Joint source-channel coding speech wireless transmission method based on regulated autoencoder

Abstract

The present invention belongs to the field of speech processing and transmission technology, and specifically provides a joint source-channel coded speech wireless transmission method based on an adjusted autoencoder, which is used to adapt a single deep joint source-channel model to different channel conditions. The autoencoder of the present invention includes: a transmitter, a channel module, and a receiver. The transmitter includes: a joint encoder and an adjustment network. The channel module includes: a signal average transmit power limiter and a wireless channel. The receiver includes: a joint decoder and a demodulation network. The adjustment network and the demodulation network generate adjustment factors based solely on the signal-to-noise ratio (SNR) value of the wireless channel, and scale the speech feature vectors in the joint encoder and joint decoder to adjust the encoding and decoding processes, thereby achieving the ability to resist channel noise of varying intensities. Based on this, the present invention can provide robust speech wireless transmission performance within a given channel SNR range, surpassing existing deep joint source-channel methods.

Description

Technical Field

[0001] The present invention belongs to the technical field of speech processing and transmission, and specifically provides a joint source-channel coding speech wireless transmission method based on an adjusted autoencoder. Background Art

[0002] Traditional communication systems, originating from Shannon's separation theorem, employ a separate design in which source coding, channel coding, and modulation are independently executed. Although this separation design performs well and dominates in current communication networks, it still suffers from two drawbacks: the cliff effect and the saturation effect. The cliff effect refers to the complete failure of transmission when the channel quality falls below the error correction capability of the channel coding, while the saturation effect refers to the inability to further improve performance when the channel quality exceeds the target channel quality.

[0003] In recent years, the deep learning (DL)-based joint source-channel coding (JSCC) paradigm has emerged. In this paradigm, the sender converts input information into channel symbols, and the receiver recovers the original information from the received information. By combining the JSCC encoder, channel, and JSCC decoder into an autoencoder and training it in an end-to-end lossy manner, the encoder-decoder group can learn to extract important semantic features, overcome channel noise, and achieve the highest possible fidelity at the receiver. Deep JSCC aims to convey the essential semantics of a given transmission task, which is closely related to the recent hot concept of semantic communication. Many recent studies have demonstrated that deep JSCC outperforms separate designs for different data type transmission tasks, such as image, video, and text.

[0004] Although deep JSCC can learn the state that best suits the current channel conditions through end-to-end training, the mismatch between channel conditions in training and deployment will lead to severe performance degradation; deploying multiple networks suitable for multiple signal-to-noise ratios can alleviate this problem, but it will greatly increase the complexity of the transceiver. In the image transmission task, the document "Xu J, Ai B, Chen W, et al. Wireless Image Transmission Using Deep Source Channel Coding With Attention Modules[J]. 2020. DOI: 10.48550 / arXiv.2012.00533." proposed the ADJSCC (Attention Deep Learning Based Joint Source-Channel Coding) network structure to overcome this shortcoming. ADJSCC is a deep learning wireless image transmission JSCC method based on the attention module. This method can successfully operate in a wide signal-to-noise ratio range during the transmission process. A series of studies based on ADJSCC have verified its effectiveness in wireless image transmission; however, due to the different data types of images and speech, the ADJSCC network structure is not suitable for speech transmission tasks; therefore, the present invention provides a joint source channel coding speech wireless transmission method based on an adjusted autoencoder. Summary of the Invention

[0005] The object of the present invention is to provide a method for wireless transmission of speech using joint source-channel coding based on an adjusted autoencoder, so that a single-depth JSCC model can adapt to different channel conditions. Specifically, the present invention proposes a regulation network and a deregulation network with a signal-to-noise ratio adjustment module, which uses the signal-to-noise ratio information of the channel to generate an adjustment factor, and adjusts the features in the encoder and decoder to control the importance of different channels, thereby obtaining the ability to resist channel noise of different intensities. Based on this, the present invention can provide robust transmission performance within a larger channel signal-to-noise ratio range, which is superior to existing deep JSCC methods.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is:

[0007] A joint source-channel coded voice wireless transmission method based on an adjustable autoencoder, characterized in that the adjustable autoencoder includes: a transmitting end, a channel module, and a receiving end; the transmitting end includes: a joint encoder and an adjustment network; the channel module includes: a signal average transmission power limiter and a wireless channel; the receiving end includes: a joint decoder and a demodulation network; the joint encoder is composed of four modules A and one module B in cascade; the joint decoder is composed of one module B, four modules A, and another module B in cascade; the adjustment network and the demodulation network adopt the same structure, each including four signal-to-noise ratio adjustment modules, and the signal-to-noise ratio adjustment modules are connected to the modules A in a one-to-one correspondence; the joint source-channel coded voice wireless transmission method includes the following steps:

[0008] Step S1: Input the signal-to-noise ratio information of the wireless channel into the adjustment network and the de-adjustment network respectively, and the adjustment network and the de-adjustment network output an adjustment factor Q and an adjustment factor P, wherein Q = {q1, ..., q4}, P = {p1, ..., p4}, q1, ..., q4 respectively represent the adjustment factors of the four modules A in the joint encoder, and p1, ..., p4 respectively represent the adjustment factors of the four modules A in the joint decoder;

[0009] Step S2: At the transmitting end, the original speech signal X is input into a normalization layer to obtain a speech signal vector X1; the speech signal X1 is resized into a two-dimensional matrix to obtain a speech signal matrix X2; the speech signal matrix X2 is encoded by a joint encoder. During the encoding process, the four modules A are adjusted by adjustment factors q1, q2, q3, and q4 respectively to obtain a speech feature vector X3 for channel transmission;

[0010] Step S3: In the channel module, the signal average transmission power limiter limits the average power of the speech feature vector X3 to obtain a transmission signal R, and transmits it to the receiving end through the wireless channel;

[0011] Step S4: At the receiving end, the transmission signal from the channel module is received to obtain a speech feature vector Y1; the speech feature vector Y1 is decoded by a joint decoder, and during the decoding process, the four modules A are adjusted respectively by adjustment factors p1, p2, p3, and p4 to obtain a reconstructed speech signal matrix Y2; the reconstructed speech signal matrix Y2 is resized into a one-dimensional vector to obtain a reconstructed speech signal vector Y3; the reconstructed speech signal Y3 is input into a denormalization layer to obtain a reconstructed speech signal Y.

[0012] Furthermore, the internal data processing flow of the module A is as follows: the input vector of module A is represented as J, and the input vector J is respectively input into four parallel two-dimensional convolutional layers to obtain feature vectors K1, K2, K3 and K4, and the feature vectors K1, K2, K3 and K4 are feature-concatenated to obtain the feature vector L, and the feature vector L is passed through the cascaded two-dimensional convolutional layer and the normalization layer to obtain the feature vector Z, and the feature vector Z is passed through the cascaded global pooling layer, the fully connected layer, the ReLU activation function layer, the fully connected layer, and the Sigmoid activation function layer to obtain the feature vector M, and the feature vector M is dimensionally expanded according to the size of the feature vector Z and multiplied with the feature vector Z, and the multiplication result is added to the input vector J, and the addition result is finally dot-multiplied with the adjustment factor output by the signal-to-noise ratio adjustment module to obtain the output of module A.

[0013] Furthermore, the internal data processing flow of the module B is as follows: the input vector of the module B is represented as N, and the input vector N is passed through a cascaded two-dimensional convolutional layer and a normalization layer to obtain the output of the module B.

[0014] Furthermore, the internal data processing flow of the signal-to-noise ratio adjustment module is as follows: the input of the signal-to-noise ratio adjustment module is the signal-to-noise ratio information of the wireless channel, and the signal-to-noise ratio information passes through the cascaded fully connected layer, ReLU activation function layer, fully connected layer and Sigmoid activation function layer in sequence to obtain the corresponding adjustment factor as the output of the signal-to-noise ratio adjustment module.

[0015] Furthermore, the regulated autoencoder is trained end-to-end, and the loss function of the training process is:

[0016]

[0017] Among them, L is the loss function, N is the number of speech signals in the training set, n is the number of sampling points of each speech signal in the training set, and s i,j represents the jth sampling point of the i-th original speech signal, s′ i,j represents the jth sampling point of the i-th reconstructed speech signal.

[0018] Based on the above technical solution, the beneficial effects of the present invention are:

[0019] The present invention provides a joint source-channel coded voice wireless transmission method based on a regulated autoencoder, wherein the autoencoder includes: a transmitter, a channel module, and a receiver, wherein the transmitter includes: an encoder and a regulation network; the channel module includes: a signal average transmission power limiter and a wireless channel; the receiver includes: a joint decoder and a demodulation network, and the regulation network and the demodulation network both include fully connected neural networks. The present invention adopts joint source-channel coding, which is more effective than traditional independent coding methods and overcomes cliff effects and saturation effects. At the transmitter, feature information is extracted from the original voice signal through the joint encoder, and the feature information enters the channel module. After the transmit signal power is constrained by the signal average transmission power limiter, it is transmitted to the receiver through the wireless channel, and the receiver reconstructs the original voice signal from the received signal through the joint decoder. In addition, during the encoding and decoding process, signal-to-noise ratio information of the current channel is introduced. Under the regulation of the regulation network and the demodulation network, more balanced resource allocation can be achieved between source coding and channel coding in voice transmission, making the voice transmission task more adaptable to current wireless channel conditions.

[0020] Furthermore, the autoencoder provided by the present invention incorporates a regulation network and a deregulation network, takes the signal-to-noise ratio information of the wireless channel as input, and adaptively generates a regulation factor, making the model more robust when facing wireless channels in different states; at the same time, the present invention uses a mean square error penalty function to train the reconstruction and regulation network of the speech signal end-to-end, effectively reducing the switching overhead of the signal-to-noise ratio scaling factor under different channel signal-to-noise ratio conditions, making it easier to deploy in practice.

[0021] In summary, the present invention, based on a deep learning method, proposes a voice compression transmission method based on joint source-channel coding of an autoencoder, providing efficient voice transmission services; introduces signal-to-noise ratio information of the wireless channel during the encoding and decoding processes, making the transmission task more adaptable to the current channel state; uses a regulation network and a deregulation network to adaptively generate a signal-to-noise ratio adjustment factor under the current channel state, further reducing the switching overhead of the adjustment factor when facing different channels, thereby achieving the effect of adaptive signal-to-noise ratio; in particular, the adaptive signal-to-noise ratio voice wireless transmission method of the present invention only requires one training for a signal-to-noise ratio within a predetermined range, has low deployment costs, and has strong adaptability to different channel transmission conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a structural diagram of a voice wireless transmission system using joint source-channel coding based on an adjusted autoencoder in Example 1 of the present invention.

[0023] Figure 2 This is a structural diagram of module A in Example 1 of the present invention.

[0024] Figure 3This is a structural diagram of the signal-to-noise ratio adjustment module according to Example 1 of the present invention.

[0025] Figure 4 Schematic diagram of the structure of the voice wireless transmission system based on the autoencoder in Comparative Example 1 of the present invention.

[0026] Figure 5 This is a comparison chart of the mean square error (MSE) between Example 1 of the present invention and Comparative Example 1.

[0027] Figure 6 1 is a comparison chart of the signal distortion ratio (SDR) of Example 1 of the present invention and Comparative Example 1.

[0028] Figure 7 1 is a comparison chart of the short-time objective intelligibility (STOI) between Example 1 of the present invention and Comparative Example 1.

[0029] Figure 8 This is a comparison chart of the mean square error (MSE) between Example 2 of the present invention and Comparative Example 2.

[0030] Figure 9 2 is a comparison chart of the signal distortion ratio (SDR) of Example 2 of the present invention and Comparative Example 2.

[0031] Figure 10 2 is a comparison chart of the short-time objective intelligibility (STOI) of Example 2 of the present invention and Comparative Example 2. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and beneficial effects of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0033] Example 1

[0034] This embodiment provides a joint source-channel coding voice wireless transmission method based on an adjustment autoencoder. The structure of the adjustment autoencoder is as follows: Figure 1 As shown, it includes: a transmitting end, a channel module and a receiving end; the transmitting end includes: a joint encoder and an adjustment network; the channel module includes: a signal average transmission power limiter and a wireless channel; the receiving end includes: a joint decoder and a demodulation network; the joint encoder is composed of four modules A and one module B in cascade; the joint decoder is composed of one module B, four modules A and another module B in cascade; the adjustment network and the demodulation network adopt the same structure, each including four signal-to-noise ratio adjustment modules, and the signal-to-noise ratio adjustment modules are connected to the modules A in a one-to-one correspondence; more specifically:

[0035] The structure of module A is as follows Figure 2As shown in the figure, its internal data processing flow is as follows: the input vector of module A is represented as J, and the input vector J is input into four parallel two-dimensional convolutional layers to obtain feature vectors K1, K2, K3 and K4 respectively. The feature vectors K1, K2, K3 and K4 are feature concatenated to obtain feature vector L. The feature vector L is passed through the cascaded two-dimensional convolutional layer and the normalization layer to obtain the feature vector Z. The feature vector Z is passed through the cascaded global pooling layer, the fully connected layer, the ReLU activation function layer, the fully connected layer, and the Sigmoid activation function layer to obtain the feature vector M. The feature vector M is dimensionally expanded according to the size of the feature vector Z and then multiplied with the feature vector Z. The multiplication result is then added to the input vector J. The addition result is finally multiplied with the adjustment factor output by the signal-to-noise ratio adjustment module to obtain the output of module A.

[0036] The module B includes a cascade of two-dimensional convolutional layers and batch normalization layers. The internal data processing flow is as follows: the input vector of the module B is represented as N, and the input vector N passes through the cascade of two-dimensional convolutional layers and normalization layers to obtain the output of the module B;

[0037] The structure of the signal-to-noise ratio adjustment module is as follows Figure 3 As shown in the figure, it includes a cascaded fully connected layer, a ReLU activation function layer, a fully connected layer and a Sigmoid activation function layer; its internal data processing flow is as follows: the input of the signal-to-noise ratio adjustment module is the signal-to-noise ratio information of the wireless channel, and the signal-to-noise ratio information passes through the cascaded fully connected layer, the ReLU activation function layer, the fully connected layer and the Sigmoid activation function layer in sequence to obtain the corresponding adjustment factor as the output of the signal-to-noise ratio adjustment module.

[0038] The regulated autoencoder is trained end-to-end. The training process uses the mean square error (MSE) between the original speech signal and the reconstructed speech signal as the penalty function L to update the parameters. The MSE calculation formula is:

[0039]

[0040] Among them, L is the loss function, N is the number of speech signals in the training set, n is the number of sampling points of each speech signal in the training set, and s i Represents the i-th original speech signal, s′ i represents the i-th reconstructed speech signal, s i,j represents the jth sampling point of the i-th original speech signal, s′ i,j represents the jth sampling point of the i-th reconstructed speech signal.

[0041] On the basis of the above-mentioned adjustment of the autoencoder, the joint source-channel coding voice wireless transmission method based on the adjustment of the autoencoder includes the following steps:

[0042] Step S1: Input the signal-to-noise ratio information of the wireless channel into the adjustment network of the transmitting end and the demodulation network of the receiving end respectively, and output the adjustment factor Q and the adjustment factor P containing the signal-to-noise ratio information through each signal-to-noise ratio adjustment module in the adjustment network and the demodulation network, where Q = {q1, ..., q4}, P = {p1, ..., p4}, q1, ..., q4 respectively represent the adjustment factors of the four modules A in the joint encoder of the transmitting end, and p1, ..., p4 respectively represent the adjustment factors of the four modules A in the joint decoder of the receiving end;

[0043] Step S2: At the transmitting end, the original speech signal X is input into a normalization layer to limit the speech signal to the range of [0, 1], thereby obtaining a speech signal vector X1; the speech signal X1 is resized to convert the one-dimensional vector into a two-dimensional matrix, thereby obtaining a speech signal matrix X2; the speech signal matrix X2 is encoded by a joint encoder, and during the encoding process, the four modules A are adjusted by adjustment factors q1, q2, q3, and q4 respectively, thereby obtaining a speech feature vector X3 for channel transmission;

[0044] Step S3: In the channel module, the signal average transmission power limiter limits the average power of the speech feature vector X3 to obtain a transmission signal R, and transmits it to the receiving end through a wireless channel (additive white Gaussian noise wireless channel);

[0045] Step S4: At the receiving end, the transmission signal from the channel module is received to obtain the speech feature vector Y1; the speech feature vector Y1 is decoded by the joint decoder, and the four modules A are adjusted respectively by the adjustment factors p1, p2, p3, and p4 during the decoding process to obtain the reconstructed speech signal matrix Y2; the reconstructed speech signal matrix Y2 is resized to convert the two-dimensional matrix into a one-dimensional vector to obtain the reconstructed speech signal vector Y3; the reconstructed speech signal Y3 is input into the denormalization layer to obtain the reconstructed speech signal Y.

[0046] Example 2

[0047] This embodiment provides a joint source-channel coded voice wireless transmission method based on an adjusted autoencoder, which differs from Embodiment 1 in that the wireless channel of the channel module adopts a Rayleigh fading channel.

[0048] Comparative Example 1

[0049] This comparative example provides a joint source-channel coding voice wireless transmission method based on an autoencoder, such as Figure 4 As shown, the difference between this embodiment and embodiment 1 is that the transmitting end and the receiving end do not include a regulation network and a deregulation network.

[0050] Comparative Example 2

[0051] This comparative example provides a joint source-channel coded voice wireless transmission method based on an autoencoder, which differs from comparative example 1 in that the wireless channel of the channel module adopts a Rayleigh fading channel.

[0052] For this wireless speech transmission method, the present invention uses a public speech dataset from the University of Edinburgh database, which stores research data generated by researchers at the university related to existing or upcoming publications. This dataset provides over 10,000 clear and noisy speech training files. The present invention preprocesses this dataset and downsamples it to 8kHz. The batch size for all embodiments and comparative examples in the present invention is 32, and the target size for speech data resizing at the transmitter is 128*128. The mean squared error between the reconstructed and original speech data is used as the penalty function for updating model parameters, and an RMSprop optimizer with a learning rate of 0.001 is used.

[0053] The present invention uses two common speech quality assessment indicators, Signal Distortion Ratio (SDR) and Short-Time Objective Intelligibility (STOI), to measure the quality of the reconstructed speech signal at the receiving end. SDR can be calculated by the following formula:

[0054]

[0055] Where s represents the original speech signal, s′ represents the reconstructed speech signal, and the higher the SDR value, the smaller the difference between the reconstructed speech signal and the original speech signal, that is, the better the performance of the voice wireless transmission system.

[0056] STOI is a speech quality assessment method designed to measure the correlation between clarity and intelligibility. It evaluates speech signal quality by comparing the spectral correlation between the original and reconstructed speech. The calculation steps are as follows: First, the original and reconstructed speech signals are divided into short time segments and subjected to a short-time Fourier transform (STFT) to convert the signal segments into spectral segments. The spectral similarity between the original and reconstructed speech signals is then calculated. Finally, the similarity across all spectral segments is averaged to obtain the STOI value between the two speech segments. The STOI value ranges from 0 to 1, with values ​​closer to 1 indicating better signal quality.

[0057] Further, such as Figure 5 、 Figure 6 and Figure 7The figures show a comparison of the MSE, SDR, and STOI of Example 1 and Comparative Example 1 using the above-mentioned public dataset in an additive white Gaussian noise wireless channel. The signal-to-noise ratio of different additive white Gaussian noise wireless channel conditions is used as the horizontal axis, and the evaluation index between the reconstructed speech signal and the original speech signal is used as the vertical axis. As can be seen from the figure, the MSE of Example 1 is smaller throughout the entire test signal-to-noise ratio range, and the SDR and STOI of Example 1 are larger throughout the entire test signal-to-noise ratio range. Comparative Example 1 performs poorly when the training signal-to-noise ratio does not match the test signal-to-noise ratio. Even at training signal-to-noise ratios within the dynamic range, the performance is still worse than that of Example 1. This shows that, under the action of the adjustment module, the voice wireless transmission system is more robust when facing additive white Gaussian noise wireless channels in different states, demonstrating the effectiveness of the method proposed in the present invention.

[0058] like Figure 8 、 Figure 9 and Figure 10 The figures show the MSE, SDR, and STOI comparisons of Example 2 and Comparative Example 2 using the above-mentioned public data set under a Rayleigh fading channel, respectively. Similarly, the signal-to-noise ratio under different Rayleigh fading channel conditions is used as the horizontal axis, and the evaluation index between the reconstructed speech signal and the original speech signal is used as the vertical axis. As can be seen from the figure, the conclusion is the same as that under the additive white Gaussian noise wireless channel. Under the Rayleigh fading channel, the adjustment module proposed in the present invention can also enhance the robustness of the system, further verifying the effectiveness of the proposed method.

[0059] In summary, the present invention provides a method for wireless voice transmission based on joint source-channel coding with an adjusted autoencoder, providing efficient voice transmission services. The method introduces wireless channel signal-to-noise ratio information during the encoding and decoding processes, and by adjusting the voice feature vectors in the joint encoder and decoder, the voice transmission task is made more adaptable to the current channel conditions. The method uses a regulation network and a deregulation network to adaptively generate a signal-to-noise ratio adjustment factor for the current channel state, further reducing the switching overhead of the signal-to-noise ratio adjustment factor when facing different channels, thereby achieving an adaptive signal-to-noise ratio effect. In particular, the adaptive signal-to-noise ratio voice wireless transmission method of the present invention requires only one training session for signal-to-noise ratios within a predetermined range, resulting in low deployment costs and strong adaptability to different channel transmission conditions.

[0060] The above description is only a specific embodiment of the present invention. Any feature disclosed in this specification, unless otherwise stated, can be replaced by other equivalent or alternative features with similar purposes; all disclosed features, or all steps in the methods or processes, except for mutually exclusive features and / or steps, can be combined in any way.

Claims

1. A method for wireless speech transmission based on joint source-channel coding and regulated autoencoder, characterized in that: The regulating autoencoder includes: a transmitting end, a channel module and a receiving end; the transmitting end includes: a joint encoder and a regulating network; the channel module includes: a signal average transmission power limiter and a wireless channel; the receiving end includes: a joint decoder and a demodulation network; the joint encoder is composed of four modules A and one module B in cascade; the joint decoder is composed of one module B, four modules A and another module B in cascade; the regulating network and the demodulation network adopt the same structure, each including four signal-to-noise ratio adjustment modules, and the signal-to-noise ratio adjustment modules are connected to the modules A in a one-to-one correspondence; the joint source-channel coded voice wireless transmission method includes the following steps: Step S1: Input the signal-to-noise ratio information of the wireless channel into the adjustment network and the de-adjustment network respectively, and the adjustment network and the de-adjustment network output an adjustment factor Q and an adjustment factor P, wherein Q = {q1, ..., q4}, P = {p1, ..., p4}, q1, ..., q4 respectively represent the adjustment factors of the four modules A in the joint encoder, and p1, ..., p4 respectively represent the adjustment factors of the four modules A in the joint decoder; Step S2: At the transmitting end, the original speech signal X is input into a normalization layer to obtain a speech signal vector X1; the speech signal X1 is resized into a two-dimensional matrix to obtain a speech signal matrix X2; the speech signal matrix X2 is encoded by a joint encoder. During the encoding process, the four modules A are adjusted by adjustment factors q1, q2, q3, and q4 respectively to obtain a speech feature vector X3 for channel transmission; Step S3: In the channel module, the signal average transmission power limiter limits the average power of the speech feature vector X3 to obtain a transmission signal R, and transmits it to the receiving end through the wireless channel; Step S4: At the receiving end, a transmission signal from the channel module is received to obtain a speech feature vector Y1; the speech feature vector Y1 is decoded by a joint decoder, and during the decoding process, the four modules A are adjusted by adjustment factors p1, p2, p3, and p4 respectively to obtain a reconstructed speech signal matrix Y2; the reconstructed speech signal matrix Y2 is resized into a one-dimensional vector to obtain a reconstructed speech signal vector Y3; the reconstructed speech signal Y3 is input into a denormalization layer to obtain a reconstructed speech signal Y; The internal data processing flow of the module A is as follows: the input vector of the module A is represented as J, the input vector J is respectively input into four parallel two-dimensional convolutional layers to obtain feature vectors K1, K2, K3 and K4, the feature vectors K1, K2, K3 and K4 are feature concatenated to obtain a feature vector L, the feature vector L is passed through a cascade of two-dimensional convolutional layers and a normalization layer to obtain a feature vector Z, the feature vector Z is passed through a cascade of a global pooling layer, a fully connected layer, a ReLU activation function layer, a fully connected layer, and a Sigmoid activation function layer to obtain a feature vector M, the feature vector M is dimensionally expanded according to the size of the feature vector Z and multiplied with the feature vector Z, the multiplication result is then added to the input vector J, and the addition result is finally dot-multiplied with the adjustment factor output by the signal-to-noise ratio adjustment module to obtain the output of the module A; The internal data processing flow of the module B is as follows: the input vector of the module B is represented as N, and the input vector N is passed through a cascade of two-dimensional convolutional layers and normalization layers to obtain the output of the module B; The internal data processing flow of the signal-to-noise ratio adjustment module is as follows: the input of the signal-to-noise ratio adjustment module is the signal-to-noise ratio information of the wireless channel. The signal-to-noise ratio information passes through the cascaded fully connected layer, ReLU activation function layer, fully connected layer and Sigmoid activation function layer in sequence to obtain the corresponding adjustment factor as the output of the signal-to-noise ratio adjustment module.

2. The method for voice wireless transmission based on joint source-channel coding and adjustment of autoencoders according to claim 1, characterized in that: The regulated autoencoder is trained end-to-end, and the loss function of the training process is: Among them, L is the loss function, N is the number of speech signals in the training set, n is the number of sampling points of each speech signal in the training set, and s i,j represents the jth sampling point of the i-th original speech signal, s′ i,j represents the jth sampling point of the i-th reconstructed speech signal.

Images