Multi-speech recognition method and system based on edge algorithm

Abstract

The invention discloses a multi-speech recognition method and system based on an edge algorithm, and relates to the technical field of speech processing, and the method comprises the steps: obtaining a speech signal through a sound pickup, and transmitting the speech signal to edge node equipment for speech recognition; extracting voice segments of which the recognition rates are lower than a recognition rate threshold in the voice signals, and extracting composite voiceprint features; recognizing environment features generated by the voice segments through a voice recognition model, and automatically adjusting a voice recognition mode; in a target recognition mode, the composite voiceprint features are recognized through a voice recognition model, multiple voice frames are output, and at least one target voice frame is determined; and outputting a control instruction according to the target voice frame. According to the method, intelligent distribution of the edge nodes is realized by setting the recognition rate threshold value, the accuracy of the control instruction sent by the user is obviously improved in cooperation with composite voiceprint feature compression transmission and cloud environment self-adaptive recognition, and the problem that the response speed and the complex environment adaptability cannot be achieved at the same time in edge voice recognition is solved.

Description

Technical Field

[0001] The present invention relates to the field of speech processing technology, and in particular to a multi-speech recognition method and system based on edge algorithms. Background Art

[0002] With the popularization of IoT devices, voice recognition needs to meet the dual needs of real-time response, namely edge computing, and high-precision recognition, namely cloud computing.

[0003] However, the recognition rate of local lightweight models plummets in complex scenarios, such as those involving noise, multi-person conversations, and specialized terminology. Furthermore, full cloud-based interaction faces three major bottlenecks: high latency, privacy breaches, and bandwidth waste. Consequently, existing edge speech recognition technologies still face reliability bottlenecks in complex scenarios, and struggle to balance responsiveness with adaptability to complex environments. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a variety of speech recognition methods and systems based on edge algorithms, aiming to solve the above-mentioned problems recorded in the prior art.

[0005] A first aspect of the present invention is to provide a plurality of speech recognition methods based on edge algorithms, the methods comprising: Acquire the voice signal output by the user through the microphone, transmit the voice signal to the edge node device for voice recognition, and output the recognition rate of the voice signal; Extracting a speech segment whose recognition rate is lower than a preset recognition rate threshold from the speech signal, and extracting a composite voiceprint feature contained in the speech segment; Transmitting the voice segment and the composite voiceprint feature to a cloud server, and calling a speech recognition model in the cloud server; Identifying environmental features of the speech segment using the speech recognition model, and automatically adjusting a pre-set speech recognition mode according to the environmental features; In target recognition mode, the composite voiceprint feature is recognized by the voice recognition model, multiple voice frames with different voiceprint features are output, and at least one target voice frame is determined; A control instruction for controlling an electronic device is output according to the target voice frame.

[0006] According to one aspect of the above technical solution, the step of identifying the environmental features of the speech segment by the speech recognition model and automatically adjusting the pre-set speech recognition mode according to the environmental features includes: According to the speech signal, the speech segment is split using the speech recognition model to obtain target speech and environmental noise; Classifying the environmental noise, and determining the environmental characteristics of the speech segment based on the environmental noise having a confidence level greater than a preset confidence threshold; According to the environmental characteristics, the pre-set speech recognition mode is automatically adjusted through the speech recognition model to obtain a target recognition mode.

[0007] According to one aspect of the above technical solution, the step of automatically adjusting a preset speech recognition mode by the speech recognition model according to the environmental characteristics to obtain a target recognition mode includes: Automatically matching feature identifiers corresponding to the environmental features according to the environmental features; According to the feature identifier corresponding to the environmental feature, the preset feature identifier of the preset speech recognition mode is adjusted through the speech recognition model to obtain the target recognition mode.

[0008] According to one aspect of the above technical solution, in target recognition mode, the steps of identifying the composite voiceprint features by the speech recognition model, outputting multiple voice frames with different voiceprint features, and determining at least one target voice frame include: In the target recognition mode, the composite voiceprint feature is recognized by the speech recognition model according to the feature identifier of the environmental feature; Splitting the composite voiceprint feature into multiple voiceprint features with different characteristics, and converting them into multiple speech frames accordingly; Performing round-robin matching of voiceprint features on the plurality of speech frames to determine at least one target speech frame.

[0009] According to one aspect of the above technical solution, the step of performing round-robin matching of voiceprint features on a plurality of speech frames to determine at least one target speech frame includes: Load the feature vector corresponding to the registered user into the preset voiceprint database. When the similarity between any voice frame and the registered user feature vector is greater than the preset similarity threshold, it is marked as a candidate voice frame to obtain a candidate queue; Converting the candidate speech frames in the candidate queue into a text sequence, setting corresponding weight coefficients of the semantics and grammatical structures of keywords in the text sequence, and calculating the instruction relevance score of the text sequence; According to the predetermined environmental feature dynamic matching interference coefficient, the highest scoring frame in the candidate queue is selected as the target speech frame.

[0010] According to one aspect of the above technical solution, the calculation expression for calculating the instruction relevance score of the text sequence is: S=α×W k +β×W p ; Where S is the instruction relevance score, α∈[0.5,0.7] is the keyword semantic weight coefficient, β∈[0.3,0.5] is the grammatical structure weight coefficient, α+β=1, W k is the weight value of keyword semantics, W p is the weight value of the grammatical structure.

[0011] A second aspect of the present invention is to provide a multi-speech recognition system based on an edge algorithm, which is applied to the method described in the above technical solution, and the system includes: A voice transmission module is used to obtain the voice signal output by the user through the microphone, transmit the voice signal to the edge node device for voice recognition, and output the recognition rate of the voice signal; A speech extraction module, configured to extract speech segments whose recognition rates are lower than a preset recognition rate threshold from the speech signal, and extract composite voiceprint features contained in the speech segments; A model calling module, configured to transmit the voice segment and the composite voiceprint feature to a cloud server, and call a speech recognition model in the cloud server; a mode adjustment module, configured to identify, by means of the speech recognition model, environmental features of the speech segment, and automatically adjust a pre-set speech recognition mode according to the environmental features; a voiceprint recognition module configured to, in a target recognition mode, identify the composite voiceprint feature using the speech recognition model, output a plurality of speech frames having different voiceprint features, and determine at least one target speech frame; The instruction generation module is used to output a control instruction for controlling the electronic device according to the target voice frame.

[0012] According to one aspect of the above technical solution, the voiceprint recognition module is specifically used to: In the target recognition mode, the composite voiceprint feature is recognized by the speech recognition model according to the feature identifier of the environmental feature; Splitting the composite voiceprint feature into multiple voiceprint features with different characteristics, and converting them into multiple speech frames accordingly; Performing round-robin matching of voiceprint features on the plurality of speech frames to determine at least one target speech frame.

[0013] The third aspect of the present invention is to provide a readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the method described in the above technical solution.

[0014] The fourth aspect of the present invention is to provide an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the above technical solution when executing the program.

[0015] Compared with the prior art, the various speech recognition methods and systems based on edge algorithms shown in the present invention have the following beneficial effects: The present invention obtains a user's voice signal through a microphone, transmits the voice signal to an edge node device for voice recognition, outputs the recognition rate of the voice signal, then extracts voice segments with a recognition rate lower than a preset recognition rate threshold from the voice signal, extracts the composite voiceprint features contained in the voice segments, and transmits the voice segments and the composite voiceprint features to a cloud server. The cloud server calls a voice recognition model, uses the voice recognition model to identify the environmental features generated by the voice segments, and automatically adjusts the pre-set voice recognition mode according to the environmental features. Finally, in the target recognition mode, the voice recognition model recognizes the composite voiceprint features, outputs multiple voice frames with different voiceprint features, determines at least one target voice frame, and outputs control instructions for controlling electronic devices according to the target voice frame. The present invention realizes intelligent traffic diversion at the edge node by setting a recognition rate threshold, and combines the compressed transmission of the composite voiceprint features with cloud environment adaptive recognition to significantly improve the accuracy of the control instructions issued by the user, solving the problem of incompatibility between response speed and adaptability to complex environments in edge voice recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which: Figure 1 Schematic diagram of the flow of multiple speech recognition methods based on edge algorithms in one embodiment of the present invention; Figure 2 This is a structural block diagram of a multi-speech recognition system based on edge algorithms in one embodiment of the present invention. DETAILED DESCRIPTION

[0017] To make the objectives, features, and advantages of the present invention more readily apparent, the following detailed description of specific embodiments of the present invention is provided in conjunction with the accompanying drawings. The accompanying drawings illustrate several embodiments of the present invention. However, the present invention may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the present invention.

[0018] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly attached to the other element or there may be an intermediate element. When an element is referred to as being "connected to" another element, it may be directly connected to the other element or there may be an intermediate element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one skilled in the art to which this invention pertains. The terms used in this specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0020] Example 1 See also Figure 1 The first embodiment of the present invention provides a multi-speech recognition method based on edge algorithms, the method comprising steps S10 to S60: Step S10: Acquire a voice signal output by the user through a microphone, transmit the voice signal to an edge node device for voice recognition, and output a recognition rate of the voice signal.

[0021] In this embodiment, the voice signal output by the user is first obtained through the microphone. The user can be a registered user or a user other than a registered user. After the voice signal is picked up by the microphone, in order to respond, the voice signal is transmitted to the edge node device, and the voice signal is recognized by the edge node device.

[0022] During this process, the recognition rate of the edge node device for the voice signal is clarified. For example, the voice signal with a recognition accuracy greater than 90% is directly responded to through the edge node device, and the corresponding control instructions are output, thereby reducing the data transmission pressure and background processing pressure; when the recognition rate is less than 90%, further judgment such as steps S20-S60 is required.

[0023] Specifically, the recognition rate threshold is set to 90%. When the recognition rate is higher than 90%, all voice signals will be handed over to the edge node device for response. When the recognition rate is lower than 90%, it means that the recognition accuracy is average, and there may be environmental interference sounds, multiple voices, etc., then all or at least part of the voice signal will be transmitted to the cloud server for response through the node edge device. For example, the corresponding model can be called in the background to overcome the corresponding problem, thereby improving the quality of voice interaction between the user and the voice system.

[0024] Step S20: extracting speech segments whose recognition rates are lower than a preset recognition rate threshold from the speech signal, and extracting composite voiceprint features contained in the speech segments.

[0025] In this embodiment, the edge node device extracts voice segments from the voice signal whose recognition rate is lower than the preset recognition rate threshold, that is, extracts part of the voice segment from the voice signal. For example, when a registered user interacts with the in-vehicle voice system through voice, the microphone also picks up some noise. The edge node device has a low recognition rate for the voice segment and is unable to respond. In this case, it is necessary to extract and mark this part of the voice segment, and extract the composite voiceprint features contained in the voice segment, including the specific voiceprint features of the registered user and other voiceprint features.

[0026] Step S30: transmitting the voice segment and the composite voiceprint feature to a cloud server, and calling a speech recognition model in the cloud server.

[0027] In this embodiment, after the composite voiceprint features are extracted based on the voice segment, the voice segment and the composite voiceprint features are transmitted to the cloud server. During this process, the complete voice signal can be comprehensively determined based on actual needs and network conditions to determine whether to be transmitted to the cloud server at the same time. Then, the cloud server calls a voice recognition model with stronger voice processing capabilities, which mainly recognizes voice segments with low recognition rates and their composite voiceprint features, thereby improving the voice recognition accuracy while reducing the pressure on data transmission.

[0028] Step S40: identifying the environmental features of the speech segment through the speech recognition model, and automatically adjusting the pre-set speech recognition mode according to the environmental features.

[0029] In this embodiment, the step of identifying the environmental features of the speech segment using the speech recognition model and automatically adjusting the pre-set speech recognition mode according to the environmental features includes: According to the speech signal, the speech segment is split using the speech recognition model to obtain target speech and environmental noise; Classifying the environmental noise, and determining the environmental characteristics of the speech segment based on the environmental noise having a confidence level greater than a preset confidence threshold; According to the environmental characteristics, the pre-set speech recognition mode is automatically adjusted through the speech recognition model to obtain a target recognition mode.

[0030] The step of automatically adjusting a preset speech recognition mode by the speech recognition model according to the environmental characteristics to obtain a target recognition mode comprises: Automatically matching feature identifiers corresponding to the environmental features according to the environmental features; According to the feature identifier corresponding to the environmental feature, the preset feature identifier of the preset speech recognition mode is adjusted through the speech recognition model to obtain the target recognition mode.

[0031] In this embodiment, the speech segment is first split by the speech recognition model to obtain the target speech corresponding to the registered user feature vector and the environmental noise. Then, the environmental noise is classified, and the environmental noise with a confidence level greater than a preset confidence threshold is selected. The environmental characteristics generated by the speech segment are determined based on the environmental noise. Then, based on the identified environmental characteristics, the pre-set general speech recognition mode is automatically adjusted by the speech recognition model to obtain the target recognition mode, for example, adjusted to the in-vehicle recognition mode, thereby obtaining a more powerful in-vehicle speech processing capability.

[0032] More specifically, the speech segment is first split through the speech recognition model to obtain the target speech corresponding to the registered user's feature vector and the environmental noise. The environmental characteristics are determined based on the classification results of the environmental noise, and then the feature identifiers corresponding to the environmental characteristics are automatically matched, such as indoor, street, driving, etc. According to the feature identifiers corresponding to the environmental characteristics, the pre-set preset feature identifiers are adjusted through the speech recognition model to obtain the target recognition mode.

[0033] It's easy to understand that the speech recognition model is trained on pre-batch data and then categorizes and generates multiple recognition modes to cope with various voice dialogue scenarios. For example, in in-vehicle recognition mode, it can automatically eliminate noises emitted by external vehicles (such as horns) and filter out the sounds of other passengers, thereby improving the quality of voice interaction between registered users and the voice system.

[0034] Step S50: in target recognition mode, the composite voiceprint feature is recognized by the voice recognition model, a plurality of voice frames with different voiceprint features are output, and at least one target voice frame is determined.

[0035] In this embodiment, after the target mode is automatically switched by the speech recognition model, the composite voiceprint features contained in the speech segment will be identified by the speech recognition model, multiple speech frames with different voiceprint features will be output, and at least one target speech frame will be determined.

[0036] For example, in a car scenario, a registered user (usually the driver) outputs a voice command of "adjust the air conditioning temperature to 23 degrees" during the interaction with the voice system, but because other passengers simultaneously say "what are we going to eat for lunch today", the voice is confused, and the initial command is generated, for example, "adjust the air conditioning temperature to 23 degrees for lunch today", but in essence, the voiceprint features of the registered user and other passengers are different. Therefore, in this embodiment, in the target recognition mode, according to the voiceprint features of different users, the voice recognition model classifies and outputs multiple voice frames, determines at least one target voice frame, and performs corresponding text conversion for the continuation of voice interaction.

[0037] Step S60: outputting a control instruction for controlling an electronic device according to the target voice frame.

[0038] Exemplarily, after the target voice frame is determined, text conversion is performed based on the target voice frame. For example, the target voice frame is converted into "adjust the air conditioning temperature to 23 degrees" through text conversion, thereby generating a control instruction for controlling the vehicle air conditioner.

[0039] Compared with the prior art, the various speech recognition methods based on edge algorithms shown in this embodiment have the following beneficial effects: In this embodiment, a user's voice signal is captured through a microphone and transmitted to an edge node device for voice recognition. The recognition rate for the voice signal is then output. Speech segments with recognition rates below a preset recognition rate threshold are then extracted from the voice signal. Composite voiceprint features contained in these segments are extracted and transmitted to a cloud server. A speech recognition model is then invoked within the cloud server, which uses the model to identify environmental features generated by the speech segments. A pre-set speech recognition mode is automatically adjusted based on these environmental features. Finally, in target recognition mode, the model identifies the composite voiceprint features, outputs multiple speech frames with differing voiceprint features, identifies at least one target speech frame, and outputs control instructions for controlling electronic devices based on the target speech frame. This embodiment achieves intelligent traffic diversion at the edge node by setting a recognition rate threshold. Combined with compressed transmission of composite voiceprint features and adaptive cloud environment recognition, the accuracy of user-issued control instructions is significantly improved, resolving the trade-off between response speed and adaptability to complex environments in edge speech recognition.

[0040] Example 2 The second embodiment of the present invention also provides a plurality of speech recognition methods based on edge algorithms. The method shown in this embodiment is basically similar to the method shown in the first embodiment, except that: In this embodiment, in the target recognition mode, the steps of identifying the composite voiceprint feature by the voice recognition model, outputting multiple voice frames with different voiceprint features, and determining at least one target voice frame include: In the target recognition mode, the composite voiceprint feature is recognized by the speech recognition model according to the feature identifier of the environmental feature; Splitting the composite voiceprint feature into multiple voiceprint features with different characteristics, and converting them into multiple speech frames accordingly; Performing round-robin matching of voiceprint features on the plurality of speech frames to determine at least one target speech frame.

[0041] The step of performing round-robin matching of voiceprint features on the plurality of voice frames to determine at least one target voice frame includes: Load the feature vector corresponding to the registered user into the preset voiceprint database. When the similarity between any voice frame and the registered user feature vector is greater than the preset similarity threshold, it is marked as a candidate voice frame to obtain a candidate queue; Converting the candidate speech frames in the candidate queue into a text sequence, setting corresponding weight coefficients of the semantics and grammatical structures of keywords in the text sequence, and calculating the instruction relevance score of the text sequence; According to the predetermined environmental feature dynamic matching interference coefficient, the highest scoring frame in the candidate queue is selected as the target speech frame.

[0042] Furthermore, the calculation expression for calculating the instruction relevance score of the text sequence is: S=α×W k +β×W p ; Where S is the instruction relevance score, α∈[0.5,0.7] is the keyword semantic weight coefficient, β∈[0.3,0.5] is the grammatical structure weight coefficient, α+β=1, W k is the weight value of keyword semantics, W p is the weight value of the grammatical structure.

[0043] In this embodiment, after the target recognition mode is obtained through adjustment of the speech recognition model, the pre-extracted composite voiceprint features are first split through the speech recognition model in this mode. Specifically, the composite voiceprint features are split into multiple voiceprint features with characteristic differences, which are converted into multiple voice frames. Then, the voiceprint features of the multiple voice frames are polled and matched, that is, the multiple voice frames are subjected to cyclic voiceprint matching, and at least one target voice frame is determined, that is, the voice frame corresponding to the registered user.

[0044] More specifically, the composite voiceprint features extracted in advance are split through the speech recognition model. Specifically, the composite voiceprint features are split into multiple voiceprint features with feature differences, which are converted into multiple voice frames accordingly. The feature vector corresponding to the registered user is loaded into the preset voiceprint database, that is, the feature vector corresponding to the pre-recorded voice is pre-recorded. When the similarity between any voice frame and the registered user feature vector is greater than the preset similarity threshold, the voice frame is marked as a candidate voice frame. If multiple frames are marked, a candidate queue is obtained, and then each candidate voice frame in the candidate queue is converted into a text sequence, and scored according to the text sequence. Specifically, the weight coefficients of the semantics and grammatical structures of the keywords in the text sequence are set respectively. According to the preset weight distribution, the instruction relevance score of the text sequence corresponding to each candidate voice frame in the candidate queue is calculated, and the interference coefficient is dynamically matched according to the preset environmental characteristics to update the instruction relevance score. The candidate voice frame with the highest score in the candidate queue is selected as the target voice frame, and text conversion and control instruction generation are subsequently performed based on the target voice frame.

[0045] Example 3 See also Figure 2 A third embodiment of the present invention provides a multi-speech recognition system based on an edge algorithm, which is applied to the method described in any of the above embodiments, and the system includes: The voice transmission module 10 is used to obtain the voice signal output by the user through the microphone, transmit the voice signal to the edge node device for voice recognition, and output the recognition rate of the voice signal.

[0046] The speech extraction module 20 is configured to extract speech segments whose recognition rates are lower than a preset recognition rate threshold from the speech signal, and extract composite voiceprint features contained in the speech segments.

[0047] The model calling module 30 is used to transmit the voice segment and the composite voiceprint feature to the cloud server, and call the voice recognition model in the cloud server.

[0048] The mode adjustment module 40 is configured to recognize the environmental features of the speech segment through the speech recognition model and automatically adjust a preset speech recognition mode according to the environmental features.

[0049] The voiceprint recognition module 50 is configured to recognize the composite voiceprint feature by using the speech recognition model in a target recognition mode, output a plurality of speech frames with different voiceprint features, and determine at least one target speech frame.

[0050] The instruction generation module 60 is configured to output a control instruction for controlling an electronic device according to the target speech frame.

[0051] The voiceprint recognition module 50 is specifically used for: In the target recognition mode, the composite voiceprint feature is recognized by the speech recognition model according to the feature identifier of the environmental feature; Splitting the composite voiceprint feature into multiple voiceprint features with different characteristics, and converting them into multiple speech frames accordingly; Performing round-robin matching of voiceprint features on the plurality of speech frames to determine at least one target speech frame.

[0052] Compared with the prior art, the various speech recognition systems based on edge algorithms shown in this embodiment have the following beneficial effects: In this embodiment, a user's voice signal is captured through a microphone and transmitted to an edge node device for voice recognition. The recognition rate for the voice signal is then output. Speech segments with recognition rates below a preset recognition rate threshold are then extracted from the voice signal. Composite voiceprint features contained in these segments are extracted and transmitted to a cloud server. A speech recognition model is then invoked within the cloud server, which uses the model to identify environmental features generated by the speech segments. A pre-set speech recognition mode is automatically adjusted based on these environmental features. Finally, in target recognition mode, the model identifies the composite voiceprint features, outputs multiple speech frames with differing voiceprint features, identifies at least one target speech frame, and outputs control instructions for controlling electronic devices based on the target speech frame. This embodiment achieves intelligent traffic diversion at the edge node by setting a recognition rate threshold. Combined with compressed transmission of composite voiceprint features and adaptive cloud environment recognition, the accuracy of user-issued control instructions is significantly improved, resolving the trade-off between response speed and adaptability to complex environments in edge speech recognition.

[0053] Example 4 A fourth embodiment of the present invention provides a readable storage medium having computer instructions stored thereon, which implement the steps of the method described in the above embodiment when executed by a processor.

[0054] Example 5 A fifth embodiment of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the above embodiment when executing the program.

[0055] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0056] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A multi-speech recognition method based on edge algorithm, characterized in that: The method comprises: Acquire the voice signal output by the user through the microphone, transmit the voice signal to the edge node device for voice recognition, and output the recognition rate of the voice signal; Extracting a speech segment whose recognition rate is lower than a preset recognition rate threshold from the speech signal, and extracting a composite voiceprint feature contained in the speech segment; Transmitting the voice segment and the composite voiceprint feature to a cloud server, and calling a speech recognition model in the cloud server; Identifying environmental features of the speech segment using the speech recognition model, and automatically adjusting a pre-set speech recognition mode according to the environmental features; In target recognition mode, the composite voiceprint feature is recognized by the voice recognition model, multiple voice frames with different voiceprint features are output, and at least one target voice frame is determined; A control instruction for controlling an electronic device is output according to the target voice frame.

2. The multi-speech recognition method based on edge algorithm according to claim 1, characterized in that: The step of identifying the environmental features of the speech segment by the speech recognition model and automatically adjusting a preset speech recognition mode according to the environmental features comprises: According to the speech signal, the speech segment is split using the speech recognition model to obtain target speech and environmental noise; Classifying the environmental noise, and determining the environmental characteristics of the speech segment based on the environmental noise having a confidence level greater than a preset confidence threshold; According to the environmental characteristics, the pre-set speech recognition mode is automatically adjusted through the speech recognition model to obtain a target recognition mode.

3. The multi-speech recognition method based on edge algorithm according to claim 2, characterized in that: The step of automatically adjusting a preset speech recognition mode by the speech recognition model according to the environmental characteristics to obtain a target recognition mode comprises: Automatically matching feature identifiers corresponding to the environmental features according to the environmental features; According to the feature identifier corresponding to the environmental feature, the preset feature identifier of the preset speech recognition mode is adjusted through the speech recognition model to obtain the target recognition mode.

4. The multi-speech recognition method based on edge algorithm according to any one of claims 1 to 3, characterized in that: In the target recognition mode, the steps of identifying the composite voiceprint feature by the voice recognition model, outputting a plurality of voice frames having different voiceprint features, and determining at least one target voice frame include: In the target recognition mode, the composite voiceprint feature is recognized by the speech recognition model according to the feature identifier of the environmental feature; Splitting the composite voiceprint feature into multiple voiceprint features with different characteristics, and converting them into multiple speech frames accordingly; Performing round-robin matching of voiceprint features on the plurality of speech frames to determine at least one target speech frame.

5. The multi-speech recognition method based on edge algorithm according to claim 4, characterized in that: The step of performing round-robin matching of voiceprint features on the plurality of speech frames to determine at least one target speech frame includes: Load the feature vector corresponding to the registered user into the preset voiceprint database. When the similarity between any voice frame and the registered user feature vector is greater than the preset similarity threshold, it is marked as a candidate voice frame to obtain a candidate queue; Converting the candidate speech frames in the candidate queue into a text sequence, setting corresponding weight coefficients of the semantics and grammatical structures of keywords in the text sequence, and calculating the instruction relevance score of the text sequence; According to the predetermined environmental feature dynamic matching interference coefficient, the highest scoring frame in the candidate queue is selected as the target speech frame.

6. The multi-speech recognition method based on edge algorithm according to claim 5, characterized in that: The calculation expression for calculating the instruction relevance score of the text sequence is: S=α×W k +β×W p ; Where S is the instruction relevance score, α∈[0.5,0.7] is the keyword semantic weight coefficient, β∈[0.3,0.5] is the grammatical structure weight coefficient, α+β=1, W k is the weight value of keyword semantics, W p is the weight value of the grammatical structure.

7. A multi-speech recognition system based on edge algorithm, characterized in that: The method according to any one of claims 1 to 6, wherein the system comprises: A voice transmission module is used to obtain the voice signal output by the user through the microphone, transmit the voice signal to the edge node device for voice recognition, and output the recognition rate of the voice signal; A speech extraction module, configured to extract speech segments whose recognition rates are lower than a preset recognition rate threshold from the speech signal, and extract composite voiceprint features contained in the speech segments; A model calling module, configured to transmit the voice segment and the composite voiceprint feature to a cloud server, and call a speech recognition model in the cloud server; a mode adjustment module, configured to identify, by means of the speech recognition model, environmental features of the speech segment, and automatically adjust a pre-set speech recognition mode according to the environmental features; a voiceprint recognition module configured to, in a target recognition mode, identify the composite voiceprint feature using the speech recognition model, output a plurality of speech frames having different voiceprint features, and determine at least one target speech frame; The instruction generation module is used to output a control instruction for controlling the electronic device according to the target voice frame.

8. The multi-speech recognition system based on edge algorithm according to claim 7, characterized in that: The voiceprint recognition module is specifically used for: In the target recognition mode, the composite voiceprint feature is recognized by the speech recognition model according to the feature identifier of the environmental feature; Splitting the composite voiceprint feature into multiple voiceprint features with different characteristics, and converting them into multiple speech frames accordingly; Performing round-robin matching of voiceprint features on the plurality of speech frames to determine at least one target speech frame.

9. A readable storage medium having computer instructions stored thereon, characterized in that: When the instruction is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method according to any one of claims 1 to 6 are implemented.

Images