Multi-speaker speech recognition method and system suitable for end-side equipment

Abstract

The invention discloses a multi-speaker speech recognition method and system suitable for an end side device, and the method comprises the steps: collecting multi-speaker audio data, decomposing the multi-speaker audio into two single-speaker audios through a sound channel separation method, carrying out the activity detection of each single-speaker audio, and carrying out the recognition of the single-speaker audios. Segmenting each single speaker audio into a plurality of effective sound segments; performing overlapping judgment on each effective voice segment on the time sequence, performing voice correction on the overlapped effective voice segments, and outputting a confidence coefficient optimization result of character transcription of each effective voice segment; aligning the text transcription result of the audio of each speaker on the time sequence, and optimally calculating the merging confidence value of the text transcription of each effective tone segment after the time sequence alignment; and according to the merging confidence value, carrying out text content splicing on the time sequence on the text transcription result of each effective tone segment, and calculating the optimal time offset in the splicing process of the transcribed text content of each single speaker by adopting a least square method.

Description

Technical Field

[0001] This invention relates to the field of speech recognition technology, and in particular to a multi-speaker speech recognition method and system suitable for edge devices. Background Technology

[0002] Currently, speech recognition scenarios on existing edge devices commonly involve multiple speakers. However, traditional speech recognition technologies are generally designed for single speakers. Multi-speaker scenarios present the following technical challenges: 1. Speech from multiple speakers is prone to overlap, and single-speaker speech recognition technologies face difficulties in separating overlapping parts; 2. Multi-speaker scenarios are often affected by silence and noise, resulting in lower accuracy in recognizing the target speaker, higher computational complexity in dialogue recognition, and incomplete recording due to dialogue fragmentation; 3. Traditional edge-side single-speaker speech recognition technologies are typically processed serially, leading to poor real-time performance in multi-speaker speech recognition on edge devices. Summary of the Invention

[0003] One objective of this invention is to provide a multi-speaker speech recognition method and system suitable for end-side devices. The method and system utilize dual-channel signal separation technology to convert multi-speaker signals into two independent single-speaker signals, and employ normalized resampling technology to resample the two independent single-speaker signals. This effectively reduces interference caused by speech overlap in multi-speaker signals, while also reducing the computational complexity of speech recognition, significantly reducing speech ambiguity recognition, and improving the accuracy of multi-speaker speech recognition.

[0004] Another objective of this invention is to provide a multi-speaker speech recognition method and system suitable for end-side devices. The method and system identify the activity boundaries of different vocal tracts through speech activity segmentation technology, and then perform speech segmentation based on the speech activity boundaries. This accurately segments different effective speech segments for each vocal tract, enabling the invention to remove noise and silence from each speaker's speech segments, reducing invalid calculations for noise and silence segments, and significantly improving the accuracy of speech recognition and computational inference efficiency for each speaker.

[0005] Another objective of this invention is to provide a multi-speaker speech recognition method and system suitable for end-side devices. After segmenting the effective audio segments of each channel, the method and system further perform overlap calculations on the temporal intersection-union ratio (IUU) of two effective audio segments, calculate the confidence level of the overlapping effective audio segments, and calibrate the overlapping effective audio segments. In subsequent single-speaker continuous effective audio segments on the time axis, a merging confidence optimization calculation is performed. Furthermore, based on the result of the merging confidence optimization calculation, the text recognition results of the continuous effective audio segments of each single speaker are spliced ​​and optimized. Thus, this invention allows single speakers to maintain continuity and consistency in time and semantics, reducing the occurrence of speech breaks.

[0006] To achieve at least one of the above-mentioned objectives, the present invention further provides a multi-speaker speech recognition method suitable for end-side devices, the method comprising:

[0007] Collect multi-speaker audio data, and decompose the multi-speaker audio into two single-speaker audios using a channel separation method. Perform activity detection on each single-speaker audio and divide each single-speaker audio into multiple effective audio segments.

[0008] Overlap detection is performed on each effective segment in the time series, and speech correction is performed on the overlapping effective segments. The confidence optimization result of the text transcription of each effective segment is output.

[0009] The transcripts of each speaker's audio are aligned in time series, and the combined confidence of the transcripts of each effective audio segment after time series alignment is optimized.

[0010] Based on the combined confidence value, the text transcription results of each effective audio segment are spliced ​​together in time series, and the optimal time offset is calculated in the splicing process of the text transcription of each single speaker using the least squares method.

[0011] According to a preferred embodiment of the present invention, the effective audio segment acquisition method includes: acquiring stereo audio data of multiple speakers, and using a separation tool to physically separate the stereo audio data of multiple speakers into audio data of two channels, wherein each channel corresponds to independent audio data of a single speaker, and performing a normalized resampling operation on each audio data of a single speaker to obtain audio data of a single speaker with consistent time-frequency resolution.

[0012] According to another preferred embodiment of the present invention, the normalized resampling formula includes:

[0013] ;

[0014] Where w mFor the normalized weight kernel, satisfying ∑w m =1, M represents the length of the interpolation window; x is the original sampling sequence; f src with f tgt These represent the original and target sampling rates, respectively; k is the target sampling rate index. This represents the normalized resampling result of the k-th target. After the normalized resampling is completed, a single speaker audio file with independent corresponding vocal channels is further obtained.

[0015] According to another preferred embodiment of the present invention, the effective audio segment acquisition method includes: after acquiring a normalized resampled single speaker audio file, converting the normalized resampled single speaker audio file to Float format, and performing activity detection on the format-converted single speaker audio file, wherein the activity detection includes calculating the energy values ​​of different speech segments in the single speaker audio file, calculating the noise and noise standard deviation of each speech segment to obtain an energy threshold, and segmenting the single speaker audio file according to the energy threshold and the energy values ​​detected by each speech segment to obtain effective audio segments after noise removal and silence removal.

[0016] According to another preferred embodiment of the present invention, after obtaining the effective audio segments of two voice channels corresponding to a single speaker, the overlap of the effective audio segments of different single speakers is judged. The overlap degree of the effective audio segments of different single speakers in the time series is calculated by the time intersection-overlap ratio. When the overlap degree is greater than a preset threshold, the corresponding overlapping speech is corrected. In the effective audio segments corresponding to the overlap correction, the transcription confidence result of speech to text is output using the following formula:

[0017] ;

[0018] in Let Δ be the ASR confidence value for the k-th valid segment. λ represents the interval between the k-th valid segment and the current time, and λ2 represents the time decay parameter. This indicates that the k-th valid segment found is the one with the highest confidence value. This is the result of the final transcription.

[0019] According to another preferred embodiment of the present invention, after the identification of the effective sound segments is completed, the overall complexity of all effective sound segments for each single speaker is further calculated, and the formula for calculating the overall complexity is as follows:

[0020] ;

[0021] Where M represents the total number of valid segments for a single speaker. Let F represent the number of frames corresponding to the i-th valid audio segment of a single speaker, F be the computational cost per frame, and O() denote the algorithm complexity function. In the above formula, the algorithm complexity and... Proportional.

[0022] According to another preferred embodiment of the present invention, the timestamps of the effective audio segments of each individual speaker are obtained, and after transcribing the effective audio segments of each individual speaker, the combined confidence of each speaker on consecutive effective audio segments in the same timeline is optimized and calculated, wherein the optimized calculation formula for the combined confidence is:

[0023] ;

[0024] ;

[0025] in This indicates the confidence level of the effective segments after merging, where n represents the number of consecutive effective segments from the same speaker to be merged. This represents the ASR confidence level of the i-th original valid segment. The nonlinear weighted index is represented by S; the acceleration ratio is represented by S. This indicates the processing time for the left channel. This indicates the processing time for the right channel. This represents the synchronization efficiency coefficient. Indicates the time taken for timeline synchronization. This indicates the time spent on merging additional overhead.

[0026] According to another preferred embodiment of the present invention, the method includes: calibrating the time offset for each valid segment of the transcription, and calculating the optimal time offset for each valid segment using the following least squares method:

[0027] ;

[0028] ;

[0029] in This represents the timestamp of the i-th valid audio segment identified. Indicates a reference timestamp. δ* represents the time offset to be optimized, n represents the time offset to be corrected, and n represents the total number of timestamp samples of valid audio segments. Indicates the corresponding time offset The formula is at its minimum when it is optimal.

[0030] To achieve at least one of the above-mentioned objectives, the present invention further provides a multi-speaker speech recognition system suitable for end-device devices, the system executing the aforementioned multi-speaker speech recognition method suitable for end-device devices. The present invention further provides a computer-readable storage medium storing a computer program, the computer program being executed by a processor to implement the aforementioned multi-speaker speech recognition method suitable for end-device devices. Attached Figure Description

[0031] Figure 1 The diagram shown is a flowchart of a multi-speaker speech recognition method applicable to end-side devices according to the present invention. Detailed Implementation

[0032] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0033] It is understood that the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0034] Please combine Figure 1 This invention discloses a multi-speaker speech recognition method and system suitable for end-side devices, wherein the method mainly includes:

[0035] S01. Collect multi-speaker audio data, and decompose the multi-speaker audio into two single-speaker audios using a channel separation method. Perform activity detection on each single-speaker audio and divide each single-speaker audio into multiple effective audio segments.

[0036] S02. For each effective segment in the time series, overlap is determined, and speech correction is performed on the overlapping effective segments. The confidence optimization result of the text transcription of each effective segment is output.

[0037] S03. Align the text transcription results of each speaker's audio on the time series, and optimize the calculation of the combined confidence value of the text transcription of each effective audio segment after time series alignment.

[0038] S04. Based on the merged confidence value, the text content of each valid phonetic transcription result is spliced ​​together in time series, and the optimal time offset is calculated in the splicing process of each single speaker's transcribed text content using the least squares method.

[0039] Specifically, this invention includes a preprocessing separation layer. The multi-speaker voice input is fed into this layer for physical separation. This invention employs dual-channel separation technology, acquiring stereo audio data from multiple speakers. An audio separation tool physically separates the stereo audio data into left and right channel audio data. The left channel audio data represents the audio data of one independent single speaker, and the right channel represents the audio data of another independent single speaker. When there are three or more speakers, the required single-speaker audio data in the left and right channels can be selected through voiceprint registration and recognition. The audio separation tool used in this invention can be, but is not limited to, FFmpeg. The single-speaker voiceprint registration and recognition can employ existing technologies, which will not be elaborated upon in detail here.

[0040] In this invention, a separation tool is used to divide the audio of multiple speakers into left and right channels, wherein the left channel audio signal is:

[0041] ;

[0042] Where L(t) and R(t) represent the left and right channel audio signals, respectively.

[0043] The aforementioned dual-channel separation technology can convert multi-speaker speech recognition tasks into two independent single-speaker speech recognition tasks, effectively reducing environmental interference.

[0044] Furthermore, to ensure data consistency during the model's data processing phase, this invention performs the following normalization resampling operation on the single-speaker audio data of the two channels:

[0045] ;

[0046] Where w m For the normalized weight kernel, satisfying ∑w m =1, M represents the length of the interpolation window; x is the original sampling sequence; f src with f tgt These represent the original and target sampling rates, respectively; k is the target sampling rate index. This represents the normalized resampling result for the k-th target. The above normalized resampling result ensures that the model input data has a consistent time-frequency resolution. Combining this with the technique of separating the vocal tract into single speakers can reduce interference caused by speech overlap and significantly reduce the computational complexity and recognition ambiguity in the subsequent recognition process, laying the foundation for accurate recognition.

[0047] The present invention further provides a core recognition layer, wherein the core recognition layer includes: speech energy recognition, noise removal, and silence segment removal, and the speech activity detection technology is used to segment single speaker speech data into effective segments, the specific method including the following steps:

[0048] First, the core model of the speech recognition layer is initialized, and multiple frames of speech segments are sampled from the single speaker's speech data. The energy of each sampled speech segment is calculated using the following formula:

[0049] ;

[0050] Where E is the average energy;

[0051] x[n] represents the nth sampling point, and N represents the frame length;

[0052] Further adjust the adaptive ambient noise threshold θ, where θ is calculated using the following formula:

[0053] ;

[0054] Where μnoise and σnoise are the mean and standard deviation of the noise energy, respectively. This is the adjustment coefficient;

[0055] When E>θ, the segment is considered a valid segment.

[0056] Through the above calculations, the present invention can accurately determine the valid speech segments, providing clean input for the subsequent recognition module.

[0057] By combining an ASR (Automatic Speech Recognition) model, this invention simultaneously acquires and processes Float format audio data and its corresponding sampling rate, providing a standardized data foundation that meets the model's input requirements for subsequent speech recognition. The system is based on the FunASR engine for speech recognition, which supports adaptive loading of multiple audio formats and can accurately parse audio information based on the acquired Float audio data and sampling rate, ensuring that audio from different sources and in different formats can be processed correctly.

[0058] Before the recognition process officially begins, the system uses speech activity detection technology to perform speech activity detection and segmentation on the single-speaker Float format audio data of the left and right channels respectively. Combined with audio sampling rate calibration, the detection accuracy is calibrated. By identifying the boundaries of speech activity, the system accurately segments out valid speech segments while filtering out invalid segments such as silence and noise. In this process, the high precision characteristics and accurate sampling rate information of Float audio data can more accurately determine the boundary between speech and non-speech, greatly reducing invalid calculations and improving overall reasoning efficiency. The output at this stage includes information containing Float format audio data and corresponding timestamps.

[0059] After segmentation, the effective Float audio segments of the left and right channels (with sampling rate and timestamp information) are respectively entered into the parallel ASR transcription processing submodule (corresponding to speaker ID1 and speaker ID2). Based on standardized Float audio data and accurate sampling rate, the ASR model completes the accurate transcription task from speech to text, and finally outputs the transcription result containing time, text content and speaker ID.

[0060] Furthermore, the invention employs a temporal intersection-union ratio (TIU) to calculate the degree of overlap of effective speech segments from different single speakers, and performs speech correction based on the degree of overlap. The temporal intersection-union ratio... The calculation formula is:

[0061] ;

[0062] Where i and j represent the numbers of the two valid segments, Let i and j represent the end times of valid segment i and j, respectively. These represent the start time of valid segment i and the start time of valid segment j, respectively. This represents a regularity constant to prevent division by zero. ()and ( ) represent taking the maximum and minimum values ​​respectively. When The system determines overlap by τ (where τ is a threshold) and triggers overlapping speech correction. Between the valid overlapping segments, the system determines the final output based on the recognition confidence level.

[0063] ;

[0064] in Let Δ be the ASR confidence value for the k-th valid segment. λ represents the interval between the k-th valid segment and the current time, and λ2 represents the time decay parameter. This indicates that the k-th valid segment found is the one with the highest confidence value. This is the result of the final transcription.

[0065] The aforementioned two-channel contrastive transcription computational processing not only significantly shortens inference latency, but also maintains the independence and clarity of multi-speaker speech content by relying on high-quality audio data, avoiding information confusion caused by single-path processing.

[0066] It is worth mentioning that, after the effective phonological segments are identified in this invention, the overall complexity of all effective phonological segments for each single speaker is further calculated. The formula for calculating the overall complexity is as follows:

[0067] ;

[0068] Where M represents the total number of valid segments for a single speaker. Let F represent the number of frames corresponding to the i-th valid audio segment of a single speaker, F be the computational cost per frame, and O() denote the algorithm complexity function. In the above formula, the algorithm complexity and... Proportional.

[0069] This invention further performs confidence calculation for multi-segment transcription merging. The method includes: obtaining the timestamp of the effective audio segments of each individual speaker; transcribing the effective audio segments of each individual speaker; and optimizing the calculation of the merging confidence of each speaker in consecutive effective audio segments according to the same timeline. The optimized calculation formula for the merging confidence is as follows:

[0070] ;

[0071] ;

[0072] in This indicates the confidence level of the effective segments after merging, where n represents the number of consecutive effective segments from the same speaker to be merged. This represents the ASR confidence level of the i-th original valid segment. The nonlinear weighted index is represented by S; the acceleration ratio is represented by S. This indicates the processing time for the left channel. This indicates the processing time for the right channel. This represents the synchronization efficiency coefficient. Indicates the time taken for timeline synchronization. This indicates the time spent on merging additional overhead.

[0073] In one preferred embodiment of the present invention, in order to reduce the probability of speech interruption, the present invention calibrates the time offset of different vocal tract speech according to the time axis. The specific method includes: calibrating the time offset of each effective segment of the transcription, and calculating the optimal time offset of each effective segment using the following least squares method:

[0074] ;

[0075] ;

[0076] in This represents the timestamp of the i-th valid audio segment identified. Indicates a reference timestamp. δ* represents the time offset to be optimized, n represents the time offset to be corrected, and n represents the total number of timestamp samples of valid audio segments. Indicates the corresponding time offset The formula is at its minimum when it is optimal.

[0077] The processes described in the flowcharts above, as disclosed in the embodiments of this invention, can be implemented as computer software programs. Embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), the methods of this application are not limited to the aforementioned functions. It should be noted that the computer-readable medium described above can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. Computer-readable storage media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wire segments, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless segments, wire segments, optical fibers, RF, etc., or any suitable combination thereof.

[0078] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0079] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments. Without departing from the stated principles, the implementation of the present invention may have any variations or modifications.

Claims

1. A multi-speaker speech recognition method suitable for end-side devices, characterized in that, The method includes: Collect multi-speaker audio data, and decompose the multi-speaker audio into two single-speaker audios using a channel separation method. Perform activity detection on each single-speaker audio and divide each single-speaker audio into multiple effective audio segments. Overlap detection is performed on each effective segment in the time series, and speech correction is performed on the overlapping effective segments. The optimized result of the confidence score of the text transcription of each effective segment is output. The transcripts of each speaker's audio are aligned on a time series, and the pooled confidence value of each effective segment transcript after time series alignment is optimized. Based on the combined confidence value, the text content of each valid phonetic transcription result is spliced ​​together in time series, and the optimal time offset in the splicing process of each single speaker's transcribed text content is calculated using the least squares method.

2. The multi-speaker speech recognition method applicable to end-side devices according to claim 1, characterized in that, The effective audio segment acquisition method includes: acquiring stereo audio data of multiple speakers, and using a separation tool to physically separate the stereo audio data of multiple speakers into audio data of two channels, wherein each channel corresponds to independent audio data of a single speaker, and performing a normalized resampling operation on each audio data of a single speaker to obtain audio data of a single speaker with consistent time-frequency resolution.

3. The multi-speaker speech recognition method applicable to end-side devices according to claim 2, characterized in that, The normalized resampling formula includes: ； Where w m For the normalized weight kernel, satisfying ∑w m =1, M represents the length of the interpolation window; x is the original sampling sequence; f src with f tgt These represent the original and target sampling rates, respectively; k is the target sampling rate index. This represents the normalized resampling result of the k-th target. After the normalized resampling is completed, a single speaker audio file with independent corresponding vocal channels is further obtained.

4. The multi-speaker speech recognition method applicable to end-side devices according to claim 3, characterized in that, The effective audio segment acquisition method includes: acquiring a normalized resampled single speaker audio file, converting the normalized resampled single speaker audio file to Float format, and performing activity detection on the converted single speaker audio file. The activity detection includes calculating the energy values ​​of different speech segments in the single speaker audio file, calculating the noise and noise standard deviation of each speech segment to obtain an energy threshold, and segmenting the single speaker audio file according to the energy threshold and the energy values ​​detected by each speech segment to obtain effective audio segments after noise removal and silence removal.

5. A multi-speaker speech recognition method suitable for end-side devices according to claim 1, characterized in that, After obtaining the effective audio segments corresponding to two speakers in each audio channel, overlap is determined for the effective audio segments of different speakers. The overlap degree of the effective audio segments of different speakers in the time series is calculated using the time intersection-overlap ratio (TIGOR). When the overlap degree is greater than a preset threshold, correction is performed on the corresponding overlapping speech. In the effective audio segments corresponding to the overlap correction, the transcription confidence result of speech to text is output using the following formula: ； in Let Δ be the ASR confidence value for the k-th valid segment. λ represents the interval between the k-th valid segment and the current time, and λ2 represents the time decay parameter. This indicates that the k-th valid segment found is the one with the highest confidence value. This is the result of the final transcription.

6. A multi-speaker speech recognition method suitable for end-side devices according to claim 1, characterized in that, After identifying the effective sound segments, the overall complexity of all effective sound segments for each single speaker is further calculated. The formula for calculating the overall complexity is as follows: ； Where M represents the total number of valid segments for a single speaker. Let F represent the number of frames corresponding to the i-th valid audio segment of a single speaker, F be the computational cost per frame, and O() denote the algorithm complexity function. In the above formula, the algorithm complexity and... Proportional.

7. A multi-speaker speech recognition method applicable to end-side devices according to claim 1, characterized in that, The timestamps of the effective audio segments for each individual speaker are obtained. After transcribing the effective audio segments for each individual speaker, the combined confidence score of each speaker on consecutive effective audio segments in the same timeline is optimized and calculated. The optimized calculation formula for the combined confidence score is as follows: ； ； in This indicates the confidence level of the effective segments after merging, where n represents the number of consecutive effective segments from the same speaker to be merged. This represents the ASR confidence level of the i-th original valid segment. The nonlinear weighted index is represented by S; the acceleration ratio is represented by S. This indicates the processing time for the left channel. This indicates the processing time for the right channel. This represents the synchronization efficiency coefficient. Indicates the time taken for timeline synchronization. This indicates the time spent on merging additional overhead.

8. A multi-speaker speech recognition method suitable for end-side devices according to claim 1, characterized in that, The method includes: calibrating the time offset for each valid segment of the transcription, and calculating the optimal time offset for each valid segment using the following least squares method: ; ; in This represents the timestamp of the i-th valid audio segment identified. Indicates a reference timestamp. δ* represents the time offset to be optimized, n represents the time offset to be corrected, and n represents the total number of timestamp samples of valid audio segments. Indicates the corresponding time offset The formula is at its minimum when it is optimal.

9. A multi-speaker speech recognition system suitable for end-side devices, characterized in that, The system performs a multi-speaker speech recognition method applicable to end-side devices as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is executed by a processor to implement a multi-speaker speech recognition method suitable for end-side devices as described in any one of claims 1-8.

Images