal into a plurality of short-time small frames through a voice framing technology, and obtaining an analog voice signal by applying a window function; extracting an initial feature vector of the digital voice signal through a Mel-frequency cepstrum coefficient; analyzing the dynamic change rate of the digital voice signal through the dynamic difference coefficient; analyzing according to the initial feature vector and the dynamic change rate to obtain a final feature vector; calculating a mean value and a standard deviation of the final feature vectors, and classifying the digital voice signals; and the speech recognition data preprocessing efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and relates to speech recognition technology, in particular to a speech recognition data preprocessing platform. Background Art

[0002] Speech recognition allows users to interact with computers or smart devices through natural language without manual input or reliance on other physical interfaces. This approach can significantly reduce input time and improve operational efficiency, especially in scenarios where large amounts of information need to be input quickly or real-time interactions need to be conducted. Speech recognition technology makes human-computer interaction more natural and intuitive. Users can communicate with devices without having to learn specific commands or language structures. Speech recognition lowers the threshold for use, allowing more users to easily use smart devices and services. For users with visual impairments or hand movement disorders, speech recognition technology provides an effective auxiliary means, enabling them to more conveniently obtain information, use devices and services. With the continuous advancement of technology, the accuracy and reliability of speech recognition systems are also constantly improving, promoting innovation and application in related fields. Speech recognition technology can collect and process large amounts of speech data in real time, providing rich resources for data analysis, mining and application, so as to help companies understand user needs and behavioral habits, and thus develop more effective market strategies and service plans.

[0003] The prior art (invention patent application with publication number CN106710587A) discloses a speech recognition data preprocessing method. The prior art provides a speech recognition data preprocessing method, including a standard audio file organization module, a standard text editing module, a pronunciation dictionary configuration module, a speech model generation module, and a standard sound feature data recognition and processing module; the final generated standard sound data model is saved to the file system, and the pre-generated data model is directly loaded in the application product to recognize and score the user recording; the prior art organizes and edits the texts such as paragraphs, sentences, and words that need to be recognized and scored, but does not take into account that editing standard texts is too cumbersome, which may lead to the problem of low efficiency of speech recognition data preprocessing.

[0004] The present invention provides a speech recognition data preprocessing platform to solve the above technical problems. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a speech recognition data preprocessing platform, which is used to solve the technical problem that the prior art organizes and edits texts such as paragraphs, sentences, words, etc. that need to be recognized and scored, but does not take into account that editing standard texts is too cumbersome, which may lead to low efficiency of speech recognition data preprocessing.

[0006] To achieve the above object, the first aspect of the present invention provides a speech recognition data preprocessing platform, comprising:

[0007] Front-end acoustic processing module: collect voice signals in the environment through audio collector; use high-pass filter to compensate for high-frequency signals of voice signals; divide voice signals into several short-time frames through voice framing technology, apply window function on several short-time frames to obtain analog voice signals; convert analog voice signals into digital voice signals, and remove interference signals in digital voice signals through filtering technology; filtering technology includes: noise suppression and echo cancellation;

[0008] Feature extraction module: extract the initial feature vector of the digital speech signal through the Mel frequency cepstrum coefficient; analyze the dynamic change rate of the digital speech signal through the dynamic differential coefficient; obtain the final feature vector based on the initial feature vector and the dynamic change rate analysis;

[0009] Data grouping module: Calculate the mean and standard deviation of the final feature vector, and classify the digital speech signal according to the mean and standard deviation.

[0010] Preferably, the compensating the high-frequency signal of the speech signal by using a high-pass filter comprises:

[0011] Divide the speech signal into several samples, select one sample from the several samples as the target sample; mark the target sample as x[n], and mark the previous sample of the target sample as x[n-1];

[0012] The value of the partial speech signal after pre-emphasis processing is calculated by the differential equation y[n]=x[n]-α×x[n-1]; wherein α is a proportional coefficient greater than 0 and less than 1;

[0013] Performing pre-emphasis processing on a number of samples in the speech signal by using a differential equation to obtain a number of partial speech signals, and integrating the number of partial speech signals into a speech signal one;

[0014] Initialize a first-order high-pass filter, and use the first-order high-pass filter to compensate for the high-frequency signal in the speech signal one to obtain the speech signal two.

[0015] It should be noted that the closer the value of α is to 1, the more obvious the enhancement effect of the high-frequency part is; when the value of α is too large, it will cause signal distortion or introduce unnecessary noise.

[0016] The present invention can restore the balance of the spectrum by using a high-pass filter to compensate for the high-frequency signal, which not only makes the voice signal have a good signal-to-noise ratio and spectrum characteristics in the entire frequency band from low frequency to high frequency, but also makes the characteristics of the voice signal more prominent, which helps to improve the auditory effect of the voice, thereby enhancing the clarity of the voice.

[0017] Preferably, the voice signal is divided into a number of short-time small frames by using the voice framing technology, including:

[0018] The second speech signal is retrieved, and a sliding window is slid on the second speech signal according to the frame length and the frame shift to obtain a frame interval, and the second speech signal is divided according to the frame interval to form a number of short-time small frames.

[0019] It should be noted that the frame shift is the time interval between consecutive frames, which is usually less than or equal to the frame length; a common frame shift is half of the frame length or less.

[0020] The present invention sets the frame shift to be half of the frame length or less, so that the signal within the frame has sufficient stability, which not only avoids the loss of information between frames, but also helps to retain the relevant information between adjacent frames.

[0021] Preferably, the step of applying a window function to a plurality of short-time small frames to obtain a simulated speech signal comprises:

[0022] Retrieve several short-duration frames and use the frame length as the length N of the window function;

[0023] The Hamming window function is calculated by the formula w(n)=0.54-0.46*cos(2*π*n / (N-1)); wherein n represents an integer from 0 to N-1;

[0024] Apply the Hamming window function to several short-time frames to obtain several frame signals after windowing; integrate the several frame signals into an analog speech signal.

[0025] The present invention realizes smoothing frame boundaries and reducing spectrum leakage by adding windows. After framing, a window function is applied to each frame to reduce the discontinuity of the frame boundary; the Hamming window is gradually attenuated to zero at both ends of the window to reduce the frame boundary effect;

[0026] Preferably, removing interference signals from digital voice signals by filtering technology includes:

[0027] Two characteristic thresholds are set by double threshold method, and characteristic values ​​of digital speech signals are extracted by fast Fourier transform technology;

[0028] Determine whether the characteristic value is between two characteristic thresholds; if yes, determine that the digital voice signal meets the requirements and retain the digital voice signal; if no, determine that the corresponding digital voice signal is an interference signal, and use filtering technology to filter the interference signal; wherein the characteristic value includes: the energy value and zero-crossing rate of the frame signal.

[0029] The present invention can significantly improve the quality of the final processed speech signal by removing or weakening those signal parts whose characteristic values ​​are not within a preset range, and help reduce the impact of background noise, electromagnetic interference, interference from other sound sources, etc. on the target speech signal; by setting a characteristic threshold and judging whether the signal is an interference signal based on it, and taking corresponding filtering measures, the risk of system performance degradation or failure due to external factors is reduced.

[0030] Preferably, the two feature thresholds are set by the double threshold method, including:

[0031] Two feature thresholds T1 and T2 are set by the double threshold method according to the short-time energy and the short-time zero-crossing rate, where T1 is used to roughly determine whether the speech has started, and T2 is used to determine the start and end time of the speech;

[0032] The acoustic model is used to identify speech and non-speech signals and predict the start and end time of speech.

[0033] It should be noted that frames in non-speech activity time periods and containing abnormal features are determined to be erroneous data or irrelevant information and are directly removed; wherein the abnormal features include: frames with too high or too low energy and zero crossing rate; the acoustic models include: speech recognition model and hidden Markov model;

[0034] The present invention sets threshold levels, uses T1 as a low threshold to preliminarily screen frame segments that may contain speech, and reduces further processing of obvious non-speech segments; uses T2 as a high threshold to more accurately determine the start and end of the speech signal, thereby improving the accuracy of detection; and uses a double threshold method to reduce unnecessary calculations, so that the system can more effectively utilize computing resources and improve system performance.

[0035] Preferably, the extracting the initial feature vector of the digital speech signal by Mel-frequency cepstral coefficients comprises:

[0036] Retrieve several digital speech signals and perform fast Fourier transform on the digital speech signals; pass the spectrum after Fourier transform through a Mel filter bank to obtain the energy distribution on the Mel scale; take the logarithm of the energy on each Mel scale and perform discrete cosine transform, and map the result to the cepstrum domain; take the first N coefficients after discrete cosine transform as the initial feature vector.

[0037] It should be noted that the Mel filter bank is densely distributed in the low-frequency area and sparsely distributed in the high-frequency area.

[0038] The present invention extracts key features related to speech recognition and other speech processing tasks from the original speech signal. These features usually have a lower dimension and can effectively describe the characteristics of the speech signal such as sound quality, pitch, length and volume. These feature vectors can represent the essential content of the speech and reduce the data dimension, which is convenient for subsequent model matching and analysis.

[0039] Preferably, analyzing the dynamic change rate of the digital voice signal by means of a dynamic differential coefficient comprises:

[0040] The differential coefficients of the digital speech signal are extracted, and the digital speech signal is identified according to the differential coefficients to obtain the dynamic change rate; wherein the differential coefficients include first-order differential coefficients and second-order differential coefficients; the first-order differential coefficients represent the change rate of the Mel-frequency cepstral coefficients over time, and the second-order differential coefficients represent the change rate of the Mel-frequency cepstral coefficients over time.

[0041] The present invention processes speech signals by differential processing, so that subtle changes in the signals can be retained and amplified, and these subtle changes may be masked in the original signals due to noise or other interference; these detailed information is crucial to improving the accuracy of recognition; and the differential coefficients can distinguish different types of speech signals, such as unvoiced sounds, voiced sounds, and transitions between different phonemes, etc. These differences are more obvious in the differential coefficients, which helps to improve the recognition system's ability to distinguish different speech signals.

[0042] Preferably, the classifying of the digital speech signal according to the mean and the standard deviation comprises:

[0043] The j-th feature value of the i-th frame is marked as Vij, and the total number of frames is marked as N;

[0044] By formula Calculate the mean of each feature vector in all speech frames;

[0045] By formula Calculate the standard deviation of each feature vector in all speech frames;

[0046] By formula Normalize the feature vector;

[0047] The normalized feature vector is used as a classification input sequence to retrieve a signal classification model; the classification input sequence is input into the signal classification model to obtain the type of the corresponding digital speech signal; wherein the signal classification model is constructed based on an artificial intelligence model.

[0048] It should be noted that when calculating the mean and standard deviation, these statistics are usually calculated using training data, and the same statistics are used during training and testing; when normalizing feature vectors, each eigenvalue is scaled to a range of mean 0 and standard deviation 1, which helps the model better learn the relationship between features.

[0049] The present invention uses the mean as the size of the DC component in the signal to reflect the overall level or baseline of the speech signal; the standard deviation can reflect the degree of fluctuation of the signal value relative to the mean. In speech signal processing, the standard deviation can reveal the stability and variability of the signal; by combining the two features of the mean and the standard deviation, a more comprehensive description and classification of the digital speech signal can be performed. This comprehensive judgment method is more accurate and reliable than single feature classification.

[0050] Preferably, the signal classification model is constructed based on an artificial intelligence model, including:

[0051] Acquire standard training data; wherein the standard training data includes standard input data consistent with the content attributes of the classification input sequence, and standard output data consistent with the content attributes of the signal classification;

[0052] The artificial intelligence model is trained using standard training data, and the trained artificial intelligence model is marked as a signal classification model; wherein the artificial intelligence model includes a convolutional neural network model or a long short-term memory neural network model.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. The present invention comprises a front-end acoustic processing module, a feature extraction module and a data grouping module, wherein the front-end acoustic processing module performs pre-emphasis processing, which can compensate for the attenuation of high-frequency signals during transmission, making the signal of the entire frequency band more balanced, which is beneficial to subsequent processing; the framing and windowing operations divide the continuous speech signal into short segments, and apply the window function on each frame to reduce the frame boundary effect, making the processing more efficient; the feature extraction module can minimize the adverse effects of factors such as environmental noise and communication channel quality differences on speech features, thereby enhancing the robustness of the system; the data grouping module can group data to make similar data more compact and heterogeneous data more dispersed, which is helpful to better capture the inherent laws and characteristics of data when training an acoustic model, and improve the generalization ability of the model.

[0055] 2. Set threshold levels, with T1 as the low threshold for preliminary screening of frames that may contain speech, reducing further processing of obvious non-speech segments; T2 as the high threshold for more accurately determining the start and end of speech signals, improving detection accuracy; using the double threshold method can reduce unnecessary calculations, enabling the system to more effectively utilize computing resources and improve system performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0057] Figure 1 A schematic diagram of the modules included in the present invention;

[0058] Figure 2 A schematic diagram of the speech recognition preprocessing process of the present invention;

[0059] Figure 3 Schematic diagram of the first step of speech recognition data preprocessing of the present invention;

[0060] Figure 4 Schematic diagram of the second step of speech recognition data preprocessing of the present invention. DETAILED DESCRIPTION

[0061] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] See also Figure 1-Figure 2 The first aspect of the present invention provides a speech recognition data preprocessing platform, including:

[0063] Front-end acoustic processing module: collect voice signals in the environment through audio collector; use high-pass filter to compensate for high-frequency signals of voice signals; divide voice signals into several short-time frames through voice framing technology, apply window function on several short-time frames to obtain analog voice signals; convert analog voice signals into digital voice signals, and remove interference signals in digital voice signals through filtering technology; filtering technology includes: noise suppression and echo cancellation;

[0064] Feature extraction module: extract the initial feature vector of the digital speech signal through the Mel frequency cepstrum coefficient; analyze the dynamic change rate of the digital speech signal through the dynamic differential coefficient; obtain the final feature vector based on the initial feature vector and the dynamic change rate analysis;

[0065] Data grouping module: Calculate the mean and standard deviation of the final feature vector, and classify the digital speech signal according to the mean and standard deviation.

[0066] See also Figure 3 , the first step of speech recognition data preprocessing is to collect speech signals in the environment through an audio collector; divide the speech signal into several samples, and select one sample from the several samples as a target sample; mark the target sample as x[n], and mark the previous sample of the target sample as x[n-1];

[0067] The value of the partial speech signal after pre-emphasis processing is calculated by the differential equation y[n]=x[n]-α×x[n-1]; wherein α is a proportional coefficient greater than 0 and less than 1;

[0068] Performing pre-emphasis processing on a number of samples in the speech signal by using a differential equation to obtain a number of partial speech signals, and integrating the number of partial speech signals into a speech signal one;

[0069] Initialize a first-order high-pass filter, use the first-order high-pass filter to compensate the high-frequency signal in the speech signal one, and obtain the speech signal two; slide the sliding window on the speech signal two according to the frame length and the frame shift to obtain the frame interval, and divide the speech signal two according to the frame interval to form a number of short-time small frames; use the frame length as the length N of the window function; calculate the Hamming window function by the formula w(n)=0.54-0.46*cos(2*π*n / (N-1)); wherein n represents an integer from 0 to N-1; apply the Hamming window function to a number of short-time small frames to obtain a number of windowed frame signals; integrate the several frame signals into an analog speech signal; according to the short-time The energy and short-time zero-crossing rate are used to set two characteristic thresholds T1 and T2 through the double threshold method, wherein T1 is used to roughly judge whether the speech has started, and T2 is used to determine the start and end time of the speech; the speech and non-speech signals are identified through the acoustic model, and the start and end time of the speech are predicted; the characteristic value of the digital speech signal is extracted through the fast Fourier transform technology; it is judged whether the characteristic value is between the two characteristic thresholds; if yes, it is judged that the digital speech signal meets the requirements and the digital speech signal is retained; if not, it is judged that the corresponding digital speech signal is an interference signal, and the interference signal is filtered using filtering technology; wherein the characteristic values ​​include: the energy value and the zero-crossing rate of the frame signal.

[0070] For example, there is a speech signal, the speech signal is divided into twenty segments of samples, sample A is selected from the twenty segments of samples as the target sample x[n], and the previous sample of sample A is recorded as Bx[n-1]; if A is set to 10, B is set to 15; the proportional coefficient α is set to 0.1; the value of the partial speech signal after pre-emphasis processing is calculated by the differential equation y[n]=x[n]-α×x[n-1] to be 11.5; the twenty segments of samples in the speech signal are pre-emphasized by using the differential equation to obtain several partial speech signals, and the twenty segments of partial speech signals are integrated into speech signal one; a first-order high-pass filter is initialized, and the high-pass filter in speech signal one is filtered by the first-order high-pass filter. The frequency signal is compensated to obtain a 120 millisecond speech signal 2; the frame length is set to 20 milliseconds, so that each frame contains 20 milliseconds of speech signal; the frame shift is set to 10 milliseconds, then after taking a frame, 10 milliseconds of time is skipped, and then the next frame is taken; according to the set frame length and frame shift sliding window, the speech signal 2 is slid to obtain the frame interval, and the speech signal 2 is divided according to the frame interval to form four short-time small frames; the frame length of 20 milliseconds is used as the length N of the window function, and the Hamming window function is obtained as: w(20)=0.54-0.46*cos(2*π*20 / 19); the Hamming window function is applied to the four short-time small frames to obtain four frame signals after windowing;

[0071] Set two feature thresholds T1 to 10 and T2 to 15; extract two feature values ​​of the digital voice signal through the fast Fourier transform technology and mark them as t1 and t2, where t1 is 9 and t2 is 13; if the feature value t1 is less than T1, the digital voice signal corresponding to t1 is an interference signal; if the feature value t2 is greater than T1 and less than T2, it is determined that the digital voice signal corresponding to t2 meets the requirements and the digital voice signal is retained.

[0072] See also Figure 4 , the second step of speech recognition data preprocessing is to retrieve several digital speech signals and perform fast Fourier transform on the digital speech signals; pass the spectrum after Fourier transform through the Mel filter group to obtain the energy distribution on the Mel scale; take the logarithm of the energy on each Mel scale and perform discrete cosine transform, and map the result to the cepstrum domain; take the first N coefficients after discrete cosine transform as the initial feature vector; extract the differential coefficient of the digital speech signal, and identify the digital speech signal according to the differential coefficient to obtain the dynamic change rate; wherein the differential coefficient includes the first-order differential coefficient and the second-order differential coefficient; the first-order differential coefficient represents the change rate of the Mel frequency cepstrum coefficient over time, and the second-order differential coefficient represents the change rate of the Mel frequency cepstrum coefficient over time; mark the j-th eigenvalue of the i-th frame as Vij, and mark the total number of frames as N;

[0073] By formula Calculate the mean of each feature vector in all speech frames;

[0074] By formula Calculate the standard deviation of each feature vector in all speech frames;

[0075] By formula Normalize the feature vector;

[0076] The normalized feature vector is used as a classification input sequence to retrieve a signal classification model; the classification input sequence is input into the signal classification model to obtain the type of the corresponding digital speech signal; wherein the signal classification model is constructed based on an artificial intelligence model.

[0077] For example, there are three existing speech frames recorded as C1, C2, and C3, and each speech frame has a feature value; the feature value corresponding to C1 is 5, the feature value corresponding to C2 is 6, and the feature value corresponding to C3 is 7;

[0078] Then through the formula Calculated

[0079] By formula Calculated

[0080] By formula Normalize the feature vector.

[0081] Part of the data in the above formula is calculated by removing the dimension and taking its numerical value. The formula is a formula closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0082] The working principle of the present invention is as follows: voice signals in an environment are collected by an audio collector; high-pass filters are used to compensate for high-frequency signals of voice signals; voice signals are divided into a number of short-time frames by voice framing technology, and window functions are applied to the short-time frames to obtain analog voice signals; the analog voice signals are converted into digital voice signals, and interference signals in the digital voice signals are removed by filtering technology; initial feature vectors of digital voice signals are extracted by Mel-frequency cepstrum coefficients; dynamic change rates of digital voice signals are analyzed by dynamic differential coefficients; final feature vectors are obtained by analyzing the initial feature vectors and the dynamic change rates; the mean and standard deviation of the final feature vectors are calculated, and the digital voice signals are classified according to the mean and standard deviation.

[0083] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. Speech recognition data preprocessing platform, characterized by: include: Front-end acoustic processing module, feature extraction module and data grouping module; Front-end acoustic processing module: collects voice signals in the environment through the audio collector; A high-pass filter is used to compensate the high-frequency signal of the speech signal; the speech signal is divided into several short-time small frames through the speech framing technology, and a window function is applied to several short-time small frames to obtain a simulated speech signal; Convert analog voice signals into digital voice signals, and remove interference signals in the digital voice signals through filtering technology; wherein filtering technology includes: noise suppression and echo cancellation; Feature extraction module: extract the initial feature vector of the digital speech signal through the Mel frequency cepstrum coefficient; analyze the dynamic change rate of the digital speech signal through the dynamic differential coefficient; obtain the final feature vector based on the initial feature vector and the dynamic change rate analysis; Data grouping module: Calculate the mean and standard deviation of the final feature vector, and classify the digital speech signal according to the mean and standard deviation.

2. The speech recognition data preprocessing platform according to claim 1, characterized in that: The method of compensating the high frequency signal of the speech signal by using a high pass filter comprises: Divide the speech signal into several samples, select one sample from the several samples as the target sample; mark the target sample as x[n], and mark the previous sample of the target sample as x[n-1]; The value of the partial speech signal after pre-emphasis processing is calculated by the differential equation y[n]=x[n]-α×x[n-1]; wherein α is a proportional coefficient greater than 0 and less than 1; Performing pre-emphasis processing on a number of samples in the speech signal by using a differential equation to obtain a number of partial speech signals, and integrating the number of partial speech signals into a speech signal one; Initialize a first-order high-pass filter, and use the first-order high-pass filter to compensate for the high-frequency signal in the speech signal one to obtain the speech signal two.

3. The speech recognition data preprocessing platform according to claim 1, characterized in that: The speech signal is divided into a number of short-time small frames by using the speech framing technology, including: The second voice signal is retrieved, and a sliding window is slid on the second voice signal according to the frame length and the frame shift to obtain a frame interval, and the second voice signal is divided according to the frame interval to form a plurality of short-time small frames.

4. The speech recognition data preprocessing platform according to claim 1, characterized in that: The step of applying a window function to a plurality of short-time small frames to obtain a simulated speech signal comprises: Retrieve several short-duration frames and use the frame length as the length N of the window function; The Hamming window function is calculated by the formula w(n)=0.54-0.46*cos(2*π*n / (N-1)); wherein n represents an integer from 0 to N-1; Apply the Hamming window function to several short-time frames to obtain several frame signals after windowing; integrate the several frame signals into an analog speech signal.

5. The speech recognition data preprocessing platform according to claim 1, characterized in that: The method of removing interference signals from digital voice signals by filtering technology includes: Two characteristic thresholds are set by double threshold method, and characteristic values ​​of digital speech signals are extracted by fast Fourier transform technology; Determine whether the characteristic value is between two characteristic thresholds; if yes, determine that the digital voice signal meets the requirements and retain the digital voice signal; if no, determine that the corresponding digital voice signal is an interference signal, and use filtering technology to filter the interference signal; wherein the characteristic value includes: the energy value and zero-crossing rate of the frame signal.

6. The speech recognition data preprocessing platform according to claim 5, characterized in that: The two feature thresholds are set by the double threshold method, including: Two feature thresholds T1 and T2 are set by the double threshold method according to the short-time energy and the short-time zero-crossing rate, where T1 is used to roughly determine whether the speech has started, and T2 is used to determine the start and end time of the speech; The acoustic model is used to identify speech and non-speech signals and predict the start and end time of speech.

7. The speech recognition data preprocessing platform according to claim 1, characterized in that: The method of extracting the initial feature vector of the digital speech signal by using the Mel-frequency cepstral coefficients includes: Retrieve several digital speech signals and perform fast Fourier transform on the digital speech signals; pass the spectrum after Fourier transform through a Mel filter bank to obtain the energy distribution on the Mel scale; take the logarithm of the energy on each Mel scale and perform discrete cosine transform, and map the result to the cepstrum domain; take the first N coefficients after discrete cosine transform as the initial feature vector.

8. The speech recognition data preprocessing platform according to claim 1, characterized in that: The method of analyzing the dynamic change rate of the digital voice signal by using the dynamic differential coefficient includes: The differential coefficients of the digital speech signal are extracted, and the digital speech signal is identified according to the differential coefficients to obtain the dynamic change rate; wherein the differential coefficients include first-order differential coefficients and second-order differential coefficients; the first-order differential coefficients represent the change rate of the Mel-frequency cepstral coefficients over time, and the second-order differential coefficients represent the change rate of the Mel-frequency cepstral coefficients over time.

9. The speech recognition data preprocessing platform according to claim 1, characterized in that: The method of classifying the digital voice signal according to the mean and the standard deviation includes: The j-th feature value of the i-th frame is marked as Vij, and the total number of frames is marked as N; By formula Calculate the mean of each feature vector in all speech frames; By formula Calculate the standard deviation of each feature vector in all speech frames; By formula Normalize the feature vector; The normalized feature vector is used as a classification input sequence to retrieve a signal classification model; the classification input sequence is input into the signal classification model to obtain the type of the corresponding digital speech signal; wherein the signal classification model is constructed based on an artificial intelligence model.

10. The speech recognition data preprocessing platform according to claim 9, characterized in that: The signal classification model is constructed based on an artificial intelligence model, including: Acquire standard training data; wherein the standard training data includes standard input data consistent with the content attributes of the classification input sequence, and standard output data consistent with the content attributes of the signal classification; The artificial intelligence model is trained using standard training data, and the trained artificial intelligence model is marked as a signal classification model; wherein the artificial intelligence model includes a convolutional neural network model or a long short-term memory neural network model.

Images