A single-channel speech echo cancellation method based on time domain neural network

Abstract

This invention discloses a single-channel speech echo cancellation method based on a temporal neural network, belonging to the field of echo cancellation technology. The method includes framing and normalizing the acquired microphone signal, extracting temporal frame sequence features, and establishing and training a temporal neural network model. The acquired microphone signal is input into the temporal neural network model for performance analysis, and processed using U-Net encoding and decoding to output the echo-cancelled speech sequence. This invention improves the training efficiency and stability of the neural network by extracting consistent temporal features; it constructs and trains a temporal neural network based on the Wave-U-Net architecture, enhancing the echo cancellation effect and adaptability to complex environments through grouped convolution and attention mechanisms; it inputs the signal into the model for performance analysis, and extracts and reconstructs features using the U-shaped encoding and decoding structure of U-Net, utilizing skip connections to retain feature information, achieving efficient echo cancellation and speech reconstruction, and improving speech clarity and naturalness.

Description

Technical Field

[0001] The invention relates to the technical field of echo cancellation, in particular to a single-channel speech echo cancellation method based on a time-domain neural network. Background Art

[0002] With the rapid development of communication technology, electronic communication devices have become an indispensable part of our lives and work, and two-way communication platforms such as Zoom, Skype and Tencent Conference have greatly improved communication efficiency and convenience. However, with the popularization of communication technology and the increase in frequency of use, users' requirements for call quality have also increased, among which voice communication quality has become one of the key factors. In two-way calls, the problem of acoustic echo is particularly prominent, which not only interferes with normal communication, but also seriously affects the clarity of the call, especially in hands-free mode.

[0003] To solve this problem, acoustic echo cancellation technology came into being. This technology aims to identify and eliminate echoes in calls, thereby improving call quality. Although acoustic echo cancellation technology has made significant progress, in practical applications, how to effectively eliminate echoes, deal with background noise and nonlinear distortion, and achieve real-time echo cancellation are still key challenges. Therefore, researching and developing more efficient, stable, and adaptable acoustic echo cancellation solutions has become one of the important research directions in the field of communication technology.

[0004] In this context, a single-channel voice echo cancellation method based on time domain neural network is proposed. This method uses neural network to process voice signals in the time domain to identify and eliminate echoes, thereby improving call quality. Therefore, the single-channel voice echo cancellation method based on time domain neural network is expected to become one of the important development directions in the field of future communication technology, providing users with a clearer and more stable call experience. Summary of the invention

[0005] In view of the problems existing in the existing acoustic echo cancellation technology, the present invention is proposed.

[0006] Therefore, the problem to be solved by the present invention is how to improve the processing capability of acoustic echo cancellation technology, especially in processing background noise and nonlinear distortion, while ensuring stability and effect in real-time calls.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] In the first aspect, an embodiment of the present invention provides a single-channel speech echo elimination method based on a time domain neural network, which includes framing and numerically normalizing the collected microphone signal to extract time domain frame sequence features; establishing and training a time domain neural network model based on the time domain frame sequence features; inputting the collected microphone signal into the time domain neural network model for performance analysis, and outputting a speech sequence after echo elimination through U-Net's U-type encoding and decoding processing; the microphone signal includes a near-end speech signal and a far-end speech signal; the time domain frame sequence features include near-end time domain sequence and far-end time domain sequence features.

[0009] As a preferred solution of the single-channel speech echo elimination method based on time domain neural network of the present invention, wherein: the framing includes frame length and frame shift; the extraction of time domain frame sequence features includes the following steps: collecting microphone signals Divide into several frames, set the sample points for the frame length and frame shift of the frame division; the relevant formula of the frame division is as follows:

[0010] ;

[0011] in, is the microphone signal collected in the i-th frame, n is the sample point index in the current frame, , i is the number of sub-frames, .

[0012] At the same time, numerical normalization is performed on several frames to extract the features of the near-end time domain sequence and the far-end time domain sequence, and a time domain neural network model is constructed; the specific formula of the numerical normalization is as follows:

[0013] ;

[0014] in, is the microphone signal collected at the i-th frame after normalization, is the mean of the i-th frame, and i is the number of sub-frames. As a preferred solution of the single-channel speech echo elimination method based on time domain neural network described in the present invention, wherein: the construction of the time domain neural network model includes the following steps: selecting the end-to-end time domain network model Wave-U-Net as the baseline architecture; in the time domain echo elimination network, the encoder and the decoder both use the group convolution technology to set the high-resolution feature extraction module; the near-end speech signal and the far-end speech signal are distinguished and processed through the coding path, the auxiliary coding path and the attention fusion mechanism; the output of the auxiliary encoder and the output of the previous layer encoder are fused through the attention fusion mechanism to fuse the time domain frame sequence features, and transmitted to the encoder of the current layer through the jump connection for encoding feature extraction; the output features of the encoder are spliced ​​with the decoder of the same layer through the jump connection, and the feature extraction is performed through the same decoding operation as the encoder, and the speech sequence features after echo elimination are calculated layer by layer.

[0015] As a preferred solution of the single-channel speech echo elimination method based on time domain neural network described in the present invention, the high-resolution feature extraction module includes the following steps: the near-end speech signal input feature is subjected to a one-dimensional convolution with a convolution kernel size of 1 and a ReLU activation function operation to obtain a new feature tensor; the new feature tensor is grouped by channel so that the input feature matches the number of output channels; for each grouped feature, a convolution layer with a convolution kernel size of 3 and a ReLU activation operation are performed to obtain a number of processed features; the several processed features are spliced ​​to obtain a merged feature tensor, and batch normalization and ReLU activation operations are performed; the activated feature tensor and the input feature are added through a residual connection to obtain the output feature of the final high-resolution feature extraction module.

[0016] As a preferred solution of the single-channel speech echo elimination method based on time domain neural network described in the present invention, the attention fusion mechanism includes the following steps: the input time domain frame sequence features are processed by one-dimensional convolution with a convolution kernel size of 1 and batch normalization layer operation, and the convolved features are spliced ​​in the channel dimension; the spliced ​​features are processed by LeakyReLU activation function, and the one-dimensional convolution operation with a convolution kernel size of 1 is performed again, and the attention weight is obtained by Sigmoid activation function; the attention weight is dynamically adjusted according to the current task of extracting echo features from remote features; the remote speech signal is processed by dual paths to extract the multi-dimensional features of the remote speech signal; the dual paths include: a main encoding path and an auxiliary encoding path; when the remote When the correlation of the multidimensional features of the speech signal is high at a certain moment, the attention fusion attention mechanism increases the feature weights transmitted from the auxiliary path; when the correlation of the multidimensional features of the remote speech signal is low at a certain moment, the feature weight values ​​transmitted by the auxiliary path tend to zero; in the feature splicing stage, the output size of the main encoding path is adjusted to the output size of the auxiliary path by downsampling; the feature stream is mixed using a one-dimensional convolutional layer and processed through the LeakyReLU activation function; the initial remote features and the attention weights are multiplied to obtain the output features of the final fusion attention mechanism; based on the output features of the final high-resolution feature extraction module and the output features of the final attention fusion attention mechanism, the final echo cancellation effect is output; the final echo cancellation effect of the time domain neural network model is evaluated through the data set.

[0017] As a preferred solution of the single-channel speech echo elimination method based on time domain neural network described in the present invention, wherein: the time domain neural network model includes an encoder, an auxiliary encoder and a decoder, wherein the encoder and the auxiliary encoder are the same, both including a one-dimensional convolutional layer, a batch normalization layer, an ELU activation function and a high-resolution feature extraction module; the encoder includes a downsampling layer and a bottleneck layer; the decoder includes an upsampling layer and an output layer.

[0018] As a preferred solution of the single-channel speech echo elimination method based on time domain neural network described in the present invention, wherein: the downsampling layer performs interval sampling of the output three-dimensional tensor in the third dimension with a step size of 2; the upsampling layer expands the output three-dimensional tensor by two times along the third dimension, and fills it using a linear interpolation method while keeping the corner points aligned; the bottleneck layer includes a one-dimensional convolution layer with a convolution kernel size of 15 and a step size of 7, a batch normalization layer and a ReLU activation function layer; the output layer includes a one-dimensional convolution layer with a convolution kernel size of 1 and a step size of 1, a normalization layer and an activation function layer.

[0019] In the second aspect, an embodiment of the present invention provides a single-channel speech echo cancellation system based on a time domain neural network, which includes: a feature extraction module, which is used to frame and numerically normalize the collected microphone signal, and extract the time domain frame sequence features; a construction module, which establishes and trains an AEC neural network model based on the time domain frame sequence features; an output module, which inputs the collected microphone signal into the time domain neural network model for performance analysis, and outputs the speech sequence after echo cancellation through the U-Net U-type codec processing.

[0020] In a third aspect, an embodiment of the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, the steps of the single-channel speech echo elimination method based on time domain neural network as described in the first aspect of the present invention are implemented.

[0021] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, the steps of the single-channel speech echo elimination method based on a time domain neural network as described in the first aspect of the present invention are implemented.

[0022] The beneficial effects of the present invention are as follows: the present invention extracts consistent time domain features by framing and normalizing microphone signals, thereby improving the training efficiency and stability of the neural network; constructs and trains a time domain neural network based on the Wave-U-Net architecture, and enhances the effect of echo cancellation and the ability to adapt to complex environments through grouped convolution and attention mechanisms; inputs the signal into the model for performance analysis, and extracts and reconstructs features through the U-Net U-type codec structure, and uses jump connections to retain feature information, thereby achieving efficient echo cancellation and speech reconstruction, and improving speech clarity and naturalness. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. 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 creative work. Among them:

[0024] Figure 1 This is a flow chart of a single-channel speech echo elimination method based on a time domain neural network in Example 1.

[0025] Figure 2 This is a deep learning echo cancellation network flow chart of Example 1 of a single-channel speech echo cancellation method based on a time domain neural network.

[0026] Figure 3This is an overall block diagram of the single-channel speech echo elimination method based on time domain neural network in Example 1.

[0027] Figure 4 This is a structural diagram of a high-resolution feature extraction module of a single-channel speech echo elimination method based on a time-domain neural network in Example 1.

[0028] Figure 5 Example 1 is a group convolution structure diagram of a single-channel speech echo elimination method based on a time domain neural network.

[0029] Figure 6 This is a structural diagram of the dual-path fusion attention mechanism of the single-channel speech echo elimination method based on time domain neural network in Example 1.

[0030] Figure 7 This is a PESQ performance diagram of Example 2 testing the single-channel speech echo cancellation method based on time domain neural network in a linear echo scenario.

[0031] Figure 8 This is a STOI performance diagram of the single-channel speech echo cancellation method based on time domain neural network in Example 2 tested in a linear echo scenario.

[0032] Fig. 9 This is a PESQ performance diagram of the single-channel speech echo cancellation method based on time domain neural network in Example 2 tested in a mixed linear and nonlinear echo scenario.

[0033] Fig.10 Example 2 is a STOI performance diagram of a single-channel speech echo cancellation method based on a time domain neural network tested in a mixed linear and nonlinear echo scenario.

[0034] Fig.11 Example 2 is a performance difference diagram of the single-channel speech echo elimination method of the present invention based on a time domain neural network.

[0035] Fig.12 Example 2 is a performance difference diagram of the existing echo cancellation algorithm of the single-channel speech echo cancellation method based on the time domain neural network DETAILED DESCRIPTION

[0036] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0037] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0038] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0039] Example 1

[0040] Reference Figure 1 to Figure 6 , which is the first embodiment of the present invention, provides a single-channel speech echo elimination method based on a time domain neural network, comprising:

[0041] S1: Frame the collected microphone signal and perform numerical normalization to extract the time domain frame sequence features.

[0042] Specifically, the collected microphone signal Divide into several frames, set the frame length and frame shift of the frame division to 16384 samples, and the relevant formula for the frame division is as follows:

[0043] ;

[0044] in, is the microphone signal collected in the i-th frame, n is the sample point index in the current frame, , i is the number of sub-frames, .

[0045] Furthermore, several frames are numerically normalized at the same time to extract the features of the near-end time domain sequence and the far-end time domain sequence, and to construct a time domain neural network model; so that the values ​​within the frame are distributed in a specific range, such as [−1, 1]. The specific formula for numerical normalization is as follows:

[0046] ;

[0047] in, for, is the mean of the i-th frame, and i is the number of frames.

[0048] Furthermore, the microphone signal includes a near-end speech signal and a far-end speech signal; and the time-domain frame sequence features include near-end time-domain sequence features and far-end time-domain sequence features.

[0049] S2: Establish and train a time domain neural network model based on the time domain frame sequence features.

[0050] Specifically, the time domain neural network model includes an encoder, an auxiliary encoder and a decoder, wherein the encoder and the auxiliary encoder are the same, both including a one-dimensional convolutional layer, a batch normalization layer, an ELU activation function and a high-resolution feature extraction module; the encoder includes a downsampling layer and a bottleneck layer; the decoder includes an upsampling layer and an output layer.

[0051] Furthermore, the specific formula of the ELU activation function is as follows:

[0052] ;

[0053] in, It is a configurable parameter, set to 1. is the value input to the current neuron.

[0054] Specifically, both the encoder and the decoder include a one-dimensional convolution layer with a convolution kernel size of 15 and a stride size of 7; the downsampling layer of the 12 encoding layers samples the output three-dimensional tensor at intervals of 2 in the third dimension; the upsampling layer of the 12 decoding layers expands the output three-dimensional tensor by two times along the third dimension and fills it with a linear interpolation method while keeping the corners aligned.

[0055] .Furthermore, the bottleneck layer includes a one-dimensional convolution layer with a convolution kernel size of 15 and a stride of 7, a batch normalization layer, and a ReLU activation function layer; the output layer includes a one-dimensional convolution layer with a convolution kernel size of 1 and a stride of 1, a normalization layer, and an activation function layer.

[0056] Furthermore, the relevant formula for the ReLU activation function layer is as follows:

[0057] ;

[0058] in, is the value input to the current neuron.

[0059] Specifically, constructing the time domain neural network model includes the following steps: selecting the end-to-end time domain network model Wave-U-Net as the baseline architecture; in the time domain echo cancellation network, both the encoder and the decoder use group convolution technology to set up a high-resolution feature extraction module; distinguish and process near-end speech signals and far-end speech signals through encoding paths, auxiliary encoding paths and attention fusion mechanisms; fuse the output of the auxiliary encoder and the output of the previous layer encoder through the attention fusion mechanism to fuse the time domain frame sequence features, and transmit them to the encoder of the current layer through a jump connection for encoding feature extraction; splicing the output features of the encoder with the decoder of the same layer through a jump connection, and extracting features through the same decoding operation as the encoder, and calculating the speech sequence features after echo cancellation layer by layer.

[0060] S3: The collected microphone signal is input into the time domain neural network model for performance analysis, and the U-net U-type codec is processed to output the speech sequence after echo cancellation.

[0061] Specifically, the high-resolution feature extraction module includes the following steps: the near-end speech signal input features are subjected to a one-dimensional convolution with a convolution kernel size of 1 and a ReLU activation function operation to obtain a new feature tensor; the new feature tensor is grouped by channel so that the input features match the number of output channels; for each grouped feature, a convolution layer with a convolution kernel size of 3 and a ReLU activation operation are performed to obtain a number of processed features; the several processed features are concatenated to obtain a merged feature tensor, and batch normalization and ReLU activation operations are performed; the activated feature tensor and the input feature are added through a residual connection to obtain the output features of the final high-resolution feature extraction module.

[0062] Furthermore, the attention fusion mechanism includes the following steps: performing feature processing on the input time domain frame sequence features through one-dimensional convolution with a convolution kernel size of 1 and batch normalization layer operations, and splicing the convolved features in the channel dimension; processing the spliced ​​features through the LeakyReLU activation function, performing a one-dimensional convolution operation with a convolution kernel size of 1 again, and obtaining the attention weight through the Sigmoid activation function; dynamically adjusting the attention weight according to the current task of extracting echo features from remote features.

[0063] Furthermore, the far-end speech signal is processed through dual paths to extract the multidimensional features of the far-end speech signal; the dual paths include: a main coding path and an auxiliary coding path; when the correlation of the multidimensional features of the far-end speech signal is high at a certain moment, the attention fusion attention mechanism increases the feature weight transmitted from the auxiliary path; when the correlation of the multidimensional features of the far-end speech signal is low at a certain moment, the feature weight value transmitted by the auxiliary path tends to zero; in the feature splicing stage, the output size of the main coding path is adjusted to the output size of the auxiliary path by downsampling; the feature stream is mixed using a one-dimensional convolutional layer and processed through a LeakyReLU activation function; the initial far-end features and the attention weights are multiplied to obtain the output features of the final fusion attention mechanism; based on the output features of the final high-resolution feature extraction module and the output features of the final attention fusion attention mechanism, the final echo cancellation effect is output; the final echo cancellation effect of the time domain neural network model is evaluated through a data set.

[0064] Furthermore, the performance analysis includes Perceptual Evaluation of Speech Quality (PESQ), Short-Time Objective Intelligibility (STOI) and Scale-Invariant Signal-to-Distortion Ratio.

[0065] Furthermore, this embodiment also provides a single-channel speech echo cancellation system based on a time domain neural network, including: a feature extraction module, which is used to frame and numerically normalize the collected microphone signal, and extract the time domain frame sequence features; a construction module, which establishes and trains an AEC neural network model based on the time domain frame sequence features; an output module, which inputs the collected microphone signal into the time domain neural network model for performance analysis, and outputs the speech sequence after echo cancellation through the U-Net U-type codec processing.

[0066] This embodiment also provides a computer device, which is applicable to the case of a single-channel voice echo elimination method based on a time-domain neural network, and includes a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the single-channel voice echo elimination method based on a time-domain neural network as proposed in the above embodiment.

[0067] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covered on the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0068] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the following steps are implemented: framing and numerically normalizing the collected microphone signal to extract time domain frame sequence features; establishing and training a time domain neural network model based on the time domain frame sequence features; inputting the collected microphone signal into the time domain neural network model for performance analysis, and outputting a speech sequence after echo cancellation through U-Net encoding and decoding processing.

[0069] In summary, the present invention extracts consistent time domain features by framing and normalizing microphone signals, thereby improving the training efficiency and stability of the neural network; constructs and trains a time domain neural network based on the Wave-U-Net architecture, and enhances the effect of echo cancellation and the ability to adapt to complex environments through group convolution and attention mechanisms; inputs the signal into the model for performance analysis, and extracts and reconstructs features through the U-Net U-type codec structure, uses jump connections to retain feature information, achieves efficient echo cancellation and speech reconstruction, and improves speech clarity and naturalness.

[0070] Example 2

[0071] Reference Figure 7 to Figure 12 , which is the second embodiment of the present invention, provides a single-channel speech echo elimination method based on time domain neural network. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0072] Specifically, the Pytorch framework was used to build the model, the batch size was set to 96, and the mean square error (MSE) was used as the loss function. The model parameters were optimized by the Adam optimizer, where the optimizer parameters were set to β1=0.9, β2=0.999, and the learning rate was 0.001. The input data of the model consisted of a time series of far-end speech frames and near-end microphone speech frames with a frame length of 16384. These data had been normalized before being input into the network for training. Under the same training set conditions, the baseline Wave-U-Net model and the time domain network with a high-resolution feature extraction module and a dual-path fusion attention mechanism were compared. The model used the same training framework and loss function during training, and the number of iterations was set to 300. The echo cancellation effect of the model was evaluated through the test set, and the performance indicators were carefully analyzed and compared.

[0073] Furthermore, as shown in Table 1, the time domain network proposed in the present invention significantly improves the performance of echo cancellation by introducing a new module. The network combines a high-resolution feature extraction module and a dual-path fusion attention mechanism, and under the same conditions, the PESQ and STOI scores are improved by 9% and 1.8%, respectively, effectively improving the feature capture and sequence processing capabilities of the network.

[0074] Table 1 Comparison table between the present invention and the prior art:

[0075]

[0076] Furthermore, by comparing the experiments under different echo scenarios, the echo is subdivided into linear echo and nonlinear echo when mathematically modeling. Figure 7 and Figure 8 As shown in the figure, the PESQ and STOI values ​​of the model increase with the increase of the signal-to-echo ratio under the condition of only linear echo, indicating that the signal-to-echo ratio is the key factor to improve the echo cancellation effect. By introducing the high-resolution feature extraction module and the attention fusion mechanism, the performance is improved under different signal-to-echo ratios, showing the effectiveness of processing echo. Fig. 9 and Fig.10 As shown, even in a more complex environment containing linear and nonlinear echoes, the time domain network proposed in the present invention still achieves excellent performance in the complex environment.

[0077] Specifically, the data sets used are all ICASSP AEC Challenge data sets to verify the performance difference between the present invention and the existing echo cancellation algorithm, such as Fig.11 As shown in the figure, the results respectively demonstrate the performance of the proposed method and various current echo cancellation algorithms in terms of the two main performance indicators of PESQ and STOI.

[0078] Furthermore, the proposed time domain network has the highest score among all compared algorithms in the PESQ score, which emphasizes its superior performance in improving speech quality. The improvement effect of the proposed method is particularly significant when compared with earlier algorithms such as LMS and NLMS, and its advantages are also reflected in the comparison with deep learning algorithms such as Conv-TasNet and Attention-Wave-U-Net. This reflects the efficiency of the proposed method in capturing and processing speech signals, as well as its accuracy in eliminating echoes from complex signals. The superior performance of the proposed time domain network is also demonstrated in the STOI score. Compared with other methods, including the recently proposed GAM-DSPM-CRN and CS-BiLSTM algorithms, the proposed method still occupies a leading position, indicating its effectiveness in maintaining speech clarity and intelligibility.

[0079] Furthermore, combined with the analysis of the two indicators of PESQ and STOI, it can be clearly seen that the proposed method has significantly improved speech quality and clarity compared with the existing technology. This demonstrates the superiority of this method in handling echo cancellation tasks.

[0080] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A single-channel speech echo cancellation method based on time domain neural network, characterized in that: include, The collected microphone signal is framed and numerically normalized to extract the time domain frame sequence features; Establish and train a time domain neural network model based on time domain frame sequence features; The collected microphone signal is input into the time domain neural network model for performance analysis, and the U-Net U-type codec is used to output the speech sequence after echo cancellation. The microphone signal includes a near-end voice signal and a far-end voice signal; The time domain frame sequence features include near-end time domain sequence features and far-end time domain sequence features; The framing includes frame length and frame shift; and the extraction of time domain frame sequence features includes the following steps: The microphone signal collected Divide into several frames, and set the sample points for the frame length and frame shift of the divided frames; The relevant formula for the framing is as follows: ； in, is the microphone signal collected in the i-th frame, is the sample index in the current frame, , i is the number of sub-frames, ; At the same time, several frames are numerically normalized, the features of the near-end time domain sequence and the far-end time domain sequence are extracted, and a time domain neural network model is constructed; The specific formula for the numerical normalization is as follows: ； in, is the microphone signal collected at the i-th frame after normalization, is the mean of the i-th frame, i is the number of sub-frames; The construction of the time domain neural network model comprises the following steps: The end-to-end time domain network model Wave-U-Net was selected as the baseline architecture; In the time-domain echo cancellation network, both the encoder and decoder use grouped convolution techniques to set up high-resolution feature extraction modules; Distinguish and process near-end and far-end speech signals through encoding pathways, auxiliary encoding pathways, and attention fusion mechanisms; The output of the auxiliary encoder and the output of the previous layer encoder are fused through the attention fusion mechanism to fuse the time domain frame sequence features, and transmitted to the encoder of the current layer through the skip connection for encoding feature extraction; The output features of the encoder are concatenated with the decoder of the same layer through skip connections, and feature extraction is performed through the same decoding operation as the encoder, and the speech sequence features after echo cancellation are calculated layer by layer; The high-resolution feature extraction module comprises the following steps: The near-end speech signal input features are subjected to a one-dimensional convolution with a convolution kernel size of 1 and a ReLU activation function operation to obtain a new feature tensor; Group the new feature tensor by channel so that the number of input features matches the number of output channels; For each grouped feature, a convolution layer with a convolution kernel size of 3 and a ReLU activation operation are performed to obtain several processed features; Concatenate several processed features to obtain a merged feature tensor, and perform batch normalization and ReLU activation operations; The activated feature tensor and input features are added through residual connection to obtain the output features of the final high-resolution feature extraction module; The attention fusion mechanism includes the following steps: The input time domain frame sequence features are processed through one-dimensional convolution with a convolution kernel size of 1 and batch normalization layer operations, and the convolved features are spliced ​​in the channel dimension; The concatenated features are processed through the LeakyReLU activation function, and a one-dimensional convolution operation with a convolution kernel size of 1 is performed again, and the attention weight is obtained through the Sigmoid activation function; Dynamically adjust the attention weights according to the current task of extracting echo features from distal features; Process the far-end speech signal through dual paths to extract multi-dimensional features of the far-end speech signal; The dual paths include: a main encoding path and an auxiliary encoding path; When the correlation of multi-dimensional features of the far-end speech signal is high at a certain moment, the attention fusion attention mechanism increases the weight of the features transferred from the auxiliary path; When the correlation of the multi-dimensional features of the far-end speech signal is low at a certain moment, the feature weight value transmitted by the auxiliary path tends to zero; In the feature concatenation stage, the output size of the main encoding path is adjusted to the output size of the auxiliary path by downsampling; Use a one-dimensional convolutional layer to mix the feature streams and process them through the LeakyReLU activation function; Multiply the initial remote features and the attention weights to obtain the final output features of the fused attention mechanism; Based on the output features of the final high-resolution feature extraction module and the output features of the final attention fusion attention mechanism, the final echo cancellation effect is output; The final echo cancellation effect of the time domain neural network model is evaluated through the dataset.

2. The single-channel speech echo cancellation method based on time domain neural network as claimed in claim 1, characterized in that: The time domain neural network model includes an encoder, an auxiliary encoder and a decoder, wherein the encoder and the auxiliary encoder are the same, both including a one-dimensional convolution layer, a batch normalization layer, an ELU activation function and a high-resolution feature extraction module; The encoder includes a downsampling layer and a bottleneck layer; the decoder includes an upsampling layer and an output layer.

3. The single-channel speech echo cancellation method based on time domain neural network as claimed in claim 2, characterized in that: The downsampling layer samples the output three-dimensional tensor at intervals of 2 in the third dimension; the upsampling layer expands the output three-dimensional tensor by two times along the third dimension and fills it using a linear interpolation method while keeping the corners aligned; the bottleneck layer includes a one-dimensional convolution layer with a convolution kernel size of 15 and a step size of 7, a batch normalization layer, and a ReLU activation function layer; the output layer includes a one-dimensional convolution layer with a convolution kernel size of 1 and a step size of 1, a normalization layer, and an activation function layer.

4. A single-channel speech echo cancellation system based on a time domain neural network, based on the single-channel speech echo cancellation method based on a time domain neural network according to any one of claims 1 to 3, characterized in that: include, A feature extraction module is used to perform framing and numerical normalization on the collected microphone signal and extract time domain frame sequence features; Building module, building and training AEC neural network model based on time domain frame sequence features; The output module inputs the collected microphone signal into the time domain neural network model for performance analysis, and outputs the speech sequence after echo cancellation through the U-type codec processing of U-Net.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the single-channel speech echo elimination method based on time domain neural network described in any one of claims 1 to 3 are implemented.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the single-channel speech echo elimination method based on time domain neural network described in any one of claims 1 to 3 are implemented.

Images