Transmission method, system and electronic device for Bluetooth audio for speech coding

Through the predictive coding model and mask prediction converter mechanism based on semantic understanding model, combined with the audio mode and communication environment, the problem of low audio coding accuracy in the prior art is solved, and the quality and accuracy of Bluetooth audio transmission is significantly improved.

CN113990334BActive Publication Date: 2024-11-01SHENZHEN JITING ERA TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111259309.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-10-28
Publication Date
2024-11-01
Estimated Expiration
2041-10-28

AI Technical Summary

Technical Problem

The existing voice encoding methods fail to effectively consider the specific mode and communication environment of the audio signal, resulting in reduced encoding accuracy and impaired audio transmission quality, especially in Bluetooth transmission.

Method used

The encoder-mask structure of the predictive encoding model based on semantic understanding model is adopted. The speech feature vector is obtained through the training stage and combined with the decoder model. The audio mode is marked by the mask prediction converter mechanism, and the communication environment factors are combined to improve the encoding. Accuracy.

Benefits of technology

When encoding, the audio mode is considered and the communication environment is combined, which significantly improves the encoding accuracy and improves the quality and accuracy of Bluetooth audio transmission.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113990334B_ABST
    Figure CN113990334B_ABST
Patent Text Reader

Abstract

The present application discloses a method, system and electronic device for transmitting Bluetooth audio for speech coding, which can obtain the speech digital signal at the receiving end by encoding the speech digital signal at the transmitting end by using an encoder-masker structure as a predictive coding model based on a semantic understanding model, thereby incorporating communication environment factors through the encoder-decoder model, and in addition, using the transformer mechanism for mask prediction of the semantic understanding model in the above encoder-masker structure, the model of the speech digital signal can be annotated based on the overall speech digital signal, thereby reflecting the audio mode. In this way, when encoding, in addition to considering the audio mode, the representation of the communication environment is also combined to improve the accuracy of encoding.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of speech coding, and more specifically, to a method for transmitting Bluetooth audio for speech coding, a system for transmitting Bluetooth audio for speech coding, and an electronic device. Background Art

[0002] Speech coding is to encode analog speech signals and convert analog signals into digital signals, thereby reducing the transmission bit rate and performing digital transmission. The commonly used speech coding is waveform coding, which samples, quantizes, and encodes the waveform signal of analog speech in the time domain to form a digital speech signal. However, this coding method does not take into account the specific audio mode of the audio signal. For example, when encoding an exciting music mode and a music mode that relaxes emotions, it is necessary to adaptively encode based on the audio mode of the audio data. Otherwise, the accuracy of the coding will be reduced, thereby affecting the quality of audio transmission.

[0003] Furthermore, when transmitting audio via Bluetooth, the communication environment, especially the communication distance and environmental noise, must be considered. Otherwise, the audio transmission quality will be impaired, and errors in the final transmitted audio may also occur.

[0004] Therefore, a transmission scheme for speech-encoded Bluetooth audio is desired. Summary of the invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide a method, system and electronic device for transmitting Bluetooth audio for speech coding, which can obtain the speech digital signal at the receiving end by encoding the speech digital signal at the transmitting end by using an encoder-masker structure as a predictive coding model based on a semantic understanding model, thereby incorporating communication environment factors through the encoder-decoder model. In addition, the transformer mechanism for mask prediction of the semantic understanding model in the above encoder-masker structure can be used to annotate the model of the speech digital signal based on the overall speech digital signal, thereby reflecting the audio mode. In this way, when encoding, in addition to considering the audio mode, the representation of the communication environment is also combined to improve the accuracy of encoding.

[0006] According to one aspect of the present application, a method for transmitting Bluetooth audio for speech coding is provided, which comprises:

[0007] The training phase includes:

[0008] Acquire a training data set, wherein the training data set includes a speech signal sent by a sending end and a speech signal received by a real receiving end;

[0009] Passing the speech signal sent by the sending end in the training data set through a semantic understanding model as a predictive coding model to obtain a plurality of speech feature vectors, wherein each speech feature vector in the plurality of semantic feature vectors corresponds to a segment of the speech signal sent by the sending end;

[0010] Inputting the plurality of speech feature vectors into a decoder model composed of a plurality of fully connected layers to recover a decoded speech signal from a speech feature vector sequence composed of the plurality of speech feature vectors;

[0011] Training the predictive coding model and the decoder model by minimizing the difference between the decoded speech signal and the actual speech signal received by the receiving end;

[0012] Passing the plurality of speech feature vectors through the converter of the predictive coding model respectively to obtain a plurality of encoded speech feature vectors;

[0013] Based on the internal structure of the converter, a mask conversion value of each encoded speech feature vector in the plurality of encoded speech feature vectors is calculated to obtain a classification feature vector composed of a plurality of mask conversion values, wherein the mask conversion value is calculated based on a distance between two encoded speech feature vectors in the plurality of encoded speech feature vectors and a representation value of whether each encoded speech feature vector in the plurality of encoded speech feature vectors is encoded;

[0014] Using a multi-label classifier to label the classification feature vector, the classification feature vector category is used to represent the category of the plurality of encoded speech feature vectors as a whole; and

[0015] Training the converter of the predictive coding model and the multi-label classifier based on the cross entropy loss function value between the category of the classification feature vector and the true category; and

[0016] The inference phase includes:

[0017] Obtain the voice signal sent by the sender;

[0018] Inputting the speech signal sent by the transmitting end into the predictive coding model trained in the training phase and the multi-label classifier to obtain a plurality of speech feature vectors and categories corresponding to the speech signal sent by the transmitting end; and

[0019] The multiple speech feature vectors and the categories corresponding to the speech signal sent by the sending end are sent to the receiving end.

[0020] According to another aspect of the present application, a transmission system for speech-encoded Bluetooth audio is provided, comprising:

[0021] Training modules include:

[0022] A training data set acquisition unit, used to acquire a training data set, wherein the training data set includes a speech signal sent by a sending end and a speech signal received by a real receiving end;

[0023] an encoding unit, configured to obtain a plurality of speech feature vectors by passing the speech signal sent by the transmitting end in the training data set obtained by the training data set obtaining unit through a semantic understanding model as a predictive coding model, wherein each speech feature vector in the plurality of semantic feature vectors corresponds to a segment of the speech signal sent by the transmitting end;

[0024] A decoding unit, configured to input the plurality of speech feature vectors obtained by the encoding unit into a decoder model composed of a plurality of fully connected layers to recover a decoded speech signal from a speech feature vector sequence composed of the plurality of speech feature vectors obtained by the encoding unit;

[0025] a difference training unit, configured to train the predictive coding model and the decoder model by minimizing the difference between the decoded speech signal obtained by the decoding unit and the actual speech signal received by the receiving end obtained by the training data set acquisition unit;

[0026] A converter processing unit, configured to respectively pass the plurality of speech feature vectors obtained by the encoding unit through a converter of the predictive coding model obtained by the difference training unit to obtain a plurality of encoded speech feature vectors;

[0027] a mask unit, configured to calculate, based on an internal structure of the converter, a mask conversion value of each of the plurality of encoded speech feature vectors obtained by the converter processing unit to obtain a classification feature vector composed of a plurality of mask conversion values, wherein the mask conversion value is calculated based on a distance between two encoded speech feature vectors in the plurality of encoded speech feature vectors and a representation value of whether each of the plurality of encoded speech feature vectors is encoded;

[0028] a classification unit, configured to label the classification feature vector obtained by the mask unit with a category of the classification feature vector through a multi-label classifier, wherein the category of the classification feature vector is used to represent the category of the plurality of encoded speech feature vectors as a whole; and

[0029] A cross entropy loss function value training unit, used to train the converter of the predictive coding model and the multi-label classifier based on the cross entropy loss function value between the category of the classification feature vector obtained by the classification unit and the true category; and

[0030] Inference modules, including:

[0031] A voice signal acquisition unit, used to acquire a voice signal sent by a sending end; and

[0032] A feature vector and category generating unit, used for inputting the speech signal sent by the sending end obtained by the speech signal acquiring unit into the predictive coding model trained in the training phase and the multi-label classifier to obtain a plurality of speech feature vectors and categories corresponding to the speech signal sent by the sending end;

[0033] The sending unit is used to send the multiple speech feature vectors obtained by the feature vector and category generating unit and the category corresponding to the speech signal sent by the sending end obtained by the feature vector and category generating unit to the receiving end.

[0034] According to another aspect of the present application, an electronic device is provided, comprising: a processor; and a memory, in which computer program instructions are stored, and when the computer program instructions are executed by the processor, the processor executes the method for transmitting Bluetooth audio for voice encoding as described above.

[0035] According to yet another aspect of the present application, a computer-readable medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the processor executes the method for transmitting Bluetooth audio for speech encoding as described above.

[0036] According to the transmission method of Bluetooth audio for speech coding, the transmission system and electronic device for Bluetooth audio for speech coding provided by the present application, by using an encoder-masker structure as a predictive coding model based on a semantic understanding model, the speech digital signal at the transmitting end can be encoded to obtain the speech digital signal at the receiving end, thereby incorporating communication environment factors through the encoder-decoder model. In addition, the transformer mechanism for mask prediction of the semantic understanding model in the above encoder-masker structure can be used to annotate the model of the speech digital signal based on the overall speech digital signal, thereby reflecting the audio mode. In this way, when encoding, in addition to considering the audio mode, the representation of the communication environment is also combined to improve the accuracy of encoding. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0038] Figure 1 The present invention is a schematic diagram of a scenario of a method for transmitting Bluetooth audio for speech coding according to an embodiment of the present application.

[0039] Figure 2A The flowchart is a training phase of a method for transmitting Bluetooth audio for speech coding according to an embodiment of the present application.

[0040] Figure 2B The flowchart is a flowchart of the inference phase in the method for transmitting Bluetooth audio for speech coding according to an embodiment of the present application.

[0041] Figure 3A Schematic diagram of the architecture of the training phase in the method for transmitting Bluetooth audio for speech coding according to an embodiment of the present application.

[0042] Figure 3B 2 is a schematic diagram of the architecture of the inference stage in the method for transmitting Bluetooth audio for speech coding according to an embodiment of the present application.

[0043] Figure 4 The block diagram is a system for transmitting Bluetooth audio for speech coding according to an embodiment of the present application.

[0044] Figure 5 is a block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0045] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0046] Scenario Overview

[0047] As mentioned above, speech coding is to encode analog speech signals and convert analog signals into digital signals, thereby reducing the transmission bit rate and performing digital transmission. The commonly used speech coding is waveform coding, which samples, quantizes, and encodes the waveform signal of analog speech in the time domain to form a digital speech signal. However, this coding method does not take into account the specific audio mode of the audio signal. For example, when encoding an exciting music mode and a music mode that relaxes emotions, it is necessary to adaptively encode based on the audio mode of the audio data. Otherwise, the accuracy of the coding will be reduced, thereby affecting the quality of audio transmission.

[0048] Furthermore, when transmitting audio via Bluetooth, the communication environment, especially the communication distance and environmental noise, needs to be considered. Otherwise, the transmission quality of the audio will be impaired, and errors in the final transmitted audio may also be caused. Therefore, a transmission scheme for Bluetooth audio with speech coding is desired.

[0049] Based on this, in the technical solution of the present application, an encoder-decoder model based on a deep neural network is used to re-encode the digitized speech signal. That is, first, a semantic understanding model such as the Bert model is used as a predictive coding model to obtain multiple speech feature vectors from the digitized speech signal sent by the transmitter in the training set, wherein each of the multiple speech feature vectors corresponds to a segment of speech signal, such as a speech signal of a word. Then, the digitized speech signal is restored from the multiple speech feature vector sequences through a decoder model composed of multiple fully connected layers, and the predictive coding model and the decoder model are trained by minimizing the difference between the restored speech signal and the speech signal received by the real receiving end in the training set, thereby obtaining an encoder model capable of feature encoding based on the audio communication environment.

[0050] Furthermore, in order to reflect the audio pattern in the encoded speech, it is necessary to label the multiple speech feature vectors as a whole as described above. Therefore, the transformer for mask prediction contained in the semantic understanding model based on, for example, the Bert model can label the contextual relationship of multiple speech feature vectors, which is also in line with the actual situation, because the speech pattern, such as the emotional pattern, needs to be expressed as a whole sentence or paragraph, rather than through a single word. Therefore, the mask conversion value of each speech feature vector is calculated based on the intrinsic structure of the converter. For example, multiple speech feature vectors are recorded as v i (i=1,…,n), then:

[0051]

[0052] where v j is a sequence of multiple feature vectors v i Eigenvectors other than j=1,…,n,j≠i (·) means for all eigenvectors v j Sum, Indicates v i With v j The distance between them, and R represents v i Whether it is masked during the predictive coding process, it can take positive or negative values ​​with equal absolute values ​​depending on the situation.

[0053] In this way, after obtaining the mask conversion value of each speech feature vector, the mask conversion value is constructed into a classification vector, and the multiple speech feature vectors are labeled as the category of the whole through a multi-label classifier.

[0054] Based on this, the present application proposes a method for transmitting Bluetooth audio for speech coding, which includes: a training phase and an inference phase. Among them, the training phase includes the steps of: obtaining a training data set, the training data set includes a speech signal sent by a transmitter and a speech signal received by a real receiver; passing the speech signal sent by the transmitter in the training data set through a semantic understanding model as a predictive coding model to obtain multiple speech feature vectors, wherein each speech feature vector in the multiple semantic feature vectors corresponds to a segment of a speech signal in the speech signal sent by the transmitter; inputting the multiple speech feature vectors into a decoder model composed of multiple fully connected layers to recover a decoded speech signal from a speech feature vector sequence composed of the multiple speech feature vectors; training the predictive coding model and the decoder model by minimizing the difference between the decoded speech signal and the speech signal received by the real receiver; passing the multiple speech feature vectors through the conversion of the predictive coding model respectively; A converter is used to obtain a plurality of encoded speech feature vectors; based on the internal structure of the converter, a mask conversion value of each encoded speech feature vector in the plurality of encoded speech feature vectors is calculated to obtain a classification feature vector composed of a plurality of mask conversion values, wherein the mask conversion value is calculated based on a distance between two encoded speech feature vectors in the plurality of encoded speech feature vectors and a representation value of whether each encoded speech feature vector in the plurality of encoded speech feature vectors is encoded; the classification feature vector is labeled with a category of the classification feature vector through a multi-label classifier, wherein the category of the classification feature vector is used to represent the category of the plurality of encoded speech feature vectors as a whole; and the converter of the predictive coding model and the multi-label classifier are trained based on a cross entropy loss function value between the category of the classification feature vector and the true category. Among them, the inference stage includes the steps of: obtaining a speech signal sent by a transmitter; inputting the speech signal sent by the transmitter into the predictive coding model and the multi-label classifier trained in the training stage to obtain multiple speech feature vectors and categories corresponding to the speech signal sent by the transmitter; and sending the multiple speech feature vectors and categories corresponding to the speech signal sent by the transmitter to a receiving end.

[0055] Figure 1 FIG. 1 is a schematic diagram of a scenario of a method for transmitting Bluetooth audio for speech coding according to an embodiment of the present application. Figure 1 As shown, in the training phase of this application scenario, first, from a terminal device (for example, Figure 1The transmitting end (eg, T1 as shown in FIG. Figure 1 ) and another terminal device (e.g., Figure 1 The receiving end (eg, as shown in T2) of Figure 1 P2) in the figure obtains a training data set, wherein the training data set includes a speech signal sent by a transmitting end and a speech signal received by a real receiving end; then, the training data set is input to a server (for example, Figure 1 S) as shown in , wherein the server is capable of training the converter of the predictive coding model and the multi-label classifier for the transmission of speech-coded Bluetooth audio using the training data set for the transmission algorithm of speech-coded Bluetooth audio.

[0056] After training is completed, in the inference phase, first, from the terminal device (e.g. Figure 1 The sending end (eg, as shown in T) of Figure 1 Then, the acquired transmitted voice signal is input to a server (for example, a server configured with a Bluetooth audio transmission algorithm for voice coding) having deployed therein. Figure 1 In the S) shown in the figure, the server can process the transmitted voice signal with a transmission algorithm of Bluetooth audio for voice coding to generate a plurality of voice feature vectors and categories corresponding to the voice signal transmitted by the transmitting end. Then, the plurality of voice feature vectors and categories corresponding to the voice signal transmitted by the transmitting end are transmitted to the receiving end.

[0057] In this application scenario, the terminal devices include but are not limited to smart phones, tablets, Bluetooth speakers, smart bracelets, sports Bluetooth headsets, etc.

[0058] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0059] Exemplary Methods

[0060] Figure 2A FIG. 1 is a flowchart of the training phase in the method for transmitting Bluetooth audio for speech coding according to an embodiment of the present application. Figure 2AAs shown, according to the embodiment of the present application, the method for transmitting Bluetooth audio for speech coding includes: a training phase, including the steps of: S110, obtaining a training data set, wherein the training data set includes a speech signal sent by a transmitter and a speech signal received by a real receiver; S120, passing the speech signal sent by the transmitter in the training data set through a semantic understanding model as a predictive coding model to obtain a plurality of speech feature vectors, wherein each speech feature vector in the plurality of semantic feature vectors corresponds to a segment of a speech signal in the speech signal sent by the transmitter; S130, inputting the plurality of speech feature vectors into a decoder model composed of a plurality of fully connected layers to recover a decoded speech signal from a speech feature vector sequence composed of the plurality of speech feature vectors; S140, training the predictive coding model and the decoder model by minimizing the difference between the decoded speech signal and the speech signal received by the real receiver; S150, inputting the plurality of speech feature vectors into a decoder model composed of a plurality of fully connected layers to recover a decoded speech signal from a speech feature vector sequence composed of the plurality of speech feature vectors; The speech feature vectors are respectively passed through the converter of the predictive coding model to obtain multiple encoded speech feature vectors; S160, based on the internal structure of the converter, the mask conversion value of each encoded speech feature vector in the multiple encoded speech feature vectors is calculated to obtain a classification feature vector composed of multiple mask conversion values, wherein the mask conversion value is calculated based on the distance between two encoded speech feature vectors in the multiple encoded speech feature vectors and the representation value of whether each encoded speech feature vector in the multiple encoded speech feature vectors is encoded; S170, the classification feature vector is passed through a multi-label classifier to label the category of the classification feature vector, wherein the category of the classification feature vector is used to represent the category of the multiple encoded speech feature vectors as a whole; and, S180, the converter of the predictive coding model and the multi-label classifier are trained based on the cross entropy loss function value between the category of the classification feature vector and the true category.

[0061] Figure 2B FIG. 1 is a flowchart of the inference phase in the method for transmitting Bluetooth audio for speech coding according to an embodiment of the present application. Figure 2B As shown, according to the transmission method of Bluetooth audio for speech coding according to the embodiment of the present application, it also includes: an inference stage, including the steps of: S210, obtaining a speech signal sent by a transmitter; S220, inputting the speech signal sent by the transmitter into the predictive coding model and the multi-label classifier trained in the training stage to obtain multiple speech feature vectors and categories corresponding to the speech signal sent by the transmitter; and, S230, sending the multiple speech feature vectors and categories corresponding to the speech signal sent by the transmitter to a receiving end.

[0062] Figure 3AFIG. 1 is a schematic diagram of the architecture of the training phase in the method for transmitting Bluetooth audio for speech coding according to an embodiment of the present application. Figure 3A As shown, in the training phase, in the network architecture, first, the obtained training data set (for example, Figure 3A The voice signal sent by the transmitting end in the IN shown in FIG. Figure 3A IN1) is represented by a semantic understanding model as a predictive coding model (e.g., Figure 3A SUM as shown in ) to obtain multiple speech feature vectors (for example, Figure 3A Then, the plurality of speech feature vectors are input into a decoder model composed of a plurality of fully connected layers (e.g., Figure 3A D) is used to recover a decoded speech signal (for example, as shown in Figure 3A Then, by minimizing the decoded speech signal and the speech signal received by the actual receiving end (for example, as Figure 3A IN2) shown in the figure is used to train the predictive coding model and the decoder model; then, the plurality of speech feature vectors are respectively passed through the converter of the predictive coding model (for example, Figure 3A C) to obtain a plurality of encoded speech feature vectors (e.g., Figure 3A Then, based on the intrinsic structure of the converter, the mask conversion value of each encoded speech feature vector in the plurality of encoded speech feature vectors is calculated to obtain a classification feature vector composed of a plurality of mask conversion values ​​(for example, Figure 3A Then, the classification feature vector is passed through a multi-label classifier (e.g., Figure 3A The circle S shown in FIG. 1 is used to label the category of the classification feature vector; and finally, the converter of the predictive coding model and the multi-label classifier are trained based on the cross entropy loss function value between the category of the classification feature vector and the true category.

[0063] Figure 3B FIG. 1 is a schematic diagram of the architecture of the inference phase in the method for transmitting Bluetooth audio for speech coding according to an embodiment of the present application. Figure 3B As shown, in the inference stage, in the network structure, first, the acquired speech signal sent by the sending end (for example, Figure 3B IN3) inputs the predictive coding model trained in the training phase (for example, Figure 3B CO) and the multi-label classifier (e.g., Figure 3B) to obtain a plurality of speech feature vectors (e.g., Figure 3B VF3) and the category corresponding to the voice signal sent by the sending end (for example, Figure 3B CA shown in FIG); then, the multiple speech feature vectors and the categories corresponding to the speech signal sent by the sending end are sent to the receiving end.

[0064] More specifically, in the training phase, in step S110, a training data set is obtained, and the training data set includes a voice signal sent by the transmitter and a voice signal received by the real receiver. As mentioned above, when performing voice encoding on Bluetooth audio, it is necessary to consider the audio mode, such as the emotional mode and the communication environment of the audio, so that information about the audio mode and the communication environment can be reflected in the encoded speech. Therefore, in the technical solution of the present application, an encoder-masker structure as a predictive coding model based on a semantic understanding model is used to encode the voice digital signal of the transmitter and train with the voice digital signal of the receiver as the true value, and here it is first necessary to obtain a training data set, wherein the training data set includes a voice signal sent by the transmitter and a voice signal received by the real receiver.

[0065] Specifically, in the embodiment of the present application, a voice signal sent by a transmitting end of a terminal device can be obtained from the transmitting end of the terminal device, and a voice signal received by a real receiving end of another terminal device can be obtained from the receiving end of the terminal device. It is worth mentioning that here, the terminal device includes but is not limited to a smart phone, a tablet, a Bluetooth speaker, a smart bracelet, a sports Bluetooth headset, etc.

[0066] More specifically, in the training phase, in step S120, the speech signal sent by the transmitter in the training data set is passed through a semantic understanding model as a predictive coding model to obtain a plurality of speech feature vectors, wherein each of the plurality of semantic feature vectors corresponds to a segment of the speech signal sent by the transmitter. It should be understood that in the technical solution of the present application, an encoder-decoder model based on a deep neural network is used to re-encode the digitized speech signal. That is, first, a semantic understanding model such as a Bert model is used as a predictive coding model to obtain a plurality of speech feature vectors from the digitized speech signal sent by the transmitter in the training set, wherein each of the plurality of speech feature vectors corresponds to a segment of the speech signal, such as the speech signal of a word.

[0067] Specifically, in an embodiment of the present application, the process of obtaining a plurality of speech feature vectors by passing the speech signal sent by the transmitting end in the training data set through a semantic understanding model as a predictive coding model includes: first, passing the speech signal sent by the transmitting end through an analog-to-digital converter to convert the speech signal sent by the transmitting end into a digital signal, so as to facilitate subsequent computer processing thereof. That is, analog-to-digital conversion is performed on the obtained speech signal sent by the transmitting end to convert the analog signal into a digital signal, specifically, it includes two steps: sampling and quantization, that is, converting the continuous waveform of the sound into discrete data points at a certain sampling rate and sampling bit number.

[0068] Next, a spectrum analyzer is used to perform spectrum analysis on the speech signal sent by the digitized transmitting end, so as to divide the speech signal sent by the digitized transmitting end into a plurality of sub-speech signals. It should be understood that sound is an analog signal, and the time domain waveform of sound only represents the relationship between the sound pressure and time, and cannot well represent the characteristics of sound. Therefore, in the technical solution of the present application, the sound waveform must be subjected to audio analysis to convert it into an acoustic feature vector. That is, in a specific example, first, a discrete Fourier transform is performed on the speech signal sent by the digitized transmitting end to convert the obtained time domain characteristics of the speech signal sent by the digitized transmitting end into the spectrum frequency domain. It should be understood that the discrete Fourier transform can extract the information of the discrete frequency band spectrum from a discrete signal. Then, the digitized speech signal sent by the transmitting end after the discrete Fourier transform is subjected to Mel filtering. It should be understood that the sensitivity of human hearing to different frequency bands is different. The human ear is not as sensitive to high frequencies as low frequencies. This dividing line is about 1000Hz. Therefore, simulating the nature of human hearing when extracting sound features can improve recognition performance. It is worth mentioning that here, the corresponding relationship between frequency (unit Hz) and Mel scale is linear below 1000Hz and logarithmic above 1000Hz. The calculation formula is as follows: Mel (f) = 1127ln (1 + f / 700). Finally, the speech signal sent by the transmitting end after Mel filtering is segmented to obtain the multiple sub-speech signals.

[0069] Then, the plurality of sub-speech signals are processed respectively using a parameter transformer to obtain the plurality of sub-speech vectors. Finally, the plurality of sub-speech vectors are passed through the semantic understanding model to obtain the plurality of speech feature vectors. It is worth mentioning that, here, each of the plurality of speech feature vectors corresponds to a segment of speech signal, such as a speech signal of a word.

[0070] More specifically, in the training phase, in step S130 and step S140, the multiple speech feature vectors are input into a decoder model composed of multiple fully connected layers to recover a decoded speech signal from a speech feature vector sequence composed of the multiple speech feature vectors, and the predictive coding model and the decoder model are trained by minimizing the difference between the decoded speech signal and the speech signal received by the real receiving end. That is, first, the multiple semantic feature vectors are composed into a semantic feature vector sequence for subsequent processing; then, the semantic feature vector sequence is passed through a decoder model composed of multiple fully connected layers to recover a digitized speech signal from the multiple speech feature vector sequences; finally, the predictive coding model and the decoder model are trained by minimizing the difference between the recovered decoded speech signal and the speech signal received by the real receiving end in the training set, thereby obtaining an encoder model capable of feature encoding based on an audio communication environment. In a specific example, the cross entropy loss function value between the decoded speech signal and the speech signal received by the real receiving end can be calculated as the difference between the two. It should be understood that the cross entropy loss function value represents the degree of consistency between the decoded speech signal and the speech signal received by the actual receiving end.

[0071] More specifically, in the training phase, in step S150 and step S160, the multiple speech feature vectors are respectively passed through the converter of the predictive coding model to obtain multiple encoded speech feature vectors, and based on the internal structure of the converter, the mask conversion value of each encoded speech feature vector in the multiple encoded speech feature vectors is calculated to obtain a classification feature vector composed of multiple mask conversion values. It should be understood that in order to be able to reflect the pattern of audio in the encoded speech, it is necessary to label the multiple speech feature vectors as a whole. And in actual scenarios, the pattern of the speech, such as the emotional pattern, needs to be expressed as a whole sentence or a whole paragraph, rather than by a single word. Therefore, in the technical solution of the present application, the transformer for mask prediction contained in the semantic understanding model based on, for example, the Bert model is used to annotate the contextual relationship of the multiple speech feature vectors. Prior to this, that is, first, the multiple speech feature vectors need to be respectively passed through the converter of the predictive coding model to obtain multiple encoded speech feature vectors. Then, based on the internal structure of the converter, a mask conversion value of each of the plurality of encoded speech feature vectors is calculated to obtain a classification feature vector composed of a plurality of mask conversion values. In a specific example, the mask conversion value is calculated based on a distance between two encoded speech feature vectors in the plurality of encoded speech feature vectors and a characterization value of whether each encoded speech feature vector in the plurality of encoded speech feature vectors is encoded.

[0072] Specifically, in an embodiment of the present application, based on the intrinsic structure of the converter, the process of calculating the mask conversion value of each of the multiple encoded speech feature vectors to obtain a classification feature vector composed of multiple mask conversion values ​​includes: based on the intrinsic structure of the converter, calculating the mask conversion value of each of the multiple encoded speech feature vectors according to the following formula to obtain a classification feature vector composed of multiple mask conversion values;

[0073] The formula is:

[0074]

[0075] where v j is the number of coded speech feature vectors except v i The coded speech feature vector other than j=1,…,n,j≠i (·) represents the speech feature vector v for all the codes j Sum, Indicates v i With v j The distance between them, and R represents v iWhether it is masked during the predictive coding process, it can take positive or negative values ​​with equal absolute values ​​depending on the situation.

[0076] More specifically, in the training phase, in step S170 and step S180, the classification feature vector is annotated with the category of the classification feature vector by a multi-label classifier, wherein the category of the classification feature vector is used to represent the category of the multiple encoded speech feature vectors as a whole, and the converter of the predictive coding model and the multi-label classifier are trained based on the cross entropy loss function value between the category of the classification feature vector and the true category. That is, in a specific example, first, the classification feature vector is input into the Softmax classification function of the multi-label classifier to obtain the probability values ​​of the classification feature vector belonging to each category respectively. It is worth mentioning that here, the category of the classification feature vector is used to represent the category of the multiple encoded speech feature vectors as a whole. Then, the category corresponding to the largest one in the probability value is determined as the category to which the classification feature vector belongs. Finally, the converter of the predictive coding model and the multi-label classifier are trained based on the cross entropy loss function value between the category of the classification feature vector and the true category. It should be understood that the cross entropy loss function value represents the degree of consistency between the category of the classification feature vector and the true category.

[0077] After the training is completed, the inference phase is entered. That is, after the converter of the predictive coding model and the multi-label classifier are trained using the transmission algorithm of the Bluetooth audio for speech coding, the trained converter of the predictive coding model and the multi-label classifier are used in an actual inference scenario.

[0078] More specifically, in the inference stage, in step S210, step S220 and step S230, the voice signal sent by the transmitter is obtained, and the voice signal sent by the transmitter is input into the predictive coding model and the multi-label classifier trained in the training stage to obtain multiple voice feature vectors and the categories corresponding to the voice signal sent by the transmitter, and then the multiple voice feature vectors and the categories corresponding to the voice signal sent by the transmitter are sent to the receiving end. That is, in a specific example, first, the voice signal sent by the transmitter can be obtained from the terminal device. It is worth mentioning that the terminal device here includes but is not limited to smart phones, tablets, Bluetooth speakers, and smart bracelets. Then, the voice signal sent by the transmitter is processed by the predictive coding model and the multi-label classifier trained in the training stage to generate the multiple voice feature vectors and the categories corresponding to the voice signal sent by the transmitter. Finally, the obtained multiple voice feature vectors and the categories corresponding to the voice signal sent by the transmitter are sent to the receiving end.

[0079] In summary, the transmission method of Bluetooth audio for speech coding based on the embodiment of the present application is explained, which can obtain the speech digital signal at the receiving end by encoding the speech digital signal at the transmitting end by using the encoder-masker structure as a predictive coding model based on the semantic understanding model, thereby incorporating the communication environment factors through the encoder-decoder model. In addition, the transformer mechanism for mask prediction of the semantic understanding model in the above encoder-masker structure can be used to annotate the model of the speech digital signal based on the overall speech digital signal, thereby reflecting the audio mode. In this way, when encoding, in addition to considering the audio mode, the representation of the communication environment is also combined to improve the accuracy of encoding.

[0080] Exemplary Systems

[0081] Figure 4 FIG. 1 is a block diagram of a transmission system for Bluetooth audio of speech coding according to an embodiment of the present application. Figure 4 As shown, the transmission system 400 for speech-encoded Bluetooth audio according to an embodiment of the present application includes: a training module 410 and an inference module 420 .

[0082] like Figure 4As shown, the training module 410 includes: a training data set acquisition unit 411, which is used to acquire a training data set, and the training data set includes a speech signal sent by the transmitter and a real speech signal received by the receiver; an encoding unit 412, which is used to obtain a plurality of speech feature vectors from the speech signal sent by the transmitter in the training data set obtained by the training data set acquisition unit 411 through a semantic understanding model as a predictive coding model, wherein each speech feature vector in the plurality of semantic feature vectors corresponds to a segment of a speech signal in the speech signal sent by the transmitter; a decoding unit 413, which is used to input the plurality of speech feature vectors obtained by the encoding unit 412 into a decoder model composed of a plurality of fully connected layers to recover a decoded speech signal from a speech feature vector sequence composed of the plurality of speech feature vectors obtained by the encoding unit 412; a difference training unit 414, which is used to train the predictive coding model and the decoder model by minimizing the difference between the decoded speech signal obtained by the decoding unit 413 and the real speech signal received by the receiver obtained by the training data set acquisition unit 411; a converter processing unit 415, which is used to convert the speech signal obtained by the encoding unit 41 2, respectively, through the converter of the predictive coding model obtained by the difference training unit 414 to obtain a plurality of encoded speech feature vectors; a mask unit 416, used to calculate the mask conversion value of each of the plurality of encoded speech feature vectors obtained by the converter processing unit 415 based on the internal structure of the converter to obtain a classification feature vector composed of a plurality of mask conversion values, wherein the mask conversion value is calculated based on the distance between two encoded speech feature vectors in the plurality of encoded speech feature vectors and the representation value of whether each encoded speech feature vector in the plurality of encoded speech feature vectors is encoded; a classification unit 417, used to annotate the classification feature vector obtained by the mask unit 416 with the category of the classification feature vector through a multi-label classifier, wherein the category of the classification feature vector is used to represent the category of the plurality of encoded speech feature vectors as a whole; and a cross entropy loss function value training unit 418, used to train the converter of the predictive coding model and the multi-label classifier based on the cross entropy loss function value between the category of the classification feature vector obtained by the classification unit 417 and the true category.

[0083] like Figure 4As shown, the inference module 420 includes: a speech signal acquisition unit 421, used to acquire the speech signal sent by the transmitter; a feature vector and category generation unit 422, used to input the speech signal sent by the transmitter obtained by the speech signal acquisition unit 421 into the predictive coding model and the multi-label classifier trained in the training stage to obtain multiple speech feature vectors and categories corresponding to the speech signal sent by the transmitter; and a sending unit 423, used to send the multiple speech feature vectors obtained by the feature vector and category generation unit 422 and the categories corresponding to the speech signal sent by the transmitter obtained by the feature vector and category generation unit to the receiving end.

[0084] In one example, in the above-mentioned Bluetooth audio transmission system 400 for speech encoding, the encoding unit 412 includes: a conversion subunit, which is used to convert the speech signal sent by the transmitter into a digital signal through an analog-to-digital converter; a spectrum analysis subunit, which is used to use a spectrum analyzer to perform spectrum analysis on the digitized speech signal sent by the transmitter obtained by the conversion subunit, so as to divide the digitized speech signal sent by the transmitter into multiple sub-speech signals; a parameter converter processing subunit, which is used to use a parameter converter to process the multiple sub-speech signals obtained by the spectrum analysis subunit respectively to obtain the multiple sub-speech vectors; and a semantic understanding model processing subunit, which is used to pass the multiple sub-speech vectors obtained by the parameter converter processing subunit through the semantic understanding model to obtain the multiple speech feature vectors.

[0085] In one example, in the above-mentioned Bluetooth audio transmission system 400 for voice encoding, the spectrum analysis subunit is further used to: perform discrete Fourier transform on the digitized voice signal sent by the transmitter; perform Mel filtering on the digitized voice signal sent by the transmitter after discrete Fourier transform; and, segment the voice signal sent by the transmitter after Mel filtering to obtain the multiple sub-voice signals.

[0086] In one example, in the above-mentioned Bluetooth audio transmission system 400 for speech coding, the difference training unit 414 is further used to: calculate the cross entropy loss function value between the decoded speech signal and the speech signal received by the actual receiving end as the difference between the two.

[0087] In one example, in the above-mentioned Bluetooth audio transmission system 400 for speech coding, the mask unit 416 is further used to: based on the inherent structure of the converter, calculate the mask conversion value of each encoded speech feature vector in the multiple encoded speech feature vectors according to the following formula to obtain a classification feature vector composed of multiple mask conversion values; the formula is:

[0088]

[0089] where v j is the number of coded speech feature vectors except v i The coded speech feature vector other than j=1,…,n,j≠i (·) represents the speech feature vector v for all the codes j Sum, Indicates v i With v j The distance between them, and R represents v i Whether to be masked during predictive coding.

[0090] In one example, in the above-mentioned Bluetooth audio transmission system 400 for speech encoding, the classification unit 417 is further used to: input the classification feature vector into the Softmax classification function of the multi-label classifier to obtain the probability values ​​of the classification feature vector belonging to each category respectively; and determine the category corresponding to the largest one among the probability values ​​as the category to which the classification feature vector belongs.

[0091] Here, those skilled in the art will appreciate that the specific functions and operations of the various units and modules in the above-mentioned Bluetooth audio transmission system 400 for speech coding have been described in detail above. Figures 1 to 3B The description of the transmission method of Bluetooth audio for speech encoding has been described in detail, and therefore, its repeated description will be omitted.

[0092] As described above, the transmission system 400 for Bluetooth audio for voice coding according to the embodiment of the present application can be implemented in various terminal devices, such as a server for the transmission algorithm of Bluetooth audio for voice coding, etc. In one example, the transmission system 400 for Bluetooth audio for voice coding according to the embodiment of the present application can be integrated into the terminal device as a software module and / or a hardware module. For example, the transmission system 400 for Bluetooth audio for voice coding can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; of course, the transmission system 400 for Bluetooth audio for voice coding can also be one of the many hardware modules of the terminal device.

[0093] Alternatively, in another example, the transmission system 400 for Bluetooth audio with voice encoding and the terminal device may be separate devices, and the transmission system 400 for Bluetooth audio with voice encoding may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0094] Exemplary Electronic Devices

[0095] Below, reference Figure 5 To describe the electronic device according to the embodiment of the present application. Figure 5 As shown, the electronic device 10 includes one or more processors 11 and a memory 12. The processor 11 may be a central processing unit (CPU) or other processing units having data processing capability and / or instruction execution capability, and may control other components in the electronic device 10 to perform desired functions.

[0096] The memory 12 may include one or more computer program products, and the computer program product may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 11 may run the program instructions to implement the functions of the transmission method for voice-encoded Bluetooth audio of the various embodiments of the present application described above and / or other desired functions. Various contents such as voice feature vectors, classification feature vectors, etc. may also be stored in the computer-readable storage medium.

[0097] In one example, the electronic device 10 may further include: an input system 13 and an output system 14, and these components are interconnected via a bus system and / or other forms of connection mechanisms (not shown).

[0098] The input system 13 may include, for example, a keyboard, a mouse, and the like.

[0099] The output system 14 can output various information to the outside, including speech feature vectors, categories corresponding to speech signals, etc. The output system 14 can include, for example, a display, a speaker, a printer, a communication network and remote output devices connected thereto, and the like.

[0100] Of course, to simplify, Figure 5Only some of the components related to the present application in the electronic device 10 are shown, and components such as a bus, an input / output interface, etc. are omitted. In addition, according to specific application situations, the electronic device 10 may also include any other appropriate components.

[0101] Exemplary computer program products and computer-readable storage media

[0102] In addition to the above-mentioned methods and devices, an embodiment of the present application may also be a computer program product, which includes computer program instructions, which, when executed by a processor, enable the processor to perform the steps in the functions of the method for transmitting Bluetooth audio for voice encoding according to various embodiments of the present application described in the above "Exemplary Method" section of this specification.

[0103] The computer program product may be written in any combination of one or more programming languages ​​to write program codes for performing the operations of the embodiments of the present application, including object-oriented programming languages, such as Java, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as an independent software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0104] In addition, an embodiment of the present application may also be a computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, cause the processor to execute the steps of the method for transmitting voice-encoded Bluetooth audio described in the above “Exemplary Method” section of this specification.

[0105] The computer readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can include, for example, but is not limited to, a system, system or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0106] The basic principles of the present application are described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, effects, etc. mentioned in the present application are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. are required by each embodiment of the present application. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, not for limitation, and the above details do not limit the present application to being implemented by adopting the above specific details.

[0107] The block diagrams of the devices, systems, equipment, and systems involved in this application are only illustrative examples and are not intended to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagram. As will be appreciated by those skilled in the art, these devices, systems, equipment, and systems can be connected, arranged, and configured in any manner. Words such as "including", "comprising", "having", etc. are open words, referring to "including but not limited to", and can be used interchangeably with them. The words "or" and "and" used here refer to the words "and / or" and can be used interchangeably with them, unless the context clearly indicates otherwise. The words "such as" used here refer to the phrase "such as but not limited to", and can be used interchangeably with them.

[0108] It should also be noted that in the system, device and method of the present application, each component or each step can be decomposed and / or recombined. Such decomposition and / or recombination should be regarded as equivalent solutions of the present application.

[0109] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present application. Therefore, the present application is not intended to be limited to the aspects shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

[0110] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. A method for transmitting Bluetooth audio for speech coding, characterized in that: include: The training phase includes: Acquire a training data set, wherein the training data set includes a speech signal sent by a sending end and a speech signal received by a real receiving end; Passing the speech signal sent by the sending end in the training data set through a semantic understanding model as a predictive coding model to obtain a plurality of speech feature vectors, wherein each speech feature vector in the plurality of speech feature vectors corresponds to a segment of the speech signal sent by the sending end; Inputting the plurality of speech feature vectors into a decoder model composed of a plurality of fully connected layers to recover a decoded speech signal from a speech feature vector sequence composed of the plurality of speech feature vectors; Training the predictive coding model and the decoder model by minimizing the difference between the decoded speech signal and the actual speech signal received by the receiving end; Passing the plurality of speech feature vectors through the converter of the predictive coding model respectively to obtain a plurality of encoded speech feature vectors; Based on the internal structure of the converter, a mask conversion value of each encoded speech feature vector in the plurality of encoded speech feature vectors is calculated to obtain a classification feature vector composed of a plurality of mask conversion values, wherein the mask conversion value is calculated based on a distance between two encoded speech feature vectors in the plurality of encoded speech feature vectors and a representation value of whether each encoded speech feature vector in the plurality of encoded speech feature vectors is encoded; Using a multi-label classifier to label the classification feature vector, the classification feature vector category is used to represent the category of the plurality of encoded speech feature vectors as a whole; and Training the converter of the predictive coding model and the multi-label classifier based on the cross entropy loss function value between the category of the classification feature vector and the true category; and The inference phase includes: Obtain the voice signal sent by the sender; Inputting the speech signal sent by the transmitting end into the predictive coding model trained in the training phase and the multi-label classifier to obtain a plurality of speech feature vectors and categories corresponding to the speech signal sent by the transmitting end; and The multiple speech feature vectors and the categories corresponding to the speech signal sent by the sending end are sent to the receiving end.

2. The method for transmitting Bluetooth audio for speech coding according to claim 1, wherein: The speech signal sent by the transmitting end in the training data set is passed through a semantic understanding model as a predictive coding model to obtain a plurality of speech feature vectors, including: The voice signal sent by the transmitting end is converted into a digital signal through an analog-to-digital converter; Using a spectrum analyzer to perform spectrum analysis on the digitized voice signal sent by the sending end, so as to divide the digitized voice signal sent by the sending end into a plurality of sub-voice signals; and Using a parameter transformer to process the plurality of sub-voice signals respectively to obtain a plurality of sub-voice vectors; and The plurality of sub-speech vectors are passed through the semantic understanding model to obtain the plurality of speech feature vectors.

3. The method for transmitting Bluetooth audio for speech coding according to claim 2, wherein: Using a spectrum analyzer to perform spectrum analysis on the digitized voice signal sent by the sending end to divide the digitized voice signal sent by the sending end into a plurality of sub-voice signals, including: Performing discrete Fourier transform on the digitized speech signal sent by the transmitting end; and Performing Mel filtering on the digitized speech signal sent by the transmitting end after discrete Fourier transformation; and The speech signal sent by the transmitting end after Mel filtering is segmented to obtain the multiple sub-speech signals.

4. The method for transmitting Bluetooth audio for speech coding according to claim 1, wherein: The predictive coding model and the decoder model are trained by minimizing the difference between the decoded speech signal and the actual speech signal received by the receiving end, comprising: A cross entropy loss function value between the decoded speech signal and the actual speech signal received by the receiving end is calculated as the difference between the two.

5. The method for transmitting Bluetooth audio for speech coding according to claim 1, wherein: Based on the intrinsic structure of the converter, a mask conversion value of each encoded speech feature vector in the plurality of encoded speech feature vectors is calculated to obtain a classification feature vector composed of the plurality of mask conversion values, comprising: Based on the intrinsic structure of the converter, the mask conversion value of each encoded speech feature vector in the plurality of encoded speech feature vectors is calculated by the following formula to obtain a classification feature vector composed of the plurality of mask conversion values; The formula is: where v j is the number of coded speech feature vectors except v i The coded speech feature vector other than j=1,…,n,j≠i (·) represents the speech feature vector v for all the codes j Sum, Indicates v i With v j The distance between them, and R represents v i Whether to be masked during predictive coding.

6. The method for transmitting Bluetooth audio for speech coding according to claim 1, wherein: The classification feature vector is labeled with a category of the classification feature vector by a multi-label classifier, including: Inputting the classification feature vector into the Softmax classification function of the multi-label classifier to obtain probability values ​​of the classification feature vector belonging to each category respectively; and The category corresponding to the largest probability value among the probability values ​​is determined as the category to which the classification feature vector belongs.

7. A Bluetooth audio transmission system for speech coding, characterized in that: include: Training modules include: A training data set acquisition unit, used to acquire a training data set, wherein the training data set includes a speech signal sent by a sending end and a speech signal received by a real receiving end; an encoding unit, configured to obtain a plurality of speech feature vectors by passing the speech signal sent by the transmitting end in the training data set obtained by the training data set obtaining unit through a semantic understanding model as a predictive coding model, wherein each speech feature vector in the plurality of speech feature vectors corresponds to a segment of the speech signal sent by the transmitting end; A decoding unit, configured to input the plurality of speech feature vectors obtained by the encoding unit into a decoder model composed of a plurality of fully connected layers to recover a decoded speech signal from a speech feature vector sequence composed of the plurality of speech feature vectors obtained by the encoding unit; a difference training unit, configured to train the predictive coding model and the decoder model by minimizing the difference between the decoded speech signal obtained by the decoding unit and the actual speech signal received by the receiving end obtained by the training data set acquisition unit; A converter processing unit, configured to respectively pass the plurality of speech feature vectors obtained by the encoding unit through a converter of the predictive coding model obtained by the difference training unit to obtain a plurality of encoded speech feature vectors; a mask unit, configured to calculate, based on an internal structure of the converter, a mask conversion value of each of the plurality of encoded speech feature vectors obtained by the converter processing unit to obtain a classification feature vector composed of a plurality of mask conversion values, wherein the mask conversion value is calculated based on a distance between two encoded speech feature vectors in the plurality of encoded speech feature vectors and a representation value of whether each of the plurality of encoded speech feature vectors is encoded; a classification unit, configured to label the classification feature vector obtained by the mask unit with a category of the classification feature vector through a multi-label classifier, wherein the category of the classification feature vector is used to represent the category of the plurality of encoded speech feature vectors as a whole; and A cross entropy loss function value training unit, used to train the converter of the predictive coding model and the multi-label classifier based on the cross entropy loss function value between the category of the classification feature vector obtained by the classification unit and the true category; and Inference modules, including: A voice signal acquisition unit, used to acquire a voice signal sent by a sending end; and A feature vector and category generating unit, used for inputting the speech signal sent by the sending end obtained by the speech signal acquiring unit into the predictive coding model trained in the training phase and the multi-label classifier to obtain a plurality of speech feature vectors and categories corresponding to the speech signal sent by the sending end; The sending unit is used to send the multiple speech feature vectors obtained by the feature vector and category generating unit and the category corresponding to the speech signal sent by the sending end obtained by the feature vector and category generating unit to the receiving end.

8. The transmission system for speech-encoded Bluetooth audio according to claim 7, wherein: The encoding unit comprises: A conversion subunit, configured to convert the voice signal sent by the sending end into a digital signal through an analog-to-digital converter; a spectrum analysis subunit, configured to use a spectrum analyzer to perform spectrum analysis on the digitized voice signal sent by the transmitting end obtained by the conversion subunit, so as to divide the digitized voice signal sent by the transmitting end into a plurality of sub-voice signals; and a parameter converter processing subunit, configured to use a parameter converter to process the plurality of sub-speech signals obtained by the spectrum analysis subunit to obtain a plurality of sub-speech vectors; and The semantic understanding model processing subunit is used to pass the multiple sub-speech vectors obtained by the parameter converter processing subunit through the semantic understanding model to obtain the multiple speech feature vectors.

9. The transmission system for speech-encoded Bluetooth audio according to claim 7, wherein: The mask unit is further used for: Based on the intrinsic structure of the converter, the mask conversion value of each encoded speech feature vector in the plurality of encoded speech feature vectors is calculated by the following formula to obtain a classification feature vector composed of the plurality of mask conversion values; The formula is: where v j is the number of coded speech feature vectors except v i The coded speech feature vector other than j=1,…,n,j≠i (·) represents the speech feature vector v for all the codes j Sum, Indicates v i With v j The distance between them, and R represents v i Whether to be masked during predictive coding.

10. An electronic device, comprising: processor; as well as A memory, in which computer program instructions are stored, and when the computer program instructions are executed by the processor, the processor executes the method for transmitting Bluetooth audio for speech coding according to any one of claims 1 to 6.