Speech recognition system and method based on artificial intelligence
By building a speech recognition system based on artificial intelligence, using feature filter groups and loop speech recognition models to process user voice data, the problem of recognition errors in existing systems in complex speech and special accent processing is solved, and higher recognition accuracy and real-timeness are achieved.
Patent Information
- Application Number
- CN202411798188.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-09
- Publication Date
- 2025-02-25
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing speech recognition systems are prone to recognition errors when dealing with complex speech, special accents, background noise, complex semantics and context analysis, and lack the ability to train and adapt to different speech and accents.
Using a speech recognition system based on artificial intelligence, the user voice data and sample data are collected through the data acquisition module, the data processing module performs data segmentation and feature extraction, builds a feature filter group and a cyclic speech recognition model, and performs cyclic recognition to obtain interactive text data.
It improves the accuracy and real-time nature of speech recognition, reduces the error rate of manual dictation, and enhances the system's ability to process complex speech and special accents.
Smart Images
Figure CN119517037A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a speech recognition system and method based on artificial intelligence. Background Art
[0002] Artificial intelligence is a field of science and technology that studies how to enable computers to think and act in a human-like intelligent way. It is committed to developing and building intelligent systems that can perceive, understand, learn, reason and make decisions, so that computers have human-like cognitive abilities and intelligence levels.
[0003] Speech recognition system is an artificial intelligence technology that aims to convert human voice input into text or commands. The goal of such a system is to accurately recognize and understand human speech in order to execute corresponding instructions or conduct further semantic analysis. The basic principle of speech recognition system is to convert speech signals into digital signals, and then use specific algorithms and models to recognize and parse these digital signals. However, there are also some disadvantages: recognition errors are prone to occur when dealing with complex speech or special accents; background noise can greatly interfere with the clarity and accuracy of speech signals, resulting in a decrease in the recognition ability of the system; for complex semantic and contextual analysis, the system may have recognition errors or interpretation errors; for different voices and accents, the training and adaptability of the system are also challenging; therefore, it is of great theoretical and practical significance to improve the shortcomings of speech recognition through artificial intelligence.
[0004] How to use artificial intelligence technology to process the collected voice data, obtain the characteristic signal spectrum, and build a cyclic speech recognition model based on the characteristic signal spectrum, and obtain interactive text data by performing cyclic recognition on the cyclic speech recognition model, is a problem we need to solve. To this end, a speech recognition system and method based on artificial intelligence is now provided. Summary of the invention
[0005] The object of the present invention is to provide a speech recognition system and method based on artificial intelligence.
[0006] The purpose of the present invention can be achieved through the following technical solutions: A speech recognition system based on artificial intelligence includes a control center, wherein the control center is connected to a data acquisition module, a data processing module, a data analysis module and a comprehensive recognition module; The process of the data acquisition module collecting user voice data and sample data includes: Setting a collection port, wherein the collection port includes a plurality of voice capture points; The collection port is used to collect the user's voice data. The voice data collected by the collection port is captured through the voice capture point to obtain the user's voice data, sample data and voice duration.
[0007] The process of the data processing module processing the user voice data and the sample data includes: Performing data segmentation on the obtained sample voice data to obtain a sample audio signal, a sample sampling frequency, and a sample voice segment, and performing data segmentation on the obtained user voice data to obtain a user audio signal, a user sampling frequency, and a user voice segment; The obtained sample audio signals are matched with the sample speech segments, and the successfully matched sample audio signals are associated with the sample speech segments; the sample speech segments are combined within the sample speech duration to obtain a sample speech segment group, and a sample segment group waveform is generated according to the sample audio signals associated with the sample speech segments in the sample speech segment group.
[0008] The process of obtaining the sample continuous audio signal and the user continuous audio signal includes: Performing signal framing on the sample audio signal according to the obtained sample sampling frequency to obtain a frame signal, and marking the length of the obtained frame signal as the frame length; Setting a signal window according to the obtained sample speech segment, and marking the length of the obtained signal window as the signal window length; Mark the obtained signal window length and frame length in the sample segment group waveform diagram, segment the waveform in the sample segment group waveform diagram according to the signal window length to obtain the signal window segment, and move and overlap the frame length of the sample segment group waveform diagram containing the signal window segment to obtain a sample continuous audio signal; The process of obtaining the sample continuous audio signal is repeated according to the obtained user audio signal, the user sampling frequency and the user voice segment to obtain the user continuous audio signal.
[0009] The process of constructing an eigenfilter bank includes: Performing fast Fourier transform on each frame signal in the sample continuous audio signal to obtain a sample frame signal spectrum, and generating a sample frame signal spectrum diagram according to the obtained sample frame signal spectrum; A characteristic frequency range is set according to the obtained sample sampling frequency, a linear frequency is obtained according to the obtained characteristic frequency range, the obtained linear frequency is mapped to a sample frame signal spectrum diagram, a frequency index is obtained, and M triangular filters are constructed according to the obtained frequency index; Each frame signal in the frame signal spectrum is subjected to feature filtering processing through M triangular filters to obtain feature spectrum coefficients, and the obtained feature spectrum coefficients are feature sorted to obtain a sample feature signal spectrum.
[0010] The process of building a recurrent speech recognition model includes: Creating a sequence label corresponding to the sample audio signal according to the obtained sample characteristic signal spectrum, and combining the sequence labels into a sequence label set according to the obtained sample speech segment group; Setting a first loop condition according to the obtained sample segment group waveform and the number of sample speech segments; Marking the obtained sample characteristic signal spectrum as an input sequence, dividing the obtained input sequence by length to obtain input subsequences, and setting an initial hidden sequence according to the obtained input subsequences; Set up input gates, transfer gates, and output gates; Upload the input subsequence through the input gate, screen the obtained input subsequence to obtain input information, upload the obtained input information to the transfer gate, extract the input information in the transfer gate to obtain retained information, upload the obtained retained information to the output gate, and fuse the obtained retained information with the initial hidden sequence through the output gate to obtain the output sequence; The obtained output sequence is used as the initial hidden sequence of the next cycle, and the process of obtaining the output sequence is repeated until the first cycle condition is met, and the obtained output sequence is counted to obtain an output column set.
[0011] The process of obtaining interactive text data includes: Input the obtained user continuous audio signal into the feature filter group to obtain the user feature signal spectrum, upload the obtained user feature signal spectrum to the cyclic speech recognition model for cyclic recognition, and obtain the user output sequence and the user output column set; The obtained user output column set is divided into sets to obtain a training set and a validation set; Select a user output sequence in the training set as a pre-training sequence, upload the obtained pre-training sequence as the initial hidden sequence to the cyclic speech recognition model, perform cyclic recognition, and obtain a predicted output sequence; The process of obtaining the prediction output column is repeated, a training output set is constructed according to the obtained prediction output column, the obtained training output set is sequence matched with the validation set, and the prediction output column with successful sequence matching is marked as interactive text data.
[0012] Based on the above-mentioned artificial intelligence-based speech recognition system, the present invention also provides an artificial intelligence-based speech recognition method, comprising the following steps: Step 1: Collect user voice data and sample data; Step 2: Process the collected user voice data and sample data to obtain a user continuous audio signal and a sample continuous audio signal, and construct a feature filter group according to the sample continuous audio signal to obtain a sample feature signal spectrum and a user feature signal spectrum; Step 3: construct a cyclic speech recognition model based on the obtained sample characteristic signal spectrum; Step 4: Perform cyclic recognition on the user characteristic signal spectrum according to the obtained cyclic speech recognition model to obtain interactive text data.
[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: the collected user voice data and sample data are processed, and a sample continuous audio signal is obtained according to the sample audio signal, the sample sampling frequency and the sample voice segment; a user continuous audio signal is obtained according to the user audio signal, the user sampling frequency and the user voice segment; Perform fast Fourier transform on the sample continuous audio signal to obtain a frame signal spectrum, construct a feature filter group according to the obtained frame signal spectrum, perform feature filtering processing on each frame signal in the obtained frame signal spectrum through M triangular filters, and obtain a sample feature signal spectrum and a user feature signal spectrum; A cyclic speech recognition model is constructed based on the sample characteristic signal spectrum, and the user characteristic signal spectrum is cyclically recognized based on the cyclic speech recognition model to obtain interactive text data; the speech data is accurately converted into text data for easy storage, which greatly improves the accuracy and real-time performance of speech recognition and effectively reduces the error rate of manual dictation. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] 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.
[0015] Figure 1 It is a schematic diagram of the present invention. DETAILED DESCRIPTION
[0016] 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.
[0017] like Figure 1As shown, a speech recognition system based on artificial intelligence includes a control center, and the control center is connected to a data acquisition module, a data processing module, a data analysis module and a comprehensive recognition module; The data collection module is used to collect user voice data and sample data, and the specific process includes: Setting a collection port, wherein the collection port includes a plurality of voice capture points; The collection port is used to collect the user's voice data, and the voice data collected by the collection port is captured through the voice capture point to obtain the user's voice data, voice duration and sample data; It should be further explained that, in the specific implementation process, the acquisition port includes a microphone, a smart phone device, and a telephone communication device.
[0018] The data processing module is used to process the collected user voice data and sample data to obtain a user continuous audio signal and a sample continuous audio signal, and to construct a feature filter group according to the sample continuous audio signal to obtain a user feature signal spectrum and a sample feature signal spectrum. The specific processing process includes: Performing data segmentation on the obtained user voice data to obtain user audio signals, user sampling frequencies, and user voice segments; Performing data segmentation on the obtained sample data to obtain a sample audio signal, a sample sampling frequency, and a sample voice segment; It should be further explained that, in the specific implementation process, the voice segments include command segments, natural segments, and scene segments; The obtained sample sampling frequency is marked as fs; Matching the obtained sample audio signals with the sample speech segments, and associating the successfully matched sample audio signals with the sample speech segments; Matching the obtained sample speech segment with the sample speech duration to obtain the sample segment duration corresponding to the sample speech segment, wherein the sample segment duration includes the segment start time and the segment end time; Generate a sample audio waveform graph according to the obtained sample audio signal, and associate the sample audio waveform graph with a sample speech segment associated with the sample audio signal; Within the sample speech duration, combining a plurality of sample speech segments according to the time sequence of the sample segment durations of the sample speech segments to obtain a sample speech segment group; Obtaining a sample segment group waveform graph according to the sample audio waveform graph associated with the sample speech segment in the obtained sample speech segment group; The sample audio signal is framed according to the obtained sample sampling frequency to obtain a frame signal, the obtained frame signal is marked as i, where i=1, 2, 3, ..., v1, v1 is a positive integer, and the length of the obtained frame signal is marked as a frame length, recorded as ZLi; A signal window is set according to the obtained sample speech segment, and the obtained signal window is marked as j, where j=1, 2, 3, ..., v2, and v2 is a positive integer, and the length of the signal window is marked as the signal window length, which is recorded as CLj; It should be further explained that, in the specific implementation process, the signal window length of this embodiment is the same as the frame length, that is, ZLi=CLj; Setting a first sample point according to the obtained signal window length, marking the obtained first sample point as n, where n=1, 2, 3, ..., v3, and v3 is a positive integer; Obtain a first sample point value according to the obtained first sample point and the signal window length, and mark the obtained first sample point value as w(n), where w(n)=0.54-0.46*, and CLj≠1; Mark the obtained signal window length and frame length in the sample segment group waveform diagram, segment the waveform in the sample segment group waveform diagram according to the signal window length to obtain signal window segments, and mark the obtained signal window segments as XL; It should be further explained that, in the specific implementation process, in this embodiment, the signal window segment is the same as the frame length, that is, XL=CL; Obtain a first window sample value according to the obtained signal window segment and the first sample point value, and mark the obtained first window sample value as CY, where CY=w(n)*XL; The frame lengths of the waveform graphs of the sample segment groups containing the signal window segments are moved and overlapped so that there are overlapping parts between adjacent signal windows, thereby obtaining a sample continuous audio signal; It should be further explained that, in the specific implementation process, the coverage rate of the movement overlap must meet 50%-75% of the entire sample segment group waveform; Similarly, the process of obtaining the sample continuous audio signal is repeated according to the sample audio signal, the sample sampling frequency and the sample voice segment to obtain the user continuous audio signal; Performing fast Fourier transform on each frame signal in the sample continuous audio signal to obtain a sample frame signal spectrum, and generating a sample frame signal spectrum diagram according to the obtained sample frame signal spectrum; It should be further explained that, in a specific implementation process, the process of repeatedly obtaining the sample frame signal spectrum is repeated according to the obtained user continuous audio signal to obtain the user frame signal spectrum; A characteristic filter group is constructed according to the obtained sample frame signal spectrum, and the construction process of the characteristic filter group includes: The number of filters is set according to the obtained sample audio signal, and the obtained number of filters is marked as M; it should be further explained that, in the specific implementation process, the number of filters in this embodiment is M=26; The characteristic frequency is set according to the obtained sample sampling frequency, and the obtained characteristic frequency is marked as MP, and the characteristic frequency range corresponding to the characteristic frequency is marked as a set {MP}, where {MP}=[0, fs*α1*B], α1 is the range coefficient, and B represents the frequency range determined by the sample frame signal spectrum, and the frequency range is the difference between the maximum frequency and the minimum frequency in the sample frame signal spectrum; Obtain the linear frequency according to the obtained characteristic frequency range, and mark the obtained linear frequency as f, where f=700*(-1); Map the obtained linear frequency to the sample frame signal spectrum diagram to obtain a frequency index, and mark the obtained frequency index as k, where k=floor[(+1)*], and N represents the number of sample points in the frame signal spectrum; M triangular filters are constructed according to the obtained frequency indexes. It should be further explained that, in a specific implementation process, each filter is determined by three consecutive frequency indexes, and the three consecutive frequency indexes are marked as a left boundary, a center frequency, and a right boundary respectively; Obtain filter responses according to the set M triangular filters, and mark the obtained filter responses as H(k,M), where H(k,M)=max[0,min(,)], and is the frequency index of the center frequency of the M filter, is the frequency index of the left boundary of the M filter, is the frequency index of the right boundary of the M filter, and is the frequency index of the M filter; Each frame signal in the obtained sample frame signal spectrum is subjected to characteristic filtering processing by M triangular filters to obtain characteristic spectrum coefficients; It should be further explained that, in a specific implementation process, the characteristic filtering process means first filtering each sample frame signal in the obtained sample frame signal spectrum through M triangular filters to obtain a filtered frame signal, performing logarithmic value sampling on the filtered frame signal to obtain a characteristic spectrum, and performing inverse discrete cosine transform on the obtained characteristic spectrum to obtain a characteristic spectrum coefficient; The obtained characteristic spectrum coefficients are characteristically sorted, the characteristic spectrum coefficients in the first e positions are characteristically represented, a sample characteristic signal spectrum is obtained, and a sample characteristic signal spectrum graph is generated according to the obtained sample characteristic signal spectrum; It should be further explained that, in the specific implementation process, e in this embodiment is 12, and the first 12 characteristic spectrum coefficients are taken from each frame signal as the characteristic representation of the sample audio signal, and the sample speech segment associated with the sample audio signal is obtained, and the obtained sample characteristic signal spectrum graph is associated with the sample speech segment corresponding to the sample audio signal; Similarly, characteristic filtering is performed on the user frame signal spectrum according to the obtained characteristic filter group to obtain the user characteristic signal spectrum.
[0019] The data analysis module is used to construct a cyclic speech recognition model according to the obtained sample characteristic signal spectrum. The specific analysis process includes: Obtaining a sample speech segment associated with the sample characteristic signal spectrogram, and creating a sequence label corresponding to the sample audio signal according to the obtained sample speech segment; It should be further explained that, in the specific implementation process, creating a sequence label corresponding to the sample audio signal means to sequence the sample speech segments, and mark the sample speech segments as corresponding sequence labels in sequence. In particular, the sequence label represents the vocabulary label corresponding to each sample speech segment, and the vocabulary in the sample speech segment is screened to obtain the most representative vocabulary as the vocabulary label of a sample speech segment; The obtained sequence label is marked as CHp, where p represents the number of the sample speech segment, and p=1, 2, 3, ..., v4, v4 is a positive integer; The obtained sequence labels are combined into a sequence label set according to the order of the sample speech segments in the sample speech segment group, and the obtained sequence label set is marked as {CH}, where {CH}=[CH1, CH2, CH3, ..., CHp]; A cyclic speech recognition model is constructed according to the obtained sequence labels, sample characteristic signal spectra and sample characteristic signal spectrograms. The process of constructing the cyclic speech recognition model includes: The number of loop layers is set according to the obtained sample segment group waveform and the number of sample speech segments, and the obtained number of loop layers is recorded as the first loop condition; Marking the obtained sample characteristic signal spectrum as an input sequence, dividing the obtained input sequence by length to obtain input subsequences; Setting an initial hidden sequence according to the obtained input subsequence; Set an input gate, a transfer gate, and an output gate, upload the obtained input subsequence through the input gate, filter the obtained input subsequence, and obtain input information; It should be further explained that, in a specific implementation process, the information screening is to screen out information that can be passed to the next time point in the input subsequence, and the next time point represents the next time point of the current input subsequence, that is, the start time of the segment of the next input subsequence; The input information obtained is uploaded to the transfer gate, and the input information is extracted in the transfer gate to obtain the retained information; It should be further explained that, in the specific implementation process, the transfer gate is provided with a retention terminal, and the portion of the input information that needs to be retained is determined through the retention terminal; The obtained retained information is uploaded to the output gate, and the obtained retained information is fused with the initial hidden sequence through the output gate to obtain the output sequence, and the obtained output sequence is associated with the corresponding sequence label; It should be further explained that, in the specific implementation process, the information fusion process includes: At each time point, the input information and the retained information obtained by the input gate, the transfer gate and the output gate are sequentially added to obtain a pre-added vector, and the obtained pre-added vector is used as the output sequence; The obtained output sequence is used as the initial hidden sequence of the next cycle, and the process of obtaining the output sequence is repeated until the first cycle condition is met, and the obtained output sequence is counted to obtain an output column set; It should be further explained that, in a specific implementation process, the output column set includes all output sequences obtained through the loop.
[0020] The comprehensive recognition module is used to perform cyclic recognition on the user characteristic signal spectrum according to the constructed cyclic speech recognition model to obtain interactive text data. The specific process includes: Upload the obtained user characteristic signal spectrum to the cyclic speech recognition model to obtain the user output sequence and the user output column set; The obtained user output column set is divided into sets to obtain a training set and a validation set; It should be further explained that, in the specific implementation process, the set division is to divide the obtained user output sequence into types according to the corresponding sequence labels, and the type division is to mark the repeated sequence labels as the same sequence type; The user output sequences contained in the same sequence type are equally divided into two groups. In particular, if the two groups differ by one, they also meet the equal division requirements. They are uploaded to the training set and the validation set respectively. The remaining output sequences of the same type are divided into different types one by one, and the output sequences of the same type are uploaded to the training set and the output set respectively. Select a user output sequence in the training set as a pre-training sequence, upload the obtained pre-training sequence as the initial hidden sequence to the cyclic speech recognition model, perform cyclic recognition, obtain the user output sequence, and mark the obtained user output sequence as a predicted output column; Select one user output sequence from the remaining training set as the pre-training sequence, and repeat the process of obtaining the predicted output sequence until all user output sequences in the training set have completed the cyclic recognition; It should be further explained that, in the specific implementation process, all user output sequences in the training set complete the loop recognition and meet the first loop condition. If the loop recognition of all user output sequences is not completed within the first loop condition, a new loop recognition is started and the first loop condition is met. Building a training output set according to the obtained prediction output column, and uploading the obtained prediction output column to the training output set; Perform sequence matching on the predicted output columns in the obtained training output set and the user output sequences in the validation set; It should be further explained that, in the specific implementation process, the sequence matching is to compare the sequence labels corresponding to the predicted output columns in the training output set with the sequence labels associated with the user output sequences in the validation set. If there is overlap, the match is successful. The predicted output columns with successful sequence matching are marked as interactive text sequences until all predicted output columns in the training output set complete sequence matching to obtain the remaining interactive text sequences. All obtained interactive text sequences are sorted in chronological order according to the end time of the fragment to obtain interactive text data.
[0021] Based on the above-mentioned artificial intelligence-based speech recognition system, the present invention also provides an artificial intelligence-based speech recognition method, comprising the following steps: Step 1: Collect user’s voice data; Step 2: Process the collected speech data to obtain a continuous audio signal, and construct a characteristic filter group based on the continuous audio signal to obtain a characteristic signal spectrum; Step 3: construct a cyclic speech recognition model based on the obtained characteristic signal spectrum; Step 4: Perform cyclic recognition on the acquired speech data according to the acquired cyclic speech recognition model to obtain interactive text data.
[0022] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.
Claims
1. A speech recognition system based on artificial intelligence, comprising a control center, characterized in that: The control center is connected to a data acquisition module, a data processing module, a data analysis module and a comprehensive identification module; The data collection module is used to collect user voice data and sample data; The data processing module is used to process the collected user voice data and sample data to obtain a user continuous audio signal and a sample continuous audio signal, and to construct a feature filter group according to the sample continuous audio signal to obtain a sample feature signal spectrum and a user feature signal spectrum; The data analysis module is used to construct a cyclic speech recognition model according to the obtained sample characteristic signal spectrum; The comprehensive recognition module is used to perform cyclic recognition on the user characteristic signal spectrum according to the obtained cyclic speech recognition model to obtain interactive text data.
2. The artificial intelligence-based speech recognition system according to claim 1, characterized in that: The process of the data acquisition module collecting user voice data and sample data includes: Setting a collection port, wherein the collection port includes a plurality of voice capture points; The collection port is used to collect the user's voice data. The voice data collected by the collection port is captured through the voice capture point to obtain the user's voice data, sample data and voice duration.
3. The artificial intelligence-based speech recognition system according to claim 2, characterized in that: The process of the data processing module processing the user voice data and the sample data includes: Performing data segmentation on the obtained sample voice data to obtain a sample audio signal, a sample sampling frequency, and a sample voice segment, and performing data segmentation on the obtained user voice data to obtain a user audio signal, a user sampling frequency, and a user voice segment; The obtained sample audio signals are matched with the sample speech segments, and the successfully matched sample audio signals are associated with the sample speech segments; the sample speech segments are combined within the sample speech duration to obtain a sample speech segment group, and a sample segment group waveform is generated according to the sample audio signals associated with the sample speech segments in the sample speech segment group.
4. The artificial intelligence-based speech recognition system according to claim 3, characterized in that: The process of obtaining the sample continuous audio signal and the user continuous audio signal includes: Performing signal framing on the sample audio signal according to the obtained sample sampling frequency to obtain a frame signal, and marking the length of the obtained frame signal as the frame length; Setting a signal window according to the obtained sample speech segment, and marking the length of the obtained signal window as the signal window length; Mark the obtained signal window length and frame length in the sample segment group waveform diagram, segment the waveform in the sample segment group waveform diagram according to the signal window length to obtain the signal window segment, and move and overlap the frame length of the sample segment group waveform diagram containing the signal window segment to obtain a sample continuous audio signal; The process of obtaining the sample continuous audio signal is repeated according to the obtained user audio signal, the user sampling frequency and the user voice segment to obtain the user continuous audio signal.
5. The artificial intelligence-based speech recognition system according to claim 4, characterized in that: The process of constructing an eigenfilter bank includes: Performing fast Fourier transform on each frame signal in the sample continuous audio signal to obtain a sample frame signal spectrum, and generating a sample frame signal spectrum diagram according to the obtained sample frame signal spectrum; A characteristic frequency range is set according to the obtained sample sampling frequency, a linear frequency is obtained according to the obtained characteristic frequency range, the obtained linear frequency is mapped to a sample frame signal spectrum diagram, a frequency index is obtained, and M triangular filters are constructed according to the obtained frequency index; Each frame signal in the frame signal spectrum is subjected to feature filtering processing through M triangular filters to obtain feature spectrum coefficients, and the obtained feature spectrum coefficients are feature sorted to obtain a sample feature signal spectrum.
6. The artificial intelligence-based speech recognition system according to claim 5, characterized in that: The process of building a recurrent speech recognition model includes: Creating a sequence label corresponding to the sample audio signal according to the obtained sample characteristic signal spectrum, and combining the sequence labels into a sequence label set according to the obtained sample speech segment group; Setting a first loop condition according to the obtained sample segment group waveform and the number of sample speech segments; Marking the obtained sample characteristic signal spectrum as an input sequence, dividing the obtained input sequence by length to obtain input subsequences, and setting an initial hidden sequence according to the obtained input subsequences; Set up input gates, transfer gates, and output gates; Upload the input subsequence through the input gate, screen the obtained input subsequence to obtain input information, upload the obtained input information to the transfer gate, extract the input information in the transfer gate to obtain retained information, upload the obtained retained information to the output gate, and fuse the obtained retained information with the initial hidden sequence through the output gate to obtain the output sequence; The obtained output sequence is used as the initial hidden sequence of the next cycle, and the process of obtaining the output sequence is repeated until the first cycle condition is met, and the obtained output sequence is counted to obtain an output column set.
7. The artificial intelligence-based speech recognition system according to claim 6, characterized in that: The process of obtaining interactive text data includes: Input the obtained user continuous audio signal into the feature filter group to obtain the user feature signal spectrum, upload the obtained user feature signal spectrum to the cyclic speech recognition model for cyclic recognition, and obtain the user output sequence and the user output column set; The obtained user output column set is divided into sets to obtain a training set and a validation set; Select a user output sequence in the training set as a pre-training sequence, upload the obtained pre-training sequence as the initial hidden sequence to the cyclic speech recognition model, perform cyclic recognition, and obtain a predicted output sequence; The process of obtaining the prediction output column is repeated, a training output set is constructed according to the obtained prediction output column, the obtained training output set is sequence matched with the validation set, and the prediction output column with successful sequence matching is marked as interactive text data.
8. A speech recognition method based on an artificial intelligence speech recognition system according to any one of claims 1 to 7, characterized in that: The following steps are involved: Step 1: Collect user voice data and sample data; Step 2: Process the collected user voice data and sample data to obtain a user continuous audio signal and a sample continuous audio signal, and construct a feature filter group according to the sample continuous audio signal to obtain a sample feature signal spectrum and a user feature signal spectrum; Step 3: construct a cyclic speech recognition model based on the obtained sample feature signal spectrum; Step 4: Perform cyclic recognition on the user characteristic signal spectrum according to the obtained cyclic speech recognition model to obtain interactive text data.