Speech recognition method and device based on large model, medium, equipment and product

Abstract

The invention relates to the technical field of voice recognition, and particularly provides a voice recognition method and device based on a large model, a medium, equipment and a product, and the method can comprise the steps: in a process of continuously receiving audio data sent by a client, when the voice state of the audio data is determined to be a speaking state, the voice data is sent to the client; generating an identification task corresponding to the audio data; caching the identification task to an audio task queue; performing voice recognition on the recognition task in the audio task queue by using a voice large model to obtain a voice recognition result; and the voice recognition result is pushed to the client. According to the embodiment of the invention, the voice can be recognized while being received, and the voice recognition efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of speech recognition technology, and more specifically, to a method, apparatus, medium, device, and product for speech recognition based on a large model. Background Technology

[0002] With the rapid development of artificial intelligence and voice interaction technology, real-time speech recognition (ASR) has become a core technology for many applications such as online conferencing, voice assistants, real-time captions, and intelligent customer service.

[0003] Traditional real-time speech recognition systems typically employ end-to-end models based on recurrent neural networks (RNNs) or connectionist temporal classification (CTC). During speech recognition, the system must wait for the user to stop speaking before it can begin recognizing the entire speech segment. This means there is a delay between the user uttering the last word and seeing the recognition result; this delay affects the real-time nature and fluency of voice interaction.

[0004] Therefore, how to provide a low-latency speech recognition method has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of some embodiments of this application is to provide a method, apparatus, medium, device and product for speech recognition based on a large model. The technical solutions of the embodiments of this application can realize speech recognition while receiving it, improve speech recognition efficiency and reduce recognition latency, and ensure the real-time and smoothness of speech interaction.

[0006] In a first aspect, some embodiments of this application provide a method for speech recognition based on a large model, comprising: during the continuous reception of audio data sent by a client, when it is determined that the speech state of the audio data is a speaking state, generating a recognition task corresponding to the audio data; caching the recognition task in an audio task queue; using a large speech model to perform speech recognition on the recognition task in the audio task queue to obtain a speech recognition result; and pushing the speech recognition result to the client.

[0007] Some embodiments of this application generate corresponding recognition tasks and cache them in an audio task queue when the voice state is determined to be speaking during the reception of audio data. Then, a large voice model is used to perform speech recognition on the recognition tasks in the audio task queue, obtaining the speech recognition results which are then pushed to the client. These embodiments enable simultaneous reception, recognition, and push of voice data, improving speech recognition efficiency, reducing recognition latency, and ensuring the real-time nature and smoothness of voice interaction.

[0008] In some embodiments, determining the voice state of the audio data as a speaking state includes: if it is determined that there are no breakpoints within a preset time period in the audio data, then the voice state is the speaking state.

[0009] Some embodiments of this application determine the voice state by checking whether there are breakpoints in the audio data during a preset time period, which can effectively identify the voice interaction situation.

[0010] In some embodiments, the method further includes: if it is confirmed that there is a breakpoint within a preset time period in the audio data, then the voice state is an end state; generating a voice dialogue recognition result corresponding to the complete audio data in the current dialogue of the client; and pushing the voice dialogue recognition result to the client.

[0011] Some embodiments of this application determine whether the audio data has reached the end state by checking whether there are breakpoints in the preset time period. After the end state, the complete voice dialogue recognition result is pushed to the client so that the client can obtain the complete and accurate voice recognition result.

[0012] In some embodiments, before receiving the audio data sent by the client, the method further includes: parsing the received request data from the client to obtain request parameters; wherein the request parameters include version information of the speech model and request item information; creating a session pool corresponding to the request parameters; wherein the session pool is used to store the audio data and the speech recognition result corresponding to the audio data.

[0013] Some embodiments of this application create corresponding session pools by receiving request parameters from the client, which can achieve effective storage of audio data and speech recognition results.

[0014] In some embodiments, before performing speech recognition on the recognition task in the audio task queue using the large speech model, the method further includes: invoking the large speech model that matches the version information.

[0015] Some embodiments of this application achieve effective and accurate speech recognition by calling corresponding large speech models.

[0016] In some embodiments, the reception and recognition of the audio data are controlled using an asynchronous architecture.

[0017] Some embodiments of this application receive and recognize audio data using an asynchronous architecture, which enables simultaneous reception and recognition, improving the efficiency of speech recognition and reducing recognition latency.

[0018] Secondly, some embodiments of this application provide a speech recognition apparatus based on a large model, comprising: a task generation module, configured to generate a recognition task corresponding to the audio data when the speech state of the audio data is determined to be a speaking state during the continuous reception of audio data sent by a client; a caching module, configured to cache the recognition task in an audio task queue; and a recognition module, configured to perform speech recognition on the recognition task in the audio task queue using a large speech model, obtain a speech recognition result, and push the speech recognition result to the client.

[0019] Thirdly, some embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the method described in any embodiment of the first aspect.

[0020] Fourthly, some embodiments of this application provide an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, can implement the method as described in any embodiment of the first aspect.

[0021] Fifthly, some embodiments of this application provide a computer program product, the computer program product including a computer program, wherein the computer program, when executed by a processor, can implement the method described in any embodiment of the first aspect. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of some embodiments of this application, the accompanying drawings used in some embodiments of this application will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A system diagram of large-model-based speech recognition provided for some embodiments of this application; Figure 2 One of the flowcharts for a large-model-based speech recognition method provided for some embodiments of this application; Figure 3 A second flowchart of a large-model-based speech recognition method provided for some embodiments of this application; Figure 4 A block diagram of a device for speech recognition based on a large model provided for some embodiments of this application; Figure 5 A schematic diagram of an electronic device provided for some embodiments of this application. Detailed Implementation

[0024] The technical solutions of some embodiments of this application will now be described with reference to the accompanying drawings.

[0025] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] In related technologies, existing streaming speech recognition services typically include the following processing flow: The client and server establish a persistent connection via WebSocket or long polling for bidirectional real-time communication. The client collects microphone audio data, performs front-end preprocessing (such as noise reduction, echo cancellation, and audio encoding), and then frames, packages, and continuously sends the binary audio stream to the server. The server receives the audio data packets and places them in a buffer for sorting and reassembly to handle network jitter and packet order issues. The VAD module performs real-time analysis of the audio buffer, accurately detecting segments in the audio stream containing human voices (i.e., the start and end points of speech activity) and filtering out silent and noisy segments. This is crucial because it avoids sending invalid audio to the recognition engine, saving computational resources. Once VAD confirms a complete speech segment (from START to END), the audio segment is sent to the ASR inference engine. The engine typically consists of an acoustic model and a language model. The acoustic model converts audio features into phonetic or phoneme probabilities, while the language model decodes the most probable text sequence based on these. Finally, the identified text results are returned to the client in real time or near real time, and the client displays or performs other subsequent processing.

[0027] While the above scheme achieves "streaming" recognition, its core mode is "segmented recognition." That is, it must wait for the VAD module to confirm the complete end of a speech segment before sending the entire segment to the ASR model for recognition. This is acceptable for short speech responses, but it has inherent limitations in interactive scenarios requiring extremely low latency and intermediate results.

[0028] As the aforementioned technologies show, existing solutions require waiting for the user to stop speaking (VAD detects STOP_TALKING) before the entire segment recognition process can begin. This means there is an unavoidable waiting time (i.e., end-point delay) between the user saying the last word and seeing the recognition result. This delay is very noticeable in long sentences or with significant network latency, disrupting the real-time nature and smoothness of the interaction. This results in bursty processing by the ASR engine. During prolonged silence, the engine is idle; when the user speaks, the engine needs to process large amounts of data instantly. This pulsed load poses challenges to the allocation and elastic scaling of computing resources, easily leading to task backlog and increased latency during peak traffic periods. True real-time interaction requires "hear and display simultaneously," meaning that for every word the user says, a corresponding word or phrase is displayed on the screen. Traditional segmented recognition cannot provide smooth intermediate results. Some solutions simulate real-time effects by cutting long segments into fixed small segments, but this easily produces incorrect truncation and absurd intermediate results at segment boundaries, resulting in a poor user experience. Since the entire speech segment needs to be cached until its end, memory usage is linear. When processing extremely long audio clips (e.g., several minutes long), memory pressure is high, and if the recognition process is interrupted, the computation of the entire segment is wasted. Moreover, the modules are tightly coupled, resulting in poor flexibility; VAD and ASR are usually tightly coupled and executed serially. The sensitivity parameters of VAD (e.g., silence duration) directly determine the recognition latency and segmentation effect, making it difficult to dynamically adjust the optimal strategy according to different scenarios (e.g., meetings, voice input methods).

[0029] In view of this, some embodiments of this application provide a speech recognition method based on a large model. In this method, when a speaking state is detected during the continuous reception of audio data sent by a client, a corresponding recognition task can be generated and cached in an audio task queue. At this time, the large speech model can read the corresponding recognition task from the audio task queue, perform speech recognition processing, obtain the speech recognition result, and push the speech recognition result to the client in real time. The embodiments of this application achieve simultaneous reception and recognition of speech data through an asynchronous pipeline of simultaneous reception and recognition, reducing recognition latency. Simultaneously, asynchronous task management based on the audio task queue enables smooth utilization of computing resources and stable real-time intermediate result generation. Dynamic configuration management of the queue and multimodal support enhance the system's flexibility and applicability.

[0030] The following is in conjunction with the appendix Figure 1 The overall structure of a large-model-based speech recognition system provided by some embodiments of this application is illustrated by way of example.

[0031] like Figure 1As shown in the diagram, some embodiments of this application provide a system diagram for speech recognition based on a large model. This system may include a client 100 and a server 200. The client 100 may request to establish a WebSocket communication connection with the server 200. After establishing the connection, the client 100 may send audio data to the server 200. While continuously receiving audio data, the server 200 may, after confirming that the speech state is speaking, establish a recognition task for the audio data and cache it in an audio task queue; then, it may call the large speech model to perform the recognition operation on the recognition task to obtain the speech recognition result; finally, it may push the speech recognition result to the client 100 in real time. Embodiments of this application can achieve simultaneous reception and recognition of audio data, reducing recognition latency and improving speech recognition efficiency.

[0032] In some embodiments of this application, the client 100 can be a mobile terminal or a non-portable computer terminal. The server 200 can be a WebSocket server or other types of server equipment; this application does not specifically limit the specific implementation of the server.

[0033] The following is in conjunction with the appendix Figure 2 The present application provides an exemplary embodiment of the implementation process of large-model-based speech recognition performed by server 200.

[0034] Please see the appendix Figure 2 , Figure 2 A flowchart of a speech recognition method based on a large model is provided for some embodiments of this application.

[0035] In some embodiments of this application, before performing the following method S210, the method for speech recognition based on a large model may further include: parsing the received request data from the client to obtain request parameters; wherein the request parameters include version information of the speech large model and request item information; creating a session pool corresponding to the request parameters; wherein the session pool is used to store the audio data and the speech recognition results corresponding to the audio data.

[0036] For example, in a specific embodiment of this application, WebSocket is defined as the transport layer, creating an independent, stateful object for each client connection to manage the client's session lifecycle and cached data. This design enables the server 200 to handle tens of thousands of independent, customizable identification sessions simultaneously and maintain its context (such as VAD history and model context) for each session, laying the foundation for high-quality, personalized identification.

[0037] Specifically, after receiving the request data URL from client 100, the relevant request parameters are extracted from the URL. For example, the URL is: http: / / 127.0.0.1:8000 / llm_asr?version=v0.1&client_id=123&company_id=2210. Here, version=v0.1 indicates that the ASR version is 0.1 (as a specific example of version information), which is identified through regular expressions; the request ID is 123 (as a unique identifier for the session), and the project space ID is 2210, indicating whether the project space belongs to an insurance project scenario or a financial project scenario. Different ASR versions represent different ASR models (i.e., large speech models). These request parameters can be used to create an independent session pool for client 100. This session pool can store the audio data sent by client 100 and the speech recognition results after recognition.

[0038] In some embodiments of this application, the large-model-based speech recognition method may include: S210, during the continuous reception of audio data sent by the client, when it is determined that the voice state of the audio data is a speaking state, a recognition task corresponding to the audio data is generated.

[0039] For example, in a specific embodiment of this application, while continuously receiving audio data sent by client 100, server 200 performs VAD (Voice Activity Detection) detection on the audio data to confirm the voice state of the audio data. If it is in the Talking state (a specific example of a speaking state), the audio data is packaged to generate a RecogTask without waiting for the audio to end. The buffer of the RecogTask stores accumulated audio, that is, the audio data from client 100 is accumulated once at a preset period (e.g., 20ms) and packaged into a RecogTask object.

[0040] The reception and recognition of audio data are controlled using an asynchronous architecture.

[0041] Specifically, this application employs a fully decoupled asynchronous pipeline, where the audio receiving (handle_audio) coroutine and the audio processing (handle_audio_process) coroutine run concurrently, communicating through a high-efficiency CustomQueue. This decoupling allows subsequent ASR recognition to proceed without waiting for the speech segment to end; while the user is still speaking, the spoken audio segment is already fed into the ASR model, and calculations and intermediate results are being output. This completely eliminates "tail point latency," achieving true "speak-as-you-go recognition." Simultaneously, the asynchronous architecture smooths the computational load and improves CPU / GPU utilization.

[0042] In some embodiments of this application, S210 may include: if it is determined that there is no breakpoint within a preset time period in the audio data, then the voice state is the speaking state.

[0043] For example, in a specific embodiment of this application, the voice state is determined to be either Talking or STOP_TALKING by analyzing whether there are any breaks in the audio data. For instance, if it is determined that there are no breaks in the audio data for a preset time period (e.g., 500ms), then the voice state of client 100 is still Talking.

[0044] S220, the recognition task is cached in the audio task queue.

[0045] For example, in a specific embodiment of this application, RecogTask is cached in the audio task queue.

[0046] In some embodiments of this application, before performing S230, the method for speech recognition based on a large model further includes: invoking the large speech model that matches the version information.

[0047] For example, in a specific embodiment of this application, a large voice model matching version=v0.1 in the call and request parameters is used.

[0048] S230, the speech recognition task in the audio task queue is performed using a large speech model to obtain the speech recognition result; and the speech recognition result is pushed to the client.

[0049] For example, in a specific embodiment of this application, a speech recognition task is read from the audio task queue using a large speech model, and the speech recognition result is obtained and pushed to the client 100 in real time. Simultaneously, the speech recognition result is also stored in the session pool.

[0050] Understandably, the large-scale speech model integrated into the server-side 200 with streaming output capabilities (e.g., a transformer-based streaming model) can progressively acquire recognition results when calling the model, rather than waiting for all processing to complete. During the progressive recognition process, previously recognized intermediate results can be initiated. Specifically, the streaming generation service using the `transcribe_streaming` function will stream text (i.e., each intermediate result). The large-scale speech model not only outputs more accurate text, but its streaming characteristics perfectly match the asynchronous pipeline architecture of this application, enabling continuous, word-by-word return of results, providing users with a smooth interactive experience.

[0051] In some embodiments of this application, the speech recognition method based on a large model may further include: if it is confirmed that there is a breakpoint within a preset time period in the audio data, then the speech state is an ended state; generating a speech dialogue recognition result corresponding to the complete audio data in the current dialogue of the client; and pushing the speech dialogue recognition result to the client.

[0052] For example, in a specific embodiment of this application, after detection, it is determined that there is a breakpoint in the audio data for a preset time period (e.g., 500ms). At this time, it indicates that the voice state of the client 100 is STOP_TALKING and the voice ends. At this time, the final task can be generated, that is, the voice dialogue recognition result of the complete audio data of the current dialogue round of the user with the client 100 can be generated.

[0053] The following is in conjunction with the appendix Figure 3 The present application provides an exemplary description of the specific process of large-model-based speech recognition provided by some embodiments of this application.

[0054] Please see the appendix Figure 3 , Figure 3 A flowchart of a speech recognition method based on a large model is provided for some embodiments of this application.

[0055] The above process is illustrated below by example.

[0056] S310, the client sends request data to the server.

[0057] S320: The server parses the request data and obtains the request parameters; it then creates a session pool based on the request parameters.

[0058] S330: The server continuously receives audio data sent by the client.

[0059] S340: Determine if there is a breakpoint in the audio data within a preset time period. If not, execute S350; otherwise, execute S380.

[0060] S350 generates a recognition task corresponding to the audio data.

[0061] The S360 uses a large speech model to perform speech recognition on recognition tasks in the audio task queue and obtain speech recognition results.

[0062] S370: The server pushes the speech recognition results to the client and returns S330.

[0063] S380 generates a speech dialogue recognition result corresponding to the complete audio data in the current conversation with the client, and pushes the speech dialogue recognition result to the client.

[0064] It is understood that the specific implementation process of S310~S380 can be referred to the method embodiment provided above. To avoid repetition, detailed descriptions are omitted here.

[0065] Through the embodiments described above, this application employs an asynchronous pipelined architecture that decouples VAD and ASR, and immediately initiates the recognition process upon detecting voice activity, reducing recognition latency and eliminating tail-point waiting. Furthermore, the asynchronous, non-blocking architecture design allows I / O operations and CPU / GPU-intensive operations (model inference) to be fully parallelized. The computing unit is not idle during network transmission and VAD processing; the network module can continue to receive data during model computation.

[0066] Please refer to Figure 4 , Figure 4 The diagram illustrates a block diagram of a large-model-based speech recognition apparatus provided in some embodiments of this application. It should be understood that this large-model-based speech recognition apparatus corresponds to the method embodiments described above and is capable of performing the various steps involved in the method embodiments. The specific functions of this large-model-based speech recognition apparatus can be found in the description above; detailed descriptions are omitted here to avoid repetition.

[0067] Figure 4 The large-model-based speech recognition device includes at least one software functional module that can be stored in a memory or embedded in the large-model-based speech recognition device in the form of software or firmware. The large-model-based speech recognition device includes: a task generation module 410, used to generate a recognition task corresponding to the audio data when the speech state of the audio data is determined to be a speaking state during the continuous reception of audio data sent by the client; a caching module 420, used to cache the recognition task to an audio task queue; and a recognition module 430, used to perform speech recognition on the recognition task in the audio task queue using a large-model speech recognition device, obtain the speech recognition result, and push the speech recognition result to the client.

[0068] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the aforementioned method, and will not be elaborated further here.

[0069] Some embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can perform the operation of any of the methods corresponding to the methods provided in the above embodiments.

[0070] Some embodiments of this application also provide a computer program product, which includes a computer program, wherein when the computer program is executed by a processor, it can implement the operation of any of the methods corresponding to the above embodiments provided in the above embodiments.

[0071] Figure 5 A schematic diagram of the structure of an electronic device 500 is shown. (See attached diagram.) Figure 5 As shown, the electronic device 500 includes a processor 510 and a memory 520, and optionally may also include a power supply 530, a display unit 540, and an input unit 550.

[0072] The processor 510 is the control center of the electronic device 500. It connects various components through various interfaces and lines, and performs various functions of the electronic device 500 by running or executing software programs and / or data stored in the memory 520, thereby performing overall monitoring of the electronic device 500.

[0073] In this embodiment, when the processor 510 calls the computer program stored in the memory 520, it executes the steps in the above embodiments.

[0074] Optionally, processor 510 may include one or more processing units; preferably, processor 510 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into processor 510. In some embodiments, the processor and memory may be implemented on a single chip; in some embodiments, they may also be implemented separately on independent chips.

[0075] The memory 520 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, various applications, etc.; the data storage area may store data created based on the use of the electronic device 500, etc. In addition, the memory 520 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0076] Electronic device 500 also includes a power supply 530 (such as a battery) that supplies power to various components. The power supply can be logically connected to processor 510 through a power management system, thereby enabling the power management system to manage functions such as charging, discharging, and power consumption.

[0077] The display unit 540 can be used to display information input by the user or information provided to the user, as well as various menus of the electronic device 500. In this embodiment of the invention, it is mainly used to display the display interfaces of various applications in the electronic device 500, as well as text, images, and other objects displayed on the display interfaces. The display unit 540 may include a display panel 541. The display panel 541 may be configured in the form of a liquid crystal display (LCD), an organic light-emitting diode (OLED), or the like.

[0078] The input unit 550 can be used to receive information such as numbers or characters input by the user. The input unit 550 may include a touch panel 551 and other input devices 552. The touch panel 551, also known as a touch screen, can collect touch operations on or near the user (such as operations performed by the user using a finger, stylus, or any suitable object or accessory on or near the touch panel 551).

[0079] Specifically, the touch panel 551 can detect user touch operations and the signals generated by these operations, convert them into touch point coordinates, send them to the processor 510, and receive and execute commands from the processor 510. Furthermore, the touch panel 551 can be implemented using various types of sensors, including resistive, capacitive, infrared, and surface acoustic wave sensors. Other input devices 552 can include, but are not limited to, one or more of the following: physical keyboard, function keys (such as volume control buttons, power buttons, etc.), trackball, mouse, joystick, etc.

[0080] Of course, the touch panel 551 can cover the display panel 541. When the touch panel 551 detects a touch operation on or near it, it transmits the information to the processor 510 to determine the type of touch event. Subsequently, the processor 510 provides corresponding visual output on the display panel 541 according to the type of touch event. Although in Figure 5 In this embodiment, the touch panel 551 and the display panel 541 are two separate components to realize the input and output functions of the electronic device 500. However, in some embodiments, the touch panel 551 and the display panel 541 can be integrated to realize the input and output functions of the electronic device 500.

[0081] The electronic device 500 may also include one or more sensors, such as a pressure sensor, a gravity acceleration sensor, a proximity sensor, etc. Of course, depending on the specific application, the electronic device 500 may also include other components such as a camera. Since these components are not the focus of this application embodiment, therefore... Figure 5 It is not shown in the text and will not be described in detail here.

[0082] Those skilled in the art will understand that Figure 5 This is merely an example of an electronic device and does not constitute a limitation on the electronic device. It may include more or fewer components than shown, or a combination of certain components, or different components.

[0083] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0084] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0085] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A speech recognition method based on a large model, characterized in that, include: During the continuous reception of audio data sent by the client, when the voice state of the audio data is determined to be a speaking state, a recognition task corresponding to the audio data is generated. The recognition task is cached in the audio task queue; The speech recognition task in the audio task queue is performed using a large speech model to obtain the speech recognition result; and the speech recognition result is pushed to the client.

2. The method as described in claim 1, characterized in that, Determining the voice state of the audio data as a speaking state includes: If it is determined that there are no breakpoints within a preset time period in the audio data, then the voice state is the speaking state.

3. The method as described in claim 1, characterized in that, The method further includes: If it is confirmed that there is a breakpoint within a preset time period in the audio data, then the voice state is the end state; Generate a voice dialogue recognition result corresponding to the complete audio data in the current conversation of the client; and push the voice dialogue recognition result to the client.

4. The method according to any one of claims 1-3, characterized in that, Before receiving the audio data sent by the client, the method further includes: The received request data from the client is parsed to obtain request parameters; wherein, the request parameters include the version information of the voice model and the request item information; Create a session pool corresponding to the request parameters; wherein the session pool is used to store the audio data and the speech recognition result corresponding to the audio data.

5. The method as described in claim 4, characterized in that, Before performing speech recognition on the recognition task in the audio task queue using the large speech model, the method further includes: Invoke the large voice model that matches the version information.

6. The method according to any one of claims 1-3, characterized in that, The reception and recognition of the audio data are controlled using an asynchronous architecture.

7. A speech recognition device based on a large model, characterized in that, include: The task generation module is used to generate a recognition task corresponding to the audio data when the voice state of the audio data is determined to be speaking state during the continuous reception of audio data sent by the client. A caching module is used to cache the recognition task to an audio task queue; The recognition module is used to perform speech recognition on the recognition tasks in the audio task queue using a large speech model, obtain the speech recognition results, and push the speech recognition results to the client.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program is executed by a processor to perform the method as described in any one of claims 1-6.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored on the memory and running on the processor, wherein the computer program is executed by the processor to perform the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program, wherein the computer program is executed by a processor to perform the method as described in any one of claims 1-6.

Images