Sound source separation method based on shallow feature reactivation and multi-stage hybrid attention

Abstract

The present invention discloses a sound source separation method based on shallow feature reactivation and multi-stage mixed attention, comprising the following steps: collecting a single-channel mixed signal and a clean target sound source signal, wherein the single-channel mixed signal corresponds to the clean target sound source signal, processing the single-channel mixed signal, and obtaining a single-channel mixed signal time domain embedding vector; based on the clean target sound source signal, calculating a weight loss function; constructing a separation network model, inputting the single-channel mixed signal time domain embedding vector into the initial separation network model, and training the separation network model using the weight loss function to obtain a trained separation network model; and using the trained separation network model to perform sound source separation on the single-channel mixed signal. The multi-stage network mode and multi-supervisory training method adopted by the present invention improve the fitting ability of the network, with a small number of parameters and high computational efficiency, and further improve the performance of sound source separation.

Description

Technical Field

[0001] The present invention relates to the field of single-channel sound source separation, and in particular to a sound source separation method based on shallow feature reactivation and multi-stage mixed attention. Background Art

[0002] Source separation aims to extract the target sound source from a mixed signal. In the cocktail party problem, it is expressed as separating the clean target speaker's speech from the mixed speech of multiple speakers. This specific source separation has many subsequent applications, such as speaker verification, voiceprint recognition, speaker speech recognition, etc. Under conditions with strong background noise, the denoising task can be regarded as another case of source separation. Noise has a great impact on the performance of tasks such as automatic speech recognition. In music signal processing, separating the clean singing source signal and the various instrument playing signals from the song can be applied to instrument type detection, singing fundamental frequency estimation, automatic lyrics recognition, automatic singer recognition, etc.

[0003] In recent years, the research on sound source separation has made great progress in academia and industry. Both the early traditional machine learning methods and the now popular deep learning research have achieved certain success. Especially for the supervised sound source separation task, the separation performance of the sound source separation model has reached a certain level, but for the needs of practical applications, the current technology still needs to be continuously innovated to mature. Although people can easily perceive an independent source in a mixed signal, it is difficult for computers to learn to recognize a certain feature, especially when the system only has a single-channel mixed signal.

[0004] Thanks to the rapid development of deep learning, sound source separation technology has made great progress in recent years, but the current separation model still has some limitations. For the commonly used CNN, it has a good learning ability for local features, but lacks the concept of temporal sequence. In comparison, recurrent neural networks can capture longer contextual relationships, but they take a long time to train, which is not conducive to the optimization of model parameters. Summary of the invention

[0005] The purpose of the present invention is to provide a sound source separation method based on shallow feature reactivation and multi-stage mixed attention to solve the problems existing in the above-mentioned prior art. The multi-stage network model and multi-supervised training method adopted are used to improve the fitting ability of the network, with small number of parameters and high computational efficiency, so as to further improve the performance of sound source separation.

[0006] To achieve the above object, the present invention provides the following scheme: a sound source separation method based on shallow feature reactivation and multi-stage mixed attention, comprising the following steps:

[0007] Collecting a single-channel mixed signal and a clean target sound source signal, wherein the single-channel mixed signal corresponds to the clean target sound source signal, and processing the single-channel mixed signal to obtain a single-channel mixed signal time domain embedding vector;

[0008] Calculating a weighted loss function based on the clean target sound source signal;

[0009] Constructing a separation network model, inputting the single-channel mixed signal time-domain embedding vector into the separation network model, and training the separation network model using the weight loss function to obtain a trained separation network model;

[0010] The trained separation network model is used to perform sound source separation on the single-channel mixed signal.

[0011] Optionally, in the process of processing the single-channel mixed signal, the single-channel mixed signal is sent to an encoder, and a one-dimensional convolutional neural network is used to process the single-channel mixed signal to obtain an embedding vector of the single-channel mixed signal.

[0012] Optionally, the process of obtaining the trained separation network model includes:

[0013] The single-channel mixed signal time domain embedding vector is input into the separation network model, and the estimated feature masking matrices of multiple target sound sources are obtained in combination with the shallow feature reactivation mechanism. Based on the estimated feature masking matrices of the multiple target sound sources, the weight loss function is used to update the parameters of the separation network model to obtain the separation network model, wherein the shallow feature reactivation mechanism uses the difference in feature information contained in the time domain feature vectors generated at different stages to prune and fuse shallow features and deep features to extract time domain features.

[0014] Optionally, using the trained separation network model to perform sound source separation on the single-channel mixed signal includes:

[0015] Based on the estimated characteristic masking matrices of the multiple target sound sources and the single-channel mixed signal time domain embedding vector, characteristic vectors of the multiple target sound source signals are obtained, and decoding is performed to obtain waveform signals of the multiple target sound source signals to complete separation.

[0016] Optionally, the separation network model includes a hybrid attention module, which is used to obtain a better fine-grained feature representation for the single-channel mixed signal time domain embedding vector.

[0017] Optionally, the hybrid attention module includes a multi-head self-attention sub-unit, an element attention sub-unit and an adaptive compression excitation sub-unit.

[0018] Optionally, obtaining the better fine-grained feature representation includes: obtaining the correlation on the mixed signal time series through the multi-head self-attention sub-unit, combining the element attention sub-unit and the adaptive compression excitation sub-unit to obtain the feature correlation between points and channels of the mixed signal, so as to obtain the better fine-grained feature representation.

[0019] Optionally, when updating the parameters of the separation network model based on the weight loss function, it also includes optimizing the parameters of the separation network model based on the fine-grained time-domain detail features and the more optimal fine-grained feature representation.

[0020] Optionally, the estimated feature masking matrix of the target sound source is obtained using a stage mixed attention mechanism network, wherein the multi-stage mixed attention mechanism network includes multiple stage mixed attention sub-modules, and the acquisition process includes: selecting the output of the last stage mixed attention sub-module as the feature masking estimation matrix of the multiple target sound sources.

[0021] Optionally, extracting feature vectors of a single-stage mixed attention module in the multiple-stage mixed attention sub-modules includes: obtaining time domain features through different types of attention mechanisms in the single-stage mixed attention module, wherein the time domain features include local time domain features and global time domain features, obtaining local features and global features according to the time domain features, fusing the local features and the global features to obtain a feature vector of the single-stage mixed attention sub-module.

[0022] The present invention discloses the following technical effects:

[0023] The sound source separation method based on shallow feature reactivation and multi-stage mixed attention provided by the present invention is

[0024] (1) A shallow feature reactivation mechanism is used to obtain a better feature representation for the network. After feature pruning and fusion operations at different stages, combined with a multi-stage hybrid attention module to aggregate different types of context information, more refined time domain features of different sound sources can be extracted. The multi-stage network model and multi-supervised training method adopted by the present invention improve the fitting ability of the network and further improve the performance of sound source separation.

[0025] (2) Experimental results show that the separation model used in the present invention can be well applied to multi-speaker separation tasks, as well as music source separation and target speech enhancement tasks under background noise conditions, and the model is more scalable.

[0026] (3) The separation network adopted by the present invention combines CNN and RNN, with small parameters and high computational efficiency, and can obtain context information over a longer range while ensuring low computational complexity of the network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0028] Figure 1 It is a flow chart of a sound source separation method based on shallow feature reactivation and multi-stage mixed attention in an embodiment of the present invention;

[0029] Figure 2 It is a schematic diagram of the overall scheme of an embodiment of the invention;

[0030] Figure 3 A schematic diagram of a hybrid attention network model according to an embodiment of the present invention;

[0031] Figure 4 This is a schematic diagram of an adaptive compression-stimulated attention unit network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] The present invention provides a sound source separation method based on shallow feature reactivation and multi-stage mixed attention, such as Figure 1 As shown, the implementation includes the following steps:

[0035] S1. Collect a single-channel mixed signal and a clean target sound source signal, process the single-channel mixed signal, and obtain a time-domain embedding vector of the single-channel mixed signal; and calculate a weight loss function based on the clean target sound source signal.

[0036] A certain number of single-channel mixed signals and their corresponding clean target sound sources (ie, clean voices of each speaker) are obtained. Specifically, for a single-channel mixed signal, there must be a clean target sound source sample to match.

[0037] The single-channel mixed time domain signal is sent to the encoder for processing to obtain the single-channel mixed signal time domain embedding vector. The encoder processing is: processing the single-channel mixed signal through a one-dimensional convolutional neural network to obtain the single-channel mixed signal time domain embedding vector.

[0038] S2. Build a separation network model, input the single-channel mixed signal time domain embedding vector into the separation network model, use the weight loss function to train the separation network model, and obtain a trained separation network model.

[0039] Building and training a separation network model is the basis for achieving target sound source separation in mixed signals. Figure 2 The shallow feature reactivation mechanism and multi-stage hybrid attention mechanism network model diagram shown in the figure use the separation network model during training and testing. When training the network model, it is necessary to update the network parameters according to the loss function and the optimization function. In addition, when training the separation network, the output of the network includes the outputs of multiple stages, and the update of the network gradient is jointly determined by the outputs of multiple stages. When performing sound source separation, the output of the network only includes the output of the last stage.

[0040] When training the sound source separation network model based on the shallow feature reactivation mechanism and the multi-stage hybrid attention mechanism (that is, training the neural network parameters), the following steps are specifically followed:

[0041] Step train1: Use the single-channel mixed signal time domain embedding vector as the input of the separation network model, use the waveform signal of the clean target sound source to calculate the loss function, and train the neural network model. The training speed depends on the configuration of the machine hardware and the size of the sample.

[0042] Specifically, a neural network is built using a shallow feature reactivation mechanism and a multi-stage hybrid attention mechanism network model, and the network is set to a trainable state (i.e., the network parameters will be continuously updated with training iterations). Based on the multi-stage hybrid attention module and the shallow feature reactivation mechanism, the estimated feature masking matrices of multiple target sound sources are obtained, and the mixed signal time domain waveform signal x(t) is used as the input of the separation network. At the same time, the time domain waveform signals s1(t), s2(t), ..., s corresponding to the clean target sound sources are l (t) is used as the supervision of the multi-stage network output to train the neural network model, and the network model learns and distinguishes the time domain features. The representation of the mixed source signal x(t) in the time domain is shown in formula (1):

[0043]

[0044] Among them, the value of I represents the number of pre-separated target sound sources, s i(t) represents the time domain waveform signal corresponding to the clean target sound source. x(t) is the target sound source to be separated. In theory, it should satisfy that the waveform signal of the mixed source is equal to the sum of the waveform signals corresponding to the target sound source.

[0045] Specifically, the time domain waveform of the input mixed signal is passed through the encoder to obtain the time domain embedding vector of the mixed signal, and the network output of multiple stages is obtained through the separation model. The multiple sound source estimation feature masking matrices outputted in the last stage are multiplied by the time domain embedding vector of the mixed signal passed through the encoder respectively, and the separated multiple target sound source time domain waveforms are obtained through the decoder. This embodiment adopts a sound source separation network based on a shallow feature reactivation mechanism and a multi-stage hybrid attention mechanism to realize training and separation. The sound source separation network model based on the shallow feature reactivation mechanism and the multi-stage hybrid attention mechanism is as follows Figure 2 shown.

[0046] In this embodiment, the separation network model includes a hybrid attention module. The hybrid attention module is used as the main feature extraction sub-network. The hybrid attention module can obtain the local and global contextual relationship, so that the network can obtain better feature representation. The structure of the hybrid attention module is as follows: Figure 3 As shown. In order to obtain contextual information between long time series, RNN or CNN structure is usually used. One method is to use a recurrent neural network to obtain contextual information, usually using a long short-term memory structure, but this structure requires a time series signal of considerable length as input, and has high computational complexity and takes a long time. Another method uses a convolutional neural network with an expansion factor, but blindly increasing the expansion rate will make the weight matrix too sparse, resulting in loss of feature information. Or increase the size of the convolution kernel, but it will increase the computational load and training time. Another way is to increase the network depth, but this will reduce computational efficiency and usually lead to problems such as gradient disappearance and network degradation. In order to effectively solve this problem, a hybrid attention module is used in this embodiment to obtain global contextual relevance for different dimensional features at the original resolution through a recurrent neural network and a self-attention mechanism, and use a convolutional neural network to obtain local information. By combining the established global relevance, the subsequent feature extraction module can better learn local time domain features.

[0047] The input of the hybrid attention module is the time domain features after block extraction. After multiple stages of feature extraction, the deep features will lose some feature information, so it is necessary to fuse the shallow features with the deep features to supplement the original feature information. The multi-stage shallow feature reactivation mechanism described in the embodiment of the present invention is described as follows:

[0048] In deep neural networks, dense connections are usually used to achieve feature reuse. However, as the connection density in the network is too high, there will be more redundant features. If the shallow features are directly involved in the new feature calculation of this layer, redundant calculations will be introduced. In order to solve this problem, in the present invention, we use the method of grouped convolution to prune the redundant features that may be generated, and reactivate the valuable redundant features to improve feature utilization, enhance the feature extraction capability of the network, and extract richer and more effective time domain features. The calculation process of the shallow feature reactivation mechanism is shown in formula (2):

[0049] y p =M p (concat[y0,y1,...,y p-1 ]) (2)

[0050] Among them, y0,y1,...,y p-1 Represents the features of each layer before layer p, M p (·) represents the feature extraction operation of the pth layer, y p Indicates that after M p (·) after the features obtained, concat[] represents the concatenation operation. All features before layer p, y0, y1, ..., y p-1 , after concatenation, is input into the pth layer for feature extraction to generate the new feature y of this layer p After the feature is reactivated, the equation (3) is obtained:

[0051]

[0052] Among them, y0, y1, ..., y p-1 represents the features of layer p and previous layers, R(·) represents the feature reactivation layer, Represents the feature after feature reactivation. Here, the feature reactivation layer R(·) is a learnable group convolution with a convolution kernel size of 1×1. Let the weight matrix be F∈R O×N×1×1 , where O is the number of output channels and N is the number of input channels. The weights are divided into g groups, F 1 , ..., F g , where the dimension of each weight is F g ∈R O×(N / g) Selectively selecting important shallow features to participate in new feature fusion can effectively reduce the computational overhead of the network, while reactivating shallow features and improving feature utilization in deep networks.

[0053] In this embodiment, the hybrid attention module structure includes three different unit structures: multi-head self-attention subunit, element-wise attention subunit and adaptive squeeze and excitation (Squeeze and Excitaion, SE) subunit. The method for the hybrid attention module to obtain a better fine-grained feature representation is as follows: obtain the correlation on the time series of the mixed signal through the multi-head self-attention subunit, combine the element-wise attention subunit and the adaptive squeeze and excitation subunit to obtain the feature correlation between points and channels of the mixed signal, and obtain the better feature representation. The multi-stage hybrid attention mechanism and the shallow feature reactivation mechanism also obtain fine-grained time domain detail features based on the differences between feature information.

[0054] The multi-head self-attention subunit, element-wiseattention subunit and adaptive squeeze and excitation (SE) subunit are as follows:

[0055] 1. Multi-head self-attention subunit

[0056] The multi-head self-attention unit aims to use the correlation between any two position features to obtain the relationship between frames in a long time series to enhance the expression of each position feature. The self-attention mechanism obtains the weight of each frame in the time frame sequence by calculating the weight value of the input time frame sequence itself, and then obtains the contextual relationship of the current frame in the time frame sequence. Given an input H∈R C×K×S , two sets of features Q and K are generated respectively through two convolution layers with a convolution kernel of 1×1, where Q and K have the same dimensions and Q, K∈R N×K×S , N represents the number of output channels, K represents the length of each frame, and S represents the dimension of the feature vector. For any channel u on Q, we can get the vector Q u ∈R K ×S Similarly, we can obtain the feature vector set K of the same channel u on K u ∈R K×S . Furthermore, the attention weight matrix is ​​obtained by combining the feature vectors of Q and K through the Affinity transformation operation. The affine transformation operation is defined as shown in formula (4):

[0057]

[0058] Here we get Wu Indicates Q u and K u The weight matrix of the degree of correlation, W∈R K×K The obtained correlation is passed through a Softmax layer to obtain the normalized weight matrix A∈R K×K In addition, the input H is also passed through a convolution layer with a convolution kernel of size 1×1 to generate V, V∈R N×K×S Based on this, the Attention operation is defined as shown in formula (5):

[0059]

[0060] Here is the k is the dimension of the feature vector, QK T is the weight matrix of the correlation between Q and K, and softmax() is the Softmax layer. Multi-head self-attention is to divide the input data into h parts and perform self-attention operation on each part of the data. The obtained weight matrix is ​​concatenated and multiplied with the original input to obtain the relationship of each frame in the global context as shown in equations (6)-(7).

[0061]

[0062] Multi Head(Q,K,V)=Concat(head1,...,head h ) (7)

[0063] Here is the head i where i is the number of longs, i = [1,…,h], is the weight corresponding to Q, K, and V after segmentation, and Multi Head (Q, K, V) is the final output feature obtained after the final concatenation. After the multi-head self-attention operation, the local features combined with the global context have a better spatial position expression.

[0064] 2. Element-wise attention subunit

[0065] RNN can model the contextual relationship of long time series and capture global features, but it does not take into account the importance and correlation between points. Therefore, the present invention adopts a recurrent neural network unit using a gating mechanism to supplement the relationship between points. For a given input H∈R N×K×S , respectively, through two bidirectional long short-term memory (Bi-directional Long Short-Term Memory, BLSTM) layers to capture the contextual relationship, and then one of them passes through the Sigmoid layer and multiplies with the other to obtain the attention feature matrix A∈RN×K×S , characterizing the element-level correlation in the feature sequence. The calculation process of the gated recurrent neural network is shown in formula (8):

[0066] A=σ(BLSTM(H))×BLSTM(H)+H (8)

[0067] Here, BLSTM(·) refers to the operation of the bidirectional long short-term memory layer, σ is the Sigmoid layer, and A is the sum of the input and the obtained features through the final residual connection. By combining the features obtained by multi-head self-attention with the features obtained by the gated neural network, the correlation between frames and the correlation between points are captured simultaneously, and the network obtains a better feature representation.

[0068] 3. Adaptive Squeeze and Excitaion (SE) subunit

[0069] The channel attention network models the contextual relationship of features from the perspective of channels, but does not take into account the importance and correlation between feature time frames. Therefore, the present invention uses a special adaptive compression-excitation convolutional neural network to simultaneously model the relationship between feature channels and time frames, such as Figure 4 As shown. For a given input H∈R N×K×S The upper part compresses the spatial dependency of the feature map through a global average pooling layer, and then scales the feature channels through a 1×1 convolutional layer and a Sigmoid layer and assigns the importance of each channel to obtain the weight matrix A. ch ∈R N×1×1 The calculation process of channel attention is shown in formula (9):

[0070] A ch =σ(W C (δW C / r (f Gap (H)))) (9)

[0071] Here Refers to the operation of the global average pooling layer, W C / r and W C is the weight matrix of two 1×1 convolutional layers, δ refers to the nonlinear activation unit ReLU, and σ is the Sigmoid layer. The obtained weight matrix A ch The output of the channel attention network can be obtained by multiplying it with the input H. The lower part uses similar operations for the K dimension. The calculation process of temporal attention is shown in formula (10):

[0072] A t =σ(W N (δW N / r (fGap (H))) (10)

[0073] The original input features, channel attention features, and time attention features are respectively passed through three 1×1 convolutional layers to obtain different degrees of adaptive weighting, and the obtained features are added to obtain the final output features. The calculation process of adaptive weighting is shown in formula (11):

[0074] Output=W α ·A ch ·H+W β ·A t ·H+W γ ·H (11)

[0075] Here W α , W β , W γ are the weight matrices of the convolution, A ch and A t They are the weight matrices of channel attention and temporal attention, respectively. By combining the temporal attention features with the channel attention features, the temporal correlation and the correlation between channels are captured simultaneously, and the network obtains a better feature representation.

[0076] exist Figure 2 The feature scales contained in the outputs of different stages in the multi-stage network described in the figure are not the same. Therefore, different weight values ​​are assigned to the estimated features of the outputs of different stages, and the loss function is calculated respectively with the reference value.

[0077] The details of the multi-stage multi-scale weight loss function in the embodiment of the present invention are shown in formula (12).

[0078]

[0079] Among them, i is the number of stages, loss is the si-snr scale-invariant signal-to-noise ratio, is the weight value. In the loss calculation, deeper layers are assigned larger weights, that is, larger weights are assigned in the loss calculation. Giving deeper layers a larger loss weight can stabilize the training process. When updating the parameters of the separation network model based on the weight loss function, it also includes optimizing the parameters of the separation network model based on fine-grained time domain detail features and better feature representation.

[0080] In the process of training the separation network model, the mixed source time domain signal is used as the input of the separation network, and the waveform signal of the clean target sound source is used for multi-stage supervision of the network. In addition, the samples used for training should be as high-quality and diverse as possible to ensure that the network has good generalization of the data.

[0081] Step train2: Save the trained network model parameters, that is, obtain the trained separation network model.

[0082] S3. Use the trained separation network model to separate the sound source of the single-channel mixed signal.

[0083] The goal of single-channel sound source separation is to isolate the target sound source signal from the single-channel mixed signal.

[0084] The specific process is: according to the estimated feature masking matrix of multiple target sound sources and the embedding vector of the single-channel mixed signal, the feature vectors of multiple target sound source signals are obtained, and the waveform signals of multiple target sound source signals are obtained through the decoder to complete the separation. Among them, the process of obtaining the estimated feature masking matrix of multiple target sound sources is: in the multi-stage mixed attention mechanism network, multiple stage mixed attention submodules are included, and the output of the last stage mixed attention submodule in the multi-stage mixed attention network is selected as the feature masking estimation matrix of multiple target sound sources. The process of extracting feature vectors of a single stage mixed attention module in multiple stage mixed attention submodules is: obtaining local and global time domain features through different types of attention mechanisms in the mixed attention module, obtaining local features and global features according to local and global time domain features, fusing local features and global features, and obtaining feature vectors after this stage mixed attention submodule.

[0085] To sum up, the advantages of the embodiments of the present invention are: it realizes the separation of multiple target sound source signals from a single-channel mixed signal, and compared with the current mainstream single-channel multi-sound source separation model, it improves the performance of sound source separation and the generalization of the model, optimizes the network model parameters, shortens the calculation time, reduces the training burden, and provides scalability for scenarios containing different sound sources (the same model can be applied to multiple different task scenarios).

[0086] The embodiments described above are only descriptions of the preferred modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A sound source separation method based on shallow feature reactivation and multi-stage mixed attention, characterized in that: The following steps are involved: Collecting a single-channel mixed signal and a clean target sound source signal, wherein the single-channel mixed signal corresponds to the clean target sound source signal, and processing the single-channel mixed signal to obtain a single-channel mixed signal time domain embedding vector; Calculating a weighted loss function based on the clean target sound source signal; Construct a separation network model, input the single-channel mixed signal time domain embedding vector into the separation network model, and train the separation network model using the weight loss function to obtain a trained separation network model: The single-channel mixed signal time domain embedding vector is input into the separation network model, and the estimated feature masking matrix of multiple target sound sources is obtained in combination with the shallow feature reactivation mechanism. Based on the estimated feature masking matrix of the multiple target sound sources, the weight loss function is used to update the parameters of the separation network model to obtain the separation network model, wherein the shallow feature reactivation mechanism uses the difference in feature information contained in the time domain feature vectors generated at different stages to prune and fuse shallow features and deep features to extract time domain features; wherein the estimated feature masking matrix of the target sound source is obtained using a stage mixed attention mechanism network, wherein the multi-stage mixed attention mechanism network includes multiple stage mixed attention submodules, and the acquisition process includes: selecting the output of the last stage mixed attention submodule as the feature masking estimation matrix of the multiple target sound sources; The trained separation network model is used to perform sound source separation on the single-channel mixed signal.

2. The sound source separation method based on shallow feature reactivation and multi-stage mixed attention according to claim 1, characterized in that: In the process of processing the single-channel mixed signal, the single-channel mixed signal is sent to an encoder, and a one-dimensional convolutional neural network is used to process the single-channel mixed signal to obtain an embedding vector of the single-channel mixed signal.

3. The sound source separation method based on shallow feature reactivation and multi-stage mixed attention according to claim 1, characterized in that: Using the trained separation network model to perform sound source separation on the single-channel mixed signal includes: Based on the estimated characteristic masking matrices of the multiple target sound sources and the single-channel mixed signal time domain embedding vector, characteristic vectors of the multiple target sound source signals are obtained, and decoding is performed to obtain waveform signals of the multiple target sound source signals to complete separation.

4. The sound source separation method based on shallow feature reactivation and multi-stage mixed attention according to claim 1, characterized in that: The separation network model includes a hybrid attention module, which is used to obtain a better fine-grained feature representation for the single-channel mixed signal time domain embedding vector.

5. The sound source separation method based on shallow feature reactivation and multi-stage mixed attention according to claim 4 is characterized in that: The hybrid attention module includes a multi-head self-attention sub-unit, an element attention sub-unit and an adaptive compression excitation sub-unit.

6. The sound source separation method based on shallow feature reactivation and multi-stage hybrid attention according to claim 5, characterized in that: Obtaining the better fine-grained feature representation includes: obtaining the correlation on the mixed signal time series through the multi-head self-attention subunit, combining the element attention subunit and the adaptive compression excitation subunit to obtain the feature correlation between points and channels of the mixed signal, and obtaining the better fine-grained feature representation.

7. The sound source separation method based on shallow feature reactivation and multi-stage mixed attention according to claim 6, characterized in that: When updating the parameters of the separation network model based on the weight loss function, it also includes optimizing the parameters of the separation network model based on the fine-grained time domain detail features and the more optimal fine-grained feature representation.

8. The sound source separation method based on shallow feature reactivation and multi-stage mixed attention according to claim 1, characterized in that: Extracting feature vectors of a single-stage mixed attention module in the multiple-stage mixed attention sub-modules includes: obtaining time domain features through different types of attention mechanisms in the single-stage mixed attention module, wherein the time domain features include local time domain features and global time domain features, obtaining local features and global features according to the time domain features, fusing the local features and the global features to obtain a feature vector of the single-stage mixed attention sub-module.

Images