Multi-zone vehicle-mounted voice wake-up and recognition method and system based on AI recognition

Abstract

The present invention relates to the technical field of sound zone speech recognition. The present invention discloses a multi-sound zone vehicle-mounted voice wake-up and recognition method and system based on AI recognition, comprising first performing sound zone separation processing on a first voice to be recognized acquired by a first microphone to obtain voice separation data, and calculating a separation influence value based on the voice separation data. If the separation influence value is greater than a preset threshold, subsequent optimization processing is performed; if the influence value is less than or equal to the threshold, the second voice to be recognized is corrected, and the original first voice to be recognized is replaced with an accurate voice to be recognized. Finally, based on the separation influence value and a preset intelligent optimization algorithm, the voice separation data is optimized. In this way, when one microphone may not be able to capture subtle sound details, the present invention can use another microphone to capture them, and can also optimize the voice separation data, thereby improving the accuracy of subsequent tone recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of audio zone speech recognition technology, and more specifically, to a multi-audio zone vehicle-mounted voice wake-up and recognition method and system based on AI recognition. Background Art

[0002] In-vehicle voice recognition systems are key technologies that provide drivers and passengers with a convenient, safe, and intelligent experience. With the development of smart cars and autonomous driving, in-vehicle voice recognition plays an important role in enabling human-vehicle interaction, improving driving safety, and user experience. Multi-zone voice recognition includes both the spatial position recognition of the sound source and the recognition of the sound tone. Spatial position recognition is used to determine the source of the command and distinguish whether the command comes from the driver or the passenger, thereby responding accurately and improving the accuracy of the interaction. Tone recognition is mostly used for sentiment analysis and tone judgment to help the system understand the user's emotions or intentions and enhance the user experience. Current innovations are mostly concentrated in the field of spatial position recognition, while innovations in tone recognition are relatively rare.

[0003] In the prior art, for example, a Chinese patent application with publication number CN117789747A provides a vehicle-mounted multi-zone interaction method, device, electronic device and storage medium. The patent obtains the signal received by the microphone and converts it into the short-time frequency domain, then performs echo cancellation. The sound zone affiliation of the signal is determined through sound zone separation processing, and then the signal of each sound zone is separated and converted into time domain output to determine the actual wake-up position. Finally, the signal of the wake-up position is input into the recognition engine to complete the voice interaction.

[0004] In the tone recognition process of the above-mentioned patent, the voice signal first needs to be processed for tone zone separation to achieve accurate voice recognition. However, due to external factors, such as damage to the internal module of the microphone in the car, the microphone may not be able to capture subtle sound details. This situation will lead to poor tone zone separation processing, thereby affecting the accuracy of subsequent tone recognition and failing to achieve the expected recognition effect.

[0005] In view of this, the present invention proposes a multi-zone in-vehicle voice wake-up and recognition method and system based on AI recognition to solve the above problems. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a multi-zone in-vehicle voice wake-up and recognition method and system based on AI recognition.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] First, a multi-zone in-vehicle voice wake-up and recognition method based on AI recognition is provided, including:

[0009] S10: performing voice zone separation processing on the first speech to be recognized to obtain speech separation data, obtaining separation evaluation data corresponding to the speech separation data, and calculating a separation influence value based on the separation evaluation data. When the separation influence value is greater than a preset separation influence threshold, the process proceeds to S20; when the separation influence value is less than or equal to the preset separation influence threshold, the process proceeds to S30. The first speech to be recognized is obtained by the first microphone;

[0010] S20: Correcting the second speech to be recognized to obtain an accurate speech to be recognized, replacing the first speech to be recognized based on the accurate speech to be recognized, and returning to S10, the second speech to be recognized is acquired by the second microphone, and the second speech to be recognized has the same semantics as the first speech to be recognized;

[0011] S30: Optimizing the speech separation data based on the separation influence value and a preset intelligent optimization algorithm to obtain optimal separation data, and performing recognition based on the optimal separation data.

[0012] Furthermore, the separation evaluation data includes a signal-to-noise ratio value, a sound region isolation degree, a residual interference value, and a separation error rate. The method for obtaining the sound region isolation degree includes:

[0013] ;

[0014] Where SIR is the sound zone isolation, is a logarithmic function with base 10, is the average power of the target tone range, is the average power of the rest of the sound range.

[0015] Furthermore, the method for obtaining the separation error rate includes:

[0016] ;

[0017] Where SER is the separation error rate, is the total number of separated signals, No. The real-time amplitude value of the separated signal, No. The standard amplitude value of the separated signal, is the similarity function.

[0018] Furthermore, the method for calculating the separation impact value based on the separation assessment data includes:

[0019] ;

[0020] Where SIV is the separation influence value, is the residual interference value, is the signal-to-noise ratio, Therefore The logarithmic function with base , is the hyperbolic cosine function, is the inverse cotangent function, Therefore The exponential function with base , is a natural constant.

[0021] Furthermore, the method for obtaining accurate speech to be recognized includes:

[0022] Difference data between the first microphone and the second microphone is obtained, a difference matrix is ​​constructed based on the difference data, and the difference matrix and the second speech to be recognized are input into a pre-built speech correction model to obtain an accurate speech to be recognized.

[0023] Furthermore, the method of constructing a difference matrix based on the difference data includes:

[0024] Based on the distance difference, the corresponding time delay and path attenuation rate are determined; based on the installation angle difference, the corresponding horizontal angle difference and vertical angle difference are determined; based on the sensitivity difference, the corresponding gain compensation value and frequency compensation value are determined; based on the built-in noise difference, the corresponding noise frequency difference and noise amplitude difference are determined; and the associated elements are sequentially used as a row in the difference matrix to construct the difference matrix, wherein the associated elements include distance-related elements, angle-related elements, sensitivity-related elements, and noise-related elements; the distance-related elements include distance difference, time delay, and path attenuation rate; the angle-related elements include installation angle difference, horizontal angle difference, and vertical angle difference; the sensitivity-related elements include sensitivity difference, gain compensation value, and frequency compensation value; and the noise-related elements include built-in noise difference, noise frequency difference, and noise amplitude difference.

[0025] Furthermore, the method for constructing the speech correction model includes:

[0026] Acquire a sample data set, wherein the sample data set includes a historical difference matrix, a historical second speech to be recognized, and a historical accurate speech to be recognized;

[0027] Divide the sample data set into a sample training set and a sample test set, and build a regression network;

[0028] The historical difference matrix and the historical second speech to be recognized in the sample training set are used as input data of the regression network, and the historical accurate speech to be recognized in the sample training set is used as output data of the regression network. The regression network is trained to obtain an initial regression network for predicting the accurate speech to be recognized;

[0029] The initial regression network is tested using a sample test set, and the initial regression network with an output smaller than a preset error value is used as the speech correction model.

[0030] Furthermore, the intelligent optimization algorithm is a whale optimization algorithm, which optimizes the speech separation data. The method for obtaining the optimal separation data includes:

[0031] S301: Define the initial number of whales as K and the maximum number of iterations , each initial whale individual is represented as a speech separation data, and the random value rand is a random number sampled from a uniform distribution of [0, 1];

[0032] S302: Obtain the separation influence value corresponding to the speech separation data, and use the separation influence value as the target fitness value of the initial whale individual;

[0033] S303: When the random value rand is less than 0.5, the shrinking and surrounding strategy is performed; when the random value rand is greater than or equal to 0.5, the spiral hunting strategy is performed;

[0034] S304: Repeat the above step S303 to continuously update the speech separation data until the maximum number of iterations is reached When the optimal speech separation data is output, the optimal speech separation data is used as the optimal separation data.

[0035] Furthermore, the method of implementing the shrinking and surrounding strategy includes:

[0036] ;

[0037] in, For the next generation of whale positions, is the current whale position, To chase the target location, is the shrinkage coefficient of the shrinkage and encirclement strategy, is the control factor of the shrinking and surrounding strategy, is the current iteration number.

[0038] In the second aspect, a multi-zone in-vehicle voice wake-up and recognition system based on AI recognition is provided, which is used to implement the above-mentioned multi-zone in-vehicle voice wake-up and recognition method based on AI recognition, including:

[0039] A data processing module is configured to perform voice zone separation processing on a first speech to be recognized to obtain speech separation data, obtain separation evaluation data corresponding to the speech separation data, calculate a separation influence value based on the separation evaluation data, and transfer the process to a speech adjustment module when the separation influence value is greater than a preset separation influence threshold; and transfer the process to a speech optimization module when the separation influence value is less than or equal to the preset separation influence threshold, wherein the first speech to be recognized is obtained by a first microphone;

[0040] A speech adjustment module is configured to correct the second speech to be recognized to obtain an accurate speech to be recognized, replace the first speech to be recognized based on the accurate speech to be recognized, and return the corrected speech to the data processing module. The second speech to be recognized is acquired by the second microphone and has the same semantics as the first speech to be recognized.

[0041] Speech optimization module: used to optimize speech separation data based on the separation impact value and the preset intelligent optimization algorithm, obtain the optimal separation data, and perform recognition based on the optimal separation data.

[0042] In a third aspect, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the above-mentioned multi-zone in-vehicle voice wake-up and recognition method based on AI recognition is implemented.

[0043] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed, the above-mentioned multi-zone in-vehicle voice wake-up and recognition method based on AI recognition is implemented.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] The present invention first performs sound zone separation processing on the first speech to be recognized acquired by the first microphone to obtain speech separation data, and calculates a separation influence value based on the speech separation data. If the separation influence value is greater than a preset threshold, subsequent optimization processing is performed; if the influence value is less than or equal to the threshold, the second speech to be recognized is corrected, and the original first speech to be recognized is replaced with the accurate speech to be recognized. Finally, based on the separation influence value and a preset intelligent optimization algorithm, the speech separation data is optimized. In this way, when one microphone may not be able to capture subtle sound details, the present invention can use another microphone to capture them, and can also optimize the speech separation data, thereby improving the accuracy of subsequent tone recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of the multi-zone in-vehicle voice wake-up and recognition method based on AI recognition in the present invention;

[0047] Figure 2 This is a schematic diagram of the structure of the multi-zone vehicle-mounted voice wake-up and recognition system based on AI recognition in the present invention;

[0048] Figure 3 A flow chart of the method for obtaining accurate speech to be recognized in the present invention;

[0049] Figure 4Schematic diagram of the process of constructing a difference matrix based on difference data in the present invention. DETAILED DESCRIPTION

[0050] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0051] Example 1

[0052] See also Figure 1 As shown, this embodiment discloses a multi-zone in-vehicle voice wake-up and recognition method based on AI recognition, including:

[0053] S10: performing voice zone separation processing on the first speech to be recognized to obtain speech separation data, obtaining separation evaluation data corresponding to the speech separation data, and calculating a separation influence value based on the separation evaluation data. When the separation influence value is greater than a preset separation influence threshold, the process proceeds to S20; when the separation influence value is less than or equal to the preset separation influence threshold, the process proceeds to S30. The first speech to be recognized is obtained by the first microphone;

[0054] In this embodiment, the first microphone refers to the microphone closest to the user. For example, if the user issues a voice command in the co-pilot area, the microphone located in the co-pilot area closest to the user serves as the first microphone. The voice zone separation processing refers to processing the voice signal to extract and separate the tone information in a specific area before performing tone recognition, so as to ensure the accuracy of tone recognition. That is to say, the purpose of the voice zone separation processing is to extract the tone features of the target voice zone (such as the co-pilot area where the user is located) in a multi-voice zone environment for subsequent recognition. The voice separation data represents the audio data containing clear tone features in the voice signal extracted from a specific voice zone (such as the co-pilot area where the user is located). These data are processed by the voice zone separation and are mainly used for subsequent tone recognition to ensure that the tone features are not affected by interference from other voice zones. Specifically, the voice separation data contains the voice signal information in the target voice zone, which can represent the tone features and voice content of the user who issues the voice in the voice zone.

[0055] It can be understood that the separation evaluation data represents the quantitative evaluation results of the quality of the speech separation data, and is used to measure the effect of the voice zone separation processing. The separation evaluation data includes but is not limited to the signal-to-noise ratio, voice zone isolation, residual interference value and separation error rate. The signal-to-noise ratio is used to measure the ratio of speech components to background noise in the separated speech signal. The higher the signal-to-noise ratio, the clearer the speech signal in the speech separation data and the less noise, which is conducive to subsequent tone recognition. The voice zone isolation is used to indicate the isolation effect between the target voice zone and other voice zones in the speech separation data. A higher isolation means that the interference components of other voice zones are effectively suppressed. The residual interference value is used to measure whether there are still residual interference signals from other voice zones in the speech separation data. A lower residual interference value indicates that the voice zone separation processing has effectively eliminated the influence of non-target voice zones. The separation error rate is used to indicate the degree of error that may occur in the voice zone separation process, and to evaluate whether there is distortion or loss of voice information in the speech separation data. A low separation error rate helps to ensure the quality of the separation data.

[0056] Methods for obtaining sound zone isolation include:

[0057] ;

[0058] Where SIR is the sound zone isolation, is a logarithmic function with base 10, is the average power of the target tone range, is the average power of the rest of the sound range.

[0059] It should be noted that the target sound zone refers to the sound zone where the user is located. For example, if the user issues a voice command in the co-pilot area, then the co-pilot area is the target sound zone. The system will focus on the signal processing of this sound zone and extract the user's voice information. The remaining sound zones refer to other areas in the car except the target sound zone. The signals in these areas are regarded as interference signals. The system will try to suppress the voices in these areas in the sound zone separation process to ensure that the signal in the target sound zone is clear and not interfered with. Both the average power of the target sound zone and the average power of the remaining sound zones can be obtained through sampling calculation, and this embodiment will not go into details about this.

[0060] Methods for obtaining the separation error rate include:

[0061] ;

[0062] Where SER is the separation error rate, is the total number of separated signals, No. The real-time amplitude value of the separated signal, No. The standard amplitude value of the separated signal, is the similarity function.

[0063] It can be understood that the total number of separated signals refers to the number of all independent separated signals located in the target sound area obtained by the system during the processing process, that is, the number of independent speech signals extracted from the target sound area. The standard amplitude value refers to the amplitude value of the separated signal of the target sound area recorded in a controlled environment. Although the semantic content corresponding to the real-time amplitude value and the standard amplitude value is different, the quality of the separated signal can still be effectively evaluated by comparing the similarity between the two (using the similarity function SIM). That is to say, even if the content is inconsistent, the system can still use the standard amplitude value as a benchmark to measure the quality of the real-time amplitude value, thereby calculating the separation error rate.

[0064] It should be added that the signal-to-noise ratio value mentioned above can be obtained by calculating the power of the separated speech signal and the background noise signal, and the residual interference value is the ratio of the average power of the residual interference signal in the remaining sound area to the average power of the separated signal in the target sound area.

[0065] Methods for calculating separation impact values ​​based on separation assessment data include:

[0066] ;

[0067] Where SIV is the separation influence value, is the residual interference value, is the signal-to-noise ratio, Therefore The logarithmic function with base , is the hyperbolic cosine function, is the inverse cotangent function, Therefore The exponential function with base , is a natural constant.

[0068] In this embodiment, Taking SIR as an example, a higher signal-to-noise ratio value indicates a clearer speech signal and less noise in the speech separation data, while the voice zone isolation is used to indicate the isolation effect between the target voice zone and other voice zones in the speech separation data. A higher voice zone isolation means that the interference components of other voice zones are effectively suppressed. Therefore, it can be seen from the above content that, in this embodiment, a larger separation influence value indicates a worse effect of separation processing on the speech separation data.

[0069] It is not difficult to understand that when the separation impact value is greater than the preset separation impact threshold, it indicates that the quality of the speech separation data is not enough to directly enter the subsequent recognition process, and may require further correction or reprocessing. When the separation impact value is less than or equal to the preset separation impact threshold, it means that the quality of the speech separation data has reached the expected level and can enter the optimization or recognition stage. This method ensures that only high-quality separation signals enter subsequent processing, thereby improving the accuracy and reliability of recognition.

[0070] This embodiment performs sound zone separation processing on the first speech to be recognized to obtain speech separation data and corresponding separation evaluation data, and then calculates the separation influence value based on these evaluation data. The combination of this series of steps enables the system to accurately evaluate the effect of the sound zone separation processing, ensure the high quality of the speech separation data, and judge whether to enter S20 or S30 based on the separation influence value. This judgment step enables the system to select the most suitable subsequent processing path according to the size of the separation influence value. If the separation influence value is large, it indicates that the separation effect is not good, and the system will enter S20 for correction; if the separation influence value is small, it indicates that the separation effect is good, and the system will enter S30 for optimization and recognition. In this way, the best processing plan can be adopted for different situations.

[0071] S20: Correcting the second speech to be recognized to obtain an accurate speech to be recognized, replacing the first speech to be recognized based on the accurate speech to be recognized, and returning to S10, the second speech to be recognized is acquired by the second microphone, and the second speech to be recognized has the same semantics as the first speech to be recognized;

[0072] It should be noted that the only difference between the first voice to be recognized and the second voice to be recognized is that the first voice to be recognized is obtained by the first microphone, and the second voice to be recognized is obtained by the second microphone. The semantic content of the first voice to be recognized and the second voice to be recognized are exactly the same. The first microphone refers to the microphone closest to the user, and the second microphone is the microphone closest to the user except the first microphone. It can be understood that when the separation impact value is greater than the preset separation impact threshold, it indicates that the quality of the voice separation data is poor. Generally speaking, there is a problem with the relevant module inside the first microphone, such as signal distortion or failure of the noise suppression module. Therefore, the voice data obtained from the second microphone needs to be subsequently processed. When the user issues a voice command, the first microphone and the second microphone work simultaneously.

[0073] like Figure 3 As shown, the method for obtaining accurate speech to be recognized includes:

[0074] Difference data between the first microphone and the second microphone is obtained, a difference matrix is ​​constructed based on the difference data, and the difference matrix and the second speech to be recognized are input into a pre-built speech correction model to obtain an accurate speech to be recognized.

[0075] In this embodiment, the difference data includes but is not limited to distance difference, installation angle difference, sensitivity difference and built-in noise difference. The distance difference refers to the physical distance difference between the first microphone and the second microphone. This difference will affect the time and volume of the signal reaching the microphone. The sensitivity difference refers to the difference in response intensity between the first microphone and the second microphone when receiving the signal. Sensitivity is the ability of a microphone to receive voice signals. The sensitivity difference indicates the difference in signal amplitude or power received by the two microphones under the same voice input. The built-in noise difference refers to the difference in background noise level generated by the first microphone and the second microphone themselves. Each microphone will generate a certain amount of noise in the absence of voice input. The built-in noise difference reflects the difference between the two microphones under the noise background.

[0076] like Figure 4 As shown, the method of constructing a difference matrix based on difference data includes:

[0077] Based on the distance difference, the corresponding time delay and path attenuation rate are determined; based on the installation angle difference, the corresponding horizontal angle difference and vertical angle difference are determined; based on the sensitivity difference, the corresponding gain compensation value and frequency compensation value are determined; based on the built-in noise difference, the corresponding noise frequency difference and noise amplitude difference are determined; and the associated elements are sequentially used as a row in the difference matrix to construct the difference matrix, wherein the associated elements include distance-related elements, angle-related elements, sensitivity-related elements, and noise-related elements; the distance-related elements include distance difference, time delay, and path attenuation rate; the angle-related elements include installation angle difference, horizontal angle difference, and vertical angle difference; the sensitivity-related elements include sensitivity difference, gain compensation value, and frequency compensation value; and the noise-related elements include built-in noise difference, noise frequency difference, and noise amplitude difference.

[0078] It can be understood that, taking the distance difference as an example, the distance difference is associated with the time delay and path attenuation rate. When the distance difference is larger, the corresponding time delay and path attenuation rate will also increase accordingly. The above-mentioned time delay and path attenuation rate can be obtained by looking up the table, and the same is true for the sensitivity difference and the built-in noise difference. This embodiment does not go into too much detail on this. The above-mentioned associated elements refer to associated data, for example, the distance difference, time delay and path attenuation rate are associated elements. Similarly, the installation angle difference, horizontal angle difference and vertical angle difference are associated elements.

[0079] The method for constructing the speech correction model includes:

[0080] Acquire a sample data set, wherein the sample data set includes a historical difference matrix, a historical second speech to be recognized, and a historical accurate speech to be recognized;

[0081] Divide the sample data set into a sample training set and a sample test set, and build a regression network;

[0082] The historical difference matrix and the historical second speech to be recognized in the sample training set are used as input data of the regression network, and the historical accurate speech to be recognized in the sample training set is used as output data of the regression network. The regression network is trained to obtain an initial regression network for predicting the accurate speech to be recognized;

[0083] The initial regression network is tested using a sample test set, and the initial regression network with an output smaller than a preset error value is used as the speech correction model. The initial regression network is a deep neural network model.

[0084] In this embodiment, the separation evaluation data and separation influence value obtained through the sound zone separation processing can preliminarily judge the quality of the speech separation data. However, processing only through the data of the first microphone may result in poor separation effect due to microphone module problems (such as signal distortion or failure of the noise suppression module). In this case, by entering step S20, the system obtains the second speech to be recognized from the second microphone for correction, so that the final precise speech to be recognized is clearer and more accurate, effectively compensating for the problem of poor separation effect caused by the problem of the first microphone module, and ensuring that only speech separation data that reaches the expected separation quality will enter the final recognition process. Through this adjustment, the system can use the second speech to be recognized from the second microphone to replace the first speech to be recognized from the first microphone with problems when necessary.

[0085] S30: Optimizing the speech separation data based on the separation influence value and a preset intelligent optimization algorithm to obtain optimal separation data, and performing recognition based on the optimal separation data.

[0086] The intelligent optimization algorithm can be a whale optimization algorithm. Then the methods for optimizing the speech separation data and obtaining the optimal separation data include:

[0087] S301: Define the initial number of whales as K and the maximum number of iterations , each initial whale individual is represented as a speech separation data, and the random value rand is a random number sampled from a uniform distribution of [0, 1];

[0088] It is worth noting that in this embodiment, the method of sound zone separation processing can be modified to a certain extent, so that K different voice separation data can be generated. Exemplarily, this embodiment can first perform sound zone separation processing according to standard rules to obtain voice separation data. After confirmation through the above-mentioned steps S10 and S20, when step S30 is performed, the standard rules are modified to generate K different voice separation data, and the optimal voice separation data is obtained from the K different voice separation data. The above-mentioned standard rules can be fixed parameters or strategies used for the preliminary division of sound zones during the sound zone separation process, such as microphone gain settings, filter parameters or directional parameters. Modifications to the standard rules may include adjustments to these parameters, such as changes in the filter range and fine-tuning of the directionality.

[0089] S302: Obtain the separation influence value corresponding to the speech separation data, and use the separation influence value as the target fitness value of the initial whale individual;

[0090] It can be understood that the separation influence value corresponds to the "fitness value" in the whale optimization algorithm, which represents the "survival ability" of an individual in the entire group. In the whale optimization algorithm, each individual (whale) has a fitness value to measure the quality of the individual. The smaller the fitness value, the closer the individual is to the optimal solution.

[0091] S303: When the random value rand is less than 0.5, the shrinking and surrounding strategy is performed; when the random value rand is greater than or equal to 0.5, the spiral hunting strategy is performed;

[0092] Among them, the shrinking and encircling strategy is implemented, and the whale will shrink and encircle around the optimal solution. The spiral hunting strategy is implemented to simulate the whale moving around the target in a spiral path.

[0093] S304: Repeat the above step S303 to continuously update the speech separation data until the maximum number of iterations is reached When the optimal speech separation data is output, the optimal speech separation data is used as the optimal separation data.

[0094] In this embodiment, in each iteration, the whale group is continuously updated until the maximum number of iterations is reached. After each iteration, the algorithm will update the weight factor combination of individual whales and calculate the new fitness value. Through repeated iterations, the weight factor combination of individual whales will be gradually improved each time, so that the whale group will gradually approach the optimal solution.

[0095] Methods for implementing the shrinking and surrounding strategy include:

[0096] ;

[0097] in, For the next generation of whale positions, is the current whale position, To chase the target location, is the shrinkage coefficient of the shrinkage and encirclement strategy, is the control factor of the shrinking and surrounding strategy, is the current iteration number.

[0098] This embodiment allows individual whales to gradually shrink when approaching the optimal solution, accurately adjust the weight factor, and perform local optimization. Through this strategy, individual whales gradually approach the current optimal solution, ensuring a detailed search in the solution space.

[0099] Methods for conducting a spiral hunting strategy include:

[0100] ;

[0101] in, is the spiral contraction coefficient of the spiral hunting strategy, A random value in the range [-1, 1].

[0102] In this embodiment, separation evaluation data corresponding to the speech separation data is obtained, a separation influence value is calculated based on the separation evaluation data, and the process proceeds to S20 or S30 based on the separation influence value. The second speech to be recognized is corrected to determine the precise speech to be recognized, the first speech to be recognized is replaced based on the precise speech to be recognized, and the process returns to S10. The speech separation data is optimized based on the separation influence value and the intelligent optimization algorithm to obtain the optimal separation data. In this way, in the case that one microphone may not be able to capture subtle sound details, the other microphone can be used to capture them, and the speech separation data can also be optimized, thereby improving the accuracy of subsequent tone recognition.

[0103] Example 2

[0104] See also Figure 2 As shown, based on the same inventive concept, this embodiment discloses a multi-zone vehicle-mounted voice wake-up and recognition system based on AI recognition. For details not detailed in this embodiment, please refer to the description of the relevant parts in Example 1. The system includes:

[0105] A data processing module is configured to perform voice zone separation processing on a first speech to be recognized to obtain speech separation data, obtain separation evaluation data corresponding to the speech separation data, calculate a separation influence value based on the separation evaluation data, and transfer the process to a speech adjustment module when the separation influence value is greater than a preset separation influence threshold; and transfer the process to a speech optimization module when the separation influence value is less than or equal to the preset separation influence threshold, wherein the first speech to be recognized is obtained by a first microphone;

[0106] It can be understood that the separation evaluation data represents the quantitative evaluation results of the quality of the speech separation data, and is used to measure the effect of the voice zone separation processing. The separation evaluation data includes but is not limited to the signal-to-noise ratio, voice zone isolation, residual interference value and separation error rate. The signal-to-noise ratio is used to measure the ratio of speech components to background noise in the separated speech signal. The higher the signal-to-noise ratio, the clearer the speech signal in the speech separation data and the less noise, which is conducive to subsequent tone recognition. The voice zone isolation is used to indicate the isolation effect between the target voice zone and other voice zones in the speech separation data. A higher isolation means that the interference components of other voice zones are effectively suppressed. The residual interference value is used to measure whether there are still residual interference signals from other voice zones in the speech separation data. A lower residual interference value indicates that the voice zone separation processing has effectively eliminated the influence of non-target voice zones. The separation error rate is used to indicate the degree of error that may occur in the voice zone separation process, and to evaluate whether there is distortion or loss of voice information in the speech separation data. A low separation error rate helps to ensure the quality of the separation data.

[0107] Methods for obtaining sound zone isolation include:

[0108] ;

[0109] Where SIR is the sound zone isolation, is a logarithmic function with base 10, is the average power of the target tone range, is the average power of the rest of the range.

[0110] Methods for obtaining the separation error rate include:

[0111] ;

[0112] Where SER is the separation error rate, is the total number of separated signals, No. The real-time amplitude value of the separated signal, No. The standard amplitude value of the separated signal, is the similarity function.

[0113] Methods for calculating separation impact values ​​based on separation assessment data include:

[0114] ;

[0115] Where SIV is the separation influence value, is the residual interference value, is the signal-to-noise ratio, Therefore The logarithmic function with base , is the hyperbolic cosine function, is the inverse cotangent function, Therefore The exponential function with base , is a natural constant.

[0116] The method for determining whether to transfer to a voice adjustment module or a voice optimization module based on the separation impact value includes:

[0117] When the separation impact value is greater than the preset separation impact threshold, the system enters the voice adjustment module; when the separation impact value is less than or equal to the preset separation impact threshold, the system enters the voice optimization module.

[0118] A speech adjustment module is configured to correct the second speech to be recognized to obtain an accurate speech to be recognized, replace the first speech to be recognized based on the accurate speech to be recognized, and return the corrected speech to the data processing module. The second speech to be recognized is acquired by the second microphone and has the same semantics as the first speech to be recognized.

[0119] Methods for obtaining accurate speech to be recognized include:

[0120] Difference data between the first microphone and the second microphone is obtained, a difference matrix is ​​constructed based on the difference data, and the difference matrix and the second speech to be recognized are input into a pre-built speech correction model to obtain an accurate speech to be recognized.

[0121] Methods for constructing a difference matrix based on difference data include:

[0122] Based on the distance difference, the corresponding time delay and path attenuation rate are determined; based on the installation angle difference, the corresponding horizontal angle difference and vertical angle difference are determined; based on the sensitivity difference, the corresponding gain compensation value and frequency compensation value are determined; based on the built-in noise difference, the corresponding noise frequency difference and noise amplitude difference are determined; and the above-mentioned associated elements are sequentially used as a row in the difference matrix to construct the difference matrix, wherein the associated elements include distance-related elements, angle-related elements, sensitivity-related elements and noise-related elements; the distance-related elements include distance difference, time delay and path attenuation rate; the angle-related elements include installation angle difference, horizontal angle difference and vertical angle difference; the sensitivity-related elements include sensitivity difference, gain compensation value and frequency compensation value; and the noise-related elements include built-in noise difference, noise frequency difference and noise amplitude difference.

[0123] The method for constructing the speech correction model includes:

[0124] Acquire a sample data set, wherein the sample data set includes a historical difference matrix, a historical second speech to be recognized, and a historical accurate speech to be recognized;

[0125] Divide the sample data set into a sample training set and a sample test set, and build a regression network;

[0126] The historical difference matrix and the historical second speech to be recognized in the sample training set are used as input data of the regression network, and the historical accurate speech to be recognized in the sample training set is used as output data of the regression network. The regression network is trained to obtain an initial regression network for predicting the accurate speech to be recognized;

[0127] The initial regression network is tested using a sample test set, and the initial regression network with an output smaller than a preset error value is used as the speech correction model. The initial regression network is a deep neural network model.

[0128] Speech optimization module: used to optimize speech separation data based on the separation impact value and the preset intelligent optimization algorithm, obtain the optimal separation data, and perform recognition based on the optimal separation data.

[0129] Example 3

[0130] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the multi-zone in-vehicle voice wake-up and recognition method based on AI recognition provided by the above methods is implemented.

[0131] Since the electronic device introduced in this embodiment is an electronic device used to implement the multi-zone vehicle-mounted voice wake-up and recognition method based on AI recognition in the embodiment of this application, based on the multi-zone vehicle-mounted voice wake-up and recognition method based on AI recognition introduced in the embodiment of this application, technical personnel in this field can understand the specific implementation method of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application will not be introduced in detail here. As long as technical personnel in this field implement the electronic device used by the multi-zone vehicle-mounted voice wake-up and recognition method based on AI recognition in the embodiment of this application, it falls within the scope of protection of this application.

[0132] Example 4

[0133] This embodiment discloses a computer-readable storage medium having a computer program stored thereon. When the computer program is executed, the multi-zone vehicle-mounted voice wake-up and recognition method based on AI recognition is implemented.

[0134] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters, weights and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0135] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in accordance with the embodiments of the present invention are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless network. The computer-readable storage medium can be any available medium accessible by a computer, or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, or magnetic tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0136] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0137] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0138] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only one type. In actual implementation, there may be other division methods, such as 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 devices or units, which can be electrical, mechanical or other forms.

[0139] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0140] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0141] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

[0142] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-zone vehicle-mounted voice wake-up and recognition method based on AI recognition, characterized in that: include: S10: performing voice zone separation processing on the first speech to be recognized to obtain speech separation data, obtaining separation evaluation data corresponding to the speech separation data, and calculating a separation influence value based on the separation evaluation data. When the separation influence value is greater than a preset separation influence threshold, the process proceeds to S20; when the separation influence value is less than or equal to the preset separation influence threshold, the process proceeds to S30. The first speech to be recognized is obtained by the first microphone; S20: Correcting the second speech to be recognized to obtain an accurate speech to be recognized, replacing the first speech to be recognized based on the accurate speech to be recognized, and returning to S10, the second speech to be recognized is acquired by the second microphone, and the second speech to be recognized has the same semantics as the first speech to be recognized; S30: Optimizing the speech separation data based on the separation influence value and a preset intelligent optimization algorithm to obtain optimal separation data, and performing recognition based on the optimal separation data; The separation evaluation data includes a signal-to-noise ratio value, a sound area isolation degree, a residual interference value, and a separation error rate. The method for obtaining the sound area isolation degree includes: SIR= ; Where SIR is the sound zone isolation, is a logarithmic function with base 10, is the average power of the target tone range, is the average power of the rest of the range; The method for calculating the separation impact value based on the separation assessment data includes: USE= ; Where SIV is the separation influence value, is the residual interference value, is the signal-to-noise ratio, Therefore The logarithmic function with base , is the hyperbolic cosine function, is the inverse cotangent function, Therefore The exponential function with base , is a natural constant; Methods for obtaining the separation error rate include: SER= ; Where SER is the separation error rate, is the total number of separated signals, No. The real-time amplitude value of the separated signal, No. The standard amplitude value of the separated signal, is the similarity function.

2. The multi-zone vehicle-mounted voice wake-up and recognition method based on AI recognition according to claim 1 is characterized in that: The method for obtaining accurate speech to be recognized comprises: Difference data between the first microphone and the second microphone is obtained, a difference matrix is ​​constructed based on the difference data, and the difference matrix and the second speech to be recognized are input into a pre-built speech correction model to obtain an accurate speech to be recognized.

3. The multi-zone vehicle-mounted voice wake-up and recognition method based on AI recognition according to claim 2 is characterized in that: The method for constructing a difference matrix based on difference data includes: Based on the distance difference, the corresponding time delay and path attenuation rate are determined; based on the installation angle difference, the corresponding horizontal angle difference and vertical angle difference are determined; based on the sensitivity difference, the corresponding gain compensation value and frequency compensation value are determined; based on the built-in noise difference, the corresponding noise frequency difference and noise amplitude difference are determined; and the associated elements are sequentially used as a row in the difference matrix to construct the difference matrix, wherein the associated elements include distance-related elements, angle-related elements, sensitivity-related elements, and noise-related elements; the distance-related elements include distance difference, time delay, and path attenuation rate; the angle-related elements include installation angle difference, horizontal angle difference, and vertical angle difference; the sensitivity-related elements include sensitivity difference, gain compensation value, and frequency compensation value; and the noise-related elements include built-in noise difference, noise frequency difference, and noise amplitude difference.

4. The multi-zone vehicle-mounted voice wake-up and recognition method based on AI recognition according to claim 3 is characterized in that: The method for constructing the speech correction model includes: Acquire a sample data set, wherein the sample data set includes a historical difference matrix, a historical second speech to be recognized, and a historical accurate speech to be recognized; Divide the sample data set into a sample training set and a sample test set, and build a regression network; The historical difference matrix and the historical second speech to be recognized in the sample training set are used as input data of the regression network, and the historical accurate speech to be recognized in the sample training set is used as output data of the regression network. The regression network is trained to obtain an initial regression network for predicting the accurate speech to be recognized; The initial regression network is tested using a sample test set, and the initial regression network with an output smaller than a preset error value is used as the speech correction model.

5. The multi-zone vehicle-mounted voice wake-up and recognition method based on AI recognition according to claim 1 is characterized in that: The intelligent optimization algorithm is a whale optimization algorithm, which optimizes the speech separation data. The method for obtaining the optimal separation data includes: S301: Define the initial number of whales as K and the maximum number of iterations , each initial whale individual is represented as a speech separation data, and the random value rand is a random number sampled from a uniform distribution of [0, 1]; S302: Obtain the separation influence value corresponding to the speech separation data, and use the separation influence value as the target fitness value of the initial whale individual; S303: When the random value rand is less than 0.5, the shrinking and surrounding strategy is performed; when the random value rand is greater than or equal to 0.5, the spiral hunting strategy is performed; S304: Repeat S303 above to continuously update the speech separation data until the maximum number of iterations is reached When the optimal speech separation data is output, the optimal speech separation data is used as the optimal separation data.

6. The multi-zone vehicle-mounted voice wake-up and recognition method based on AI recognition according to claim 5 is characterized in that: The method for performing the shrinking and surrounding strategy includes: ； in, For the next generation of whale positions, is the current whale position, To chase the target location, is the shrinkage coefficient of the shrinkage and encirclement strategy, is the control factor of the shrinking and surrounding strategy, is the current iteration number.

7. A multi-zone vehicle-mounted voice wake-up and recognition system based on AI recognition, which is used to implement the multi-zone vehicle-mounted voice wake-up and recognition method based on AI recognition according to any one of claims 1 to 6, characterized in that: include: A data processing module is configured to perform voice zone separation processing on a first speech to be recognized to obtain speech separation data, obtain separation evaluation data corresponding to the speech separation data, calculate a separation influence value based on the separation evaluation data, and transfer the process to a speech adjustment module when the separation influence value is greater than a preset separation influence threshold; and transfer the process to a speech optimization module when the separation influence value is less than or equal to the preset separation influence threshold, wherein the first speech to be recognized is obtained by a first microphone; A speech adjustment module is configured to correct the second speech to be recognized to obtain an accurate speech to be recognized, replace the first speech to be recognized based on the accurate speech to be recognized, and return the corrected speech to the data processing module. The second speech to be recognized is acquired by the second microphone and has the same semantics as the first speech to be recognized. Speech optimization module: used to optimize speech separation data based on the separation impact value and the preset intelligent optimization algorithm, obtain the optimal separation data, and perform recognition based on the optimal separation data.

8. 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 computer program, the multi-zone in-vehicle voice wake-up and recognition method based on AI recognition is implemented as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed, implements the multi-zone vehicle-mounted voice wake-up and recognition method based on AI recognition according to any one of claims 1 to 6.

Images