Vehicle-mounted voice activation detection method and device, equipment and storage medium

Abstract

The invention discloses a vehicle-mounted voice activation detection method, device and equipment and a storage medium, and the method comprises the steps: determining a current audio block and a previous audio block according to an in-vehicle audio stream; acquiring historical audio information corresponding to the previous audio block; and inputting the current audio block and the historical audio information to a preset classification model to obtain the category of the current audio block output by the preset classification model, the category of the current audio block including a voice head end, a voice middle segment, a voice tail end and a non-voice segment. The historical audio information is added on the basis of the current audio block and is input into the preset classification model, so that the preset classification model can more accurately classify the current audio block on the basis of the historical audio information, and the stability and accuracy of voice activation detection in the vehicle-mounted environment are improved.

Description

Technical Field

[0001] The present application belongs to the field of voice activation detection technology, and in particular, relates to a vehicle-mounted voice activation detection method, device, equipment and storage medium. Background Art

[0002] Voice Activity Detection (VAD) refers to the detection of speech segments in an audio stream. In a car environment, the audio stream in the car needs to be detected in real time, which places higher requirements on the accuracy of the detection.

[0003] Traditional voice activation detection solutions mainly determine the energy difference between the voice signal and the noise signal in the audio stream, and then determine the voice segment and the noise segment based on the energy threshold. In the vehicle environment, the background audio changes frequently. The above traditional solutions are difficult to meet the transient environment and are prone to misjudgment. Therefore, how to improve the accuracy of voice activation detection in the vehicle environment is an urgent problem to be solved. Summary of the invention

[0004] The embodiments of the present application provide a method, apparatus, device and storage medium for in-vehicle voice activation detection, thereby improving the accuracy of voice activation detection in a vehicle environment at least to a certain extent.

[0005] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by the practice of the present application.

[0006] According to a first aspect of an embodiment of the present application, a vehicle-mounted voice activation detection method is provided, comprising:

[0007] Determine a current audio block and a previous audio block according to the in-vehicle audio stream;

[0008] Acquire historical audio information corresponding to the previous audio block;

[0009] The current audio block and the historical audio information are input into a preset classification model to obtain a category of the current audio block output by the preset classification model, wherein the categories of the current audio block include: a speech start end, a speech middle segment, a speech end end, and a non-speech segment.

[0010] In some embodiments of the present application, based on the aforementioned solution, the preset model includes a convolutional layer, a multi-head attention layer, a long short-term memory network, and a linear classification layer connected in sequence, and the current audio block and the historical audio information are input into the preset classification model to obtain the category of the current audio block output by the preset classification model, including:

[0011] Inputting the current audio block into the convolutional layer to obtain a current spectrogram;

[0012] Input the current spectrogram into the multi-head attention layer to obtain a current attention vector;

[0013] Inputting the current attention vector and the historical audio information into the long short-term memory network to obtain a current output value and a current state of the long short-term memory network;

[0014] The current output value is input into the linear classification layer to obtain the category of the current audio block.

[0015] In some embodiments of the present application, based on the above scheme, inputting the current spectrogram into the multi-head attention layer to obtain the current attention vector includes:

[0016] The current spectrogram is input into the multi-head attention layer so that the multi-head attention layer extracts features from the current spectrogram, performs pre-classification based on the extracted features, and outputs a current attention vector corresponding to the pre-classification result.

[0017] In some embodiments of the present application, based on the above solution, the acquiring the historical audio information corresponding to the previous audio block includes:

[0018] A previous output value and a previous state obtained by the long short-term memory network based on the previous audio block are determined as historical audio information corresponding to the previous audio block.

[0019] In some embodiments of the present application, based on the above solution, before determining the current audio block and the previous audio block according to the in-car audio stream, the method further includes:

[0020] Training the convolutional layer using sample audio blocks to obtain a sample spectrogram;

[0021] After the convolutional layer training is completed, the multi-head attention layer and the long short-term memory network are trained using the sample spectrogram to obtain the preset classification model.

[0022] In some embodiments of the present application, based on the aforementioned scheme, the objective function of the convolution layer is a mean square error loss function.

[0023] In some embodiments of the present application, based on the aforementioned solution, the using the sample spectrogram to train the multi-head attention layer and the long short-term memory network includes:

[0024] The weights of the beginning and the end of the speech in the objective function controlling the multi-head attention layer and the long short-term memory network are both greater than the weights of the middle segment of the speech and the non-speech segment.

[0025] According to a second aspect of an embodiment of the present application, there is provided a vehicle-mounted voice activation detection device, comprising:

[0026] A current audio block acquisition module, used to determine a current audio block and a previous audio block according to an in-car audio stream;

[0027] A historical audio block acquisition module, used to acquire historical audio information corresponding to the previous audio block;

[0028] The current audio block classification module is used to input the current audio block and the historical audio information into a preset classification model to obtain the category of the current audio block output by the preset classification model, wherein the categories of the current audio block include: speech beginning, speech middle segment, speech end and non-speech segment.

[0029] In some embodiments of the present application, based on the aforementioned scheme, the preset model includes a convolutional layer, a multi-head attention layer, a long short-term memory network and a linear classification layer connected in sequence, and the current audio block classification module is also used to input the current audio block into the convolutional layer to obtain the current spectrogram; input the current spectrogram into the multi-head attention layer to obtain the current attention vector; input the current attention vector and the historical audio information into the long short-term memory network to obtain the current output value and current state of the long short-term memory network; input the current output value into the linear classification layer to obtain the category of the current audio block.

[0030] In some embodiments of the present application, based on the aforementioned scheme, the current audio block classification module is also used to input the current spectrogram into the multi-head attention layer, so that the multi-head attention layer can extract features from the current spectrogram, perform pre-classification according to the extracted features, and output the current attention vector corresponding to the pre-classification result.

[0031] In some embodiments of the present application, based on the aforementioned scheme, the historical audio block acquisition module is further used to determine the previous output value and the previous state obtained by the long short-term memory network based on the previous audio block as the historical audio information corresponding to the previous audio block.

[0032] In some embodiments of the present application, based on the aforementioned scheme, the current audio block classification module is also used to train the convolution layer using the sample audio block to obtain a sample spectrogram; after the convolution layer training is completed, the multi-head attention layer and the long short-term memory network are trained using the sample spectrogram to obtain the preset classification model.

[0033] In some embodiments of the present application, based on the aforementioned scheme, the objective function of the convolution layer is a mean square error loss function.

[0034] In some embodiments of the present application, based on the aforementioned scheme, the current audio block classification module is also used to control the weights of the beginning and the end of the speech in the objective function of the multi-head attention layer and the long short-term memory network to be greater than the weights of the middle segment of the speech and the non-speech segment.

[0035] According to a third aspect of an embodiment of the present application, a vehicle-mounted voice activation detection device is provided, comprising a processor and a memory, wherein the memory stores computer program instructions that can be executed by the processor, and when the processor executes the computer program instructions, the steps of the method described in any one of the first aspects above are implemented.

[0036] According to a fourth aspect of an embodiment of the present application, a computer-readable storage medium is provided, in which computer program instructions are stored. When the computer program instructions are executed by a processor, the processor is prompted to implement the steps of the method described in any one of the first aspects above.

[0037] In the present application, the current audio block and the previous audio block are determined according to the in-vehicle audio stream; the historical audio information corresponding to the previous audio block is obtained; the current audio block and the historical audio information are input into a preset classification model to obtain the category of the current audio block output by the preset classification model, wherein the categories of the current audio block include: speech start, speech middle, speech end, and non-speech segment. Among them, by adding historical audio information to the preset classification model on the basis of the current audio block, the preset classification model can more accurately classify the current audio block based on the historical audio information, thereby improving the stability and accuracy of voice activation detection in a vehicle environment.

[0038] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The drawings herein are incorporated into the specification and constitute a part of the specification, showing embodiments consistent with the present application, and together with the specification, are used to explain the principles of the present application. Obviously, the drawings described below are only some embodiments of the present application, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0040] Figure 1 A schematic diagram of a process flow of a vehicle-mounted voice activation detection method in one embodiment is shown;

[0041] Figure 2 A schematic diagram showing a voice start end, a voice middle segment, a voice end, and a non-voice segment in an in-vehicle audio stream in one embodiment is shown;

[0042] Figure 3 A detailed schematic diagram of an in-vehicle voice activation detection method in one embodiment is shown;

[0043] Figure 4 A block diagram of a vehicle-mounted voice activation detection device in one embodiment is shown;

[0044] Figure 5 A schematic structural diagram of a vehicle-mounted voice activation detection device in one embodiment is shown. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0046] In addition, described feature, structure or characteristic can be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present application. However, those skilled in the art will appreciate that the technical scheme of the present application can be put into practice without one or more of the specific details, or other methods, components, devices, steps, etc. can be adopted. In other cases, known methods, devices, realizations or operations are not shown or described in detail to avoid blurring the various aspects of the application.

[0047] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0048] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0049] Figure 1 FIG. 2 shows a flow chart of a vehicle-mounted voice activation detection method in one embodiment. Figure 1 As shown, the in-vehicle voice activation detection method may include the following steps 101 to 103.

[0050] In step 101, a current audio block and a previous audio block are determined according to an in-car audio stream.

[0051] It is understandable that the in-car audio stream refers to an audio stream collected by a device such as an in-car microphone. An audio block may be an in-car audio stream of a certain duration, for example, an in-car audio stream of 50 ms to 200 ms.

[0052] During implementation, the audio block currently to be classified may be used as the current audio block, and the last classified audio block may be used as the previous audio block.

[0053] In step 102, historical audio information corresponding to a previous audio block is obtained.

[0054] It can be understood that the historical audio information includes compressed information of all historical audio blocks since the vehicle computer was turned on, and this information can assist in the classification of the current audio block.

[0055] In step 103, the current audio block and historical audio information are input into a preset classification model to obtain the category of the current audio block output by the preset classification model, wherein the categories of the current audio block include: speech beginning, speech middle segment, speech end and non-speech segment.

[0056] Figure 2 A schematic diagram of the voice beginning, voice middle, voice end and non-voice segment in an in-car audio stream in one embodiment is shown. Figure 2 As shown in the figure, the beginning of the speech is an audio block with no speech in the first part and speech in the second part; the end of the speech is an audio block with speech in the first part and no speech in the second part. Accurately modeling the beginning and end of the speech can greatly improve the smoothness and real-time performance of downstream applications.

[0057] In some embodiments, the preset model may include a convolutional layer, a multi-head attention layer, a long short-term memory network and a linear classification layer connected in sequence, and the current audio block and the historical audio information are input into the preset classification model to obtain the category of the current audio block output by the preset classification model, including: inputting the current audio block into the convolutional layer to obtain the current spectrogram; inputting the current spectrogram into the multi-head attention layer to obtain the current attention vector; inputting the current attention vector and the historical audio information into the long short-term memory network to obtain the current output value and current state of the long short-term memory network; inputting the current output value into the linear classification layer to obtain the category of the current audio block.

[0058] Among them, the convolution layer is used to simulate Fourier transform. The current audio block is a time domain signal, which can be converted into the current spectrogram in the frequency domain after the transformation processing of the convolution layer to better represent the characteristics of the current audio block.

[0059] The multi-head attention layer is used to extract features from the current spectrogram, perform pre-classification based on the extracted features, and output the current attention vector corresponding to the pre-classification result.

[0060] It is understandable that the beginning of speech, the middle of speech, the end of speech, and the non-speech segment all have their own characteristics. For example, the first part of the beginning of speech has no speech and the second part has speech, the middle of speech has only speech, the first part of the end of speech has speech and the second part has no speech, and the non-speech segment has no speech. The multi-head attention layer can pre-classify the features extracted from the current spectrogram, determine which category it belongs to among the beginning of speech, the middle of speech, the end of speech, and the non-speech segment, and output the corresponding current attention vector. By introducing a multi-layer attention mechanism, the beginning and end of the speech segment can be located more accurately, improving the detection accuracy.

[0061] The long short-term memory network is used to link the information of all historical audio blocks. It will compress the current audio block and the historical audio block information, and the output result is used as the historical audio block information required for the calculation of the next audio block. By using the long short-term memory network to model the long-term dependence of the in-vehicle audio stream, even in high-noise and high-speed vehicle scenarios, it can accurately distinguish between speech segments (including the beginning, middle and end of speech) and non-speech segments, thereby improving the anti-interference ability of detection in vehicle scenarios. At the same time, the solution can adapt to transient changes in the environment, adaptively adjust the classification results according to changes in the environment, and improve the stability and accuracy of detection in the vehicle environment.

[0062] In some embodiments, obtaining historical audio information corresponding to a previous audio block includes: determining a previous output value and a previous state obtained by a long short-term memory network based on the previous audio block as the historical audio information corresponding to the previous audio block.

[0063] Figure 3 FIG. 2 shows a detailed schematic diagram of a vehicle-mounted voice activation detection method in an embodiment. Figure 3 As shown, the Long Short-Term Memory (LSTM) network obtains the previous output value h(t-1) and the previous state c(t-1) based on the previous audio block t-1. The previous output value h(t-1) and the previous state c(t-1) are input into the Long Short-Term Memory network as historical audio information. The Long Short-Term Memory network has two outputs: the current output value h(t) and the current state c(t). The current output value h(t) is input into the linear classification layer for classification. The current output value h(t) and the current state c(t) are input into the Long Short-Term Memory network as the next historical audio information to assist in classifying the next audio block.

[0064] In some embodiments, before obtaining the current audio block and the previous audio block from the in-vehicle audio stream, the in-vehicle voice activation detection method also includes: training the convolution layer with the sample audio block to obtain a sample spectrogram; after the convolution layer training is completed, training the multi-head attention layer and the long short-term memory network with the sample spectrogram to obtain a preset classification model.

[0065] During the training process, the sample audio block can be input into the convolution layer, the output of the Fourier operator of the sample audio is used as the label of the convolution layer, and the mean square error loss function is used as the objective function for training. When the objective function reaches convergence, the training of the convolution layer is completed, and the sample spectrogram is obtained. The sample spectrogram is then used to train the multi-head attention layer and the long short-term memory network, making full use of the selectivity of the attention mechanism and the temporal memory characteristics of the long short-term memory network to achieve category detection of the sample audio block.

[0066] During the training of the multi-head attention layer and the long short-term memory network, the weights of the beginning and the end of the speech in the objective functions of the multi-head attention layer and the long short-term memory network can be controlled to be greater than the weights of the middle and non-speech segments.

[0067] It is understandable that there are more audio blocks as the middle segment of speech and non-speech segment than the audio blocks as the beginning and end of speech. Therefore, in the process of training the multi-head attention layer and the long short-term memory network, the sample size of the audio blocks at the beginning and end of speech is less than that of the audio blocks at the middle segment of speech and non-speech segment. At this time, when constructing the objective function, it is necessary to adjust the weights to make the beginning and end of speech have greater weights to ensure that the detection of the beginning and end of speech has significant feedback during the training process.

[0068] The specific value design of the weight is not limited in this embodiment. In some embodiments, the weights of the middle segment of the speech and the non-speech segment can be set to 0.2, and the weights of the beginning and the end of the speech can be set to 0.3.

[0069] By using a preset classification model to detect the beginning of speech, the beginning of the speech segment can be detected in advance, reducing the burr phenomenon between the speech segment and the non-speech segment, and avoiding affecting the subsequent speech processing function; by using a preset classification model to detect the end of speech, the end of the speech segment can be detected later to ensure the integrity of the entire speech segment and avoid fragmentation of the speech segment; by adjusting the weight of the objective function, the probability of false detection is significantly reduced, making the detection result more reliable.

[0070] The in-vehicle voice activation detection method in the present application determines the current audio block and the previous audio block according to the in-vehicle audio stream; obtains the historical audio information corresponding to the previous audio block; inputs the current audio block and the historical audio information into a preset classification model to obtain the category of the current audio block output by the preset classification model, wherein the categories of the current audio block include: speech start, speech middle, speech end, and non-speech segment. Among them, by adding historical audio information to the preset classification model on the basis of the current audio block, the preset classification model can more accurately classify the current audio block based on the historical audio information, thereby improving the stability and accuracy of voice activation detection in a vehicle environment. At the same time, the present application ensures the immediacy and continuity of detection by processing the real-time audio stream, and reduces the probability of false triggering and missed detection by accurately detecting the speech start and speech end.

[0071] The following describes an apparatus embodiment of the present application, which can be used to execute the vehicle-mounted voice activation detection method in the above-mentioned embodiment of the present application. For details not disclosed in the apparatus embodiment of the present application, please refer to the above-mentioned embodiment of the vehicle-mounted voice activation detection method of the present application.

[0072] Figure 4 FIG. 2 shows a block diagram of a vehicle-mounted voice activation detection device in one embodiment. Figure 4 As shown, the vehicle-mounted voice activation detection device of an embodiment of the present application includes: a current audio block acquisition module 401, a historical audio block acquisition module 402 and a current audio block classification module 403, wherein the current audio block acquisition module 401 is used to determine the current audio block and the previous audio block according to the in-vehicle audio stream; the historical audio block acquisition module 402 is used to obtain the historical audio information corresponding to the previous audio block; the current audio block classification module 403 is used to input the current audio block and the historical audio information into a preset classification model to obtain the category of the current audio block output by the preset classification model, wherein the category of the current audio block includes: speech beginning, speech middle segment, speech end and non-speech segment.

[0073] In some embodiments of the present application, based on the aforementioned scheme, the preset model includes a convolutional layer, a multi-head attention layer, a long short-term memory network and a linear classification layer connected in sequence, and the current audio block classification module 403 is also used to input the current audio block into the convolutional layer to obtain a current spectrogram; input the current spectrogram into the multi-head attention layer to obtain a current attention vector; input the current attention vector and the historical audio information into the long short-term memory network to obtain a current output value and a current state of the long short-term memory network; input the current output value into the linear classification layer to obtain a category of the current audio block.

[0074] In some embodiments of the present application, based on the aforementioned scheme, the current audio block classification module 403 is also used to input the current spectrogram into the multi-head attention layer, so that the multi-head attention layer extracts features from the current spectrogram, performs pre-classification according to the extracted features, and outputs the current attention vector corresponding to the pre-classification result.

[0075] In some embodiments of the present application, based on the aforementioned scheme, the historical audio block acquisition module 402 is further used to determine the previous output value and the previous state obtained by the long short-term memory network based on the previous audio block as the historical audio information corresponding to the previous audio block.

[0076] In some embodiments of the present application, based on the aforementioned scheme, the current audio block classification module 403 is also used to train the convolution layer using the sample audio block to obtain a sample spectrogram; after the convolution layer training is completed, the multi-head attention layer and the long short-term memory network are trained using the sample spectrogram to obtain the preset classification model.

[0077] In some embodiments of the present application, based on the aforementioned scheme, the objective function of the convolution layer is a mean square error loss function.

[0078] In some embodiments of the present application, based on the aforementioned scheme, the current audio block classification module 403 is also used to control the weights of the beginning and the end of the speech in the objective function of the multi-head attention layer and the long short-term memory network to be greater than the weights of the middle segment of the speech and the non-speech segment.

[0079] Based on the same inventive concept, the present application embodiment also provides a vehicle-mounted voice activation detection device, referring to Figure 5 , shows a schematic diagram of the structure of an in-vehicle voice activation detection device in an embodiment of the present application, wherein the in-vehicle voice activation detection device includes one or more memories 504, one or more processors 502, and at least one computer program (computer program instruction) stored in the memory 504 and executable on the processor 502, and the processor 502 implements the method described above when executing the computer program.

[0080] Among them, Figure 5In the embodiment of the present invention, a bus architecture (represented by bus 500) is shown, which may include any number of interconnected buses and bridges, and bus 500 links various circuits including one or more processors represented by processor 502 and memory represented by memory 504. Bus 500 may also link various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and are therefore not further described herein. Bus interface 505 provides an interface between bus 500 and receiver 501 and transmitter 503. Receiver 501 and transmitter 503 may be the same element, namely a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 502 is responsible for managing bus 500 and general processing, while memory 504 may be used to store data used by processor 502 when performing operations.

[0081] Based on the same inventive concept, an embodiment of the present application provides a computer-readable storage medium, in which computer program instructions are stored. When the computer program instructions are executed by a processor, the processor is prompted to implement the steps of the method as described above.

[0082] The functions described herein may be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions may be stored as one or more instructions or codes on a computer-readable medium or transmitted via a computer-readable medium. Other examples and implementations are within the scope and spirit of the present application and the appended claims. For example, due to the nature of software, the functions described above may be implemented using software executed by a processor, hardware, firmware, hard wiring, or a combination of any of these. In addition, each functional unit may be integrated into a processing unit, each unit may exist physically separately, or two or more units may be integrated into one unit.

[0083] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0084] The units described as separate components may or may not be physically separated, and the components of the control device may or may not be physical units, that is, they may be located in one place or distributed in multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0085] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly or all or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk, etc., which can store computer program instructions.

[0086] The above description is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present application shall be included in the scope of the claims of the present application.

Claims

1. A vehicle-mounted voice activation detection method, It is characterized in that include: Determine a current audio block and a previous audio block according to the in-vehicle audio stream; Acquire historical audio information corresponding to the previous audio block; The current audio block and the historical audio information are input into a preset classification model to obtain a category of the current audio block output by the preset classification model, wherein the categories of the current audio block include: a speech start end, a speech middle segment, a speech end end, and a non-speech segment.

2. The vehicle-mounted voice activation detection method according to claim 1, It is characterized in that The preset model includes a convolutional layer, a multi-head attention layer, a long short-term memory network, and a linear classification layer connected in sequence, and the current audio block and the historical audio information are input into the preset classification model to obtain the category of the current audio block output by the preset classification model, including: Inputting the current audio block into the convolutional layer to obtain a current spectrogram; Input the current spectrogram into the multi-head attention layer to obtain a current attention vector; Inputting the current attention vector and the historical audio information into the long short-term memory network to obtain a current output value and a current state of the long short-term memory network; The current output value is input into the linear classification layer to obtain the category of the current audio block.

3. The vehicle-mounted voice activation detection method according to claim 2, It is characterized in that The step of inputting the current spectrogram into the multi-head attention layer to obtain a current attention vector includes: The current spectrogram is input into the multi-head attention layer so that the multi-head attention layer extracts features from the current spectrogram, performs pre-classification based on the extracted features, and outputs a current attention vector corresponding to the pre-classification result.

4. The vehicle-mounted voice activation detection method according to claim 2, It is characterized in that The obtaining of historical audio information corresponding to the previous audio block includes: A previous output value and a previous state obtained by the long short-term memory network based on the previous audio block are determined as historical audio information corresponding to the previous audio block.

5. The vehicle-mounted voice activation detection method according to claim 2, It is characterized in that Before determining the current audio block and the previous audio block according to the in-vehicle audio stream, the method further includes: Training the convolutional layer using sample audio blocks to obtain a sample spectrogram; After the convolutional layer training is completed, the multi-head attention layer and the long short-term memory network are trained using the sample spectrogram to obtain the preset classification model.

6. The vehicle-mounted voice activation detection method according to claim 5, It is characterized in that The objective function of the convolution layer is the mean square error loss function.

7. The vehicle-mounted voice activation detection method according to claim 5, It is characterized in that The using the sample spectrogram to train the multi-head attention layer and the long short-term memory network includes: The weights of the beginning and the end of the speech in the objective function controlling the multi-head attention layer and the long short-term memory network are both greater than the weights of the middle segment of the speech and the non-speech segment.

8. A vehicle-mounted voice activation detection device, It is characterized in that include: A current audio block acquisition module, used to determine a current audio block and a previous audio block according to an in-car audio stream; A historical audio block acquisition module, used to acquire historical audio information corresponding to the previous audio block; The current audio block classification module is used to input the current audio block and the historical audio information into a preset classification model to obtain the category of the current audio block output by the preset classification model, wherein the categories of the current audio block include: speech beginning, speech middle segment, speech end and non-speech segment.

9. A vehicle-mounted voice activation detection device, comprising a processor and a memory, It is characterized in that The memory stores computer program instructions that can be executed by the processor, and when the processor executes the computer program instructions, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores computer program instructions, which, when executed by a processor, prompt the processor to implement the steps of the method according to any one of claims 1 to 7.

Images