Echo cancellation algorithm

Through DNSMOS tool filtering data, progressive training and self-attention mechanism and other technologies, combined with automatic labeling and knowledge distillation, the shortcomings of the existing echo cancellation algorithm in complex environments are solved, and high-precision and strong generalization echo cancellation effect is achieved. .

CN119007737APending Publication Date: 2024-11-22SHANGHAI SHENGHAN INFORMATION TECH CO LTD
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411075719.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-07
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

Existing echo cancellation algorithms are not effective when dealing with long delays, complex noise and reverberation environments, high data quality requirements, inaccurate delay estimation, insufficient alignment module performance, limited echo cancellation capabilities in complex environments, and lack effective joints. Training mechanism.

Method used

DNSMOS tool is used to filter high-quality data, adopt a progressive training method, use self-attention mechanism and deep one-dimensional convolution to extract features, combine automatic labeling and knowledge distillation, and combine multiple auxiliary information to distillate teacher-student knowledge, optimize alignment Module and loss function, audio state detection function is introduced.

Benefits of technology

It significantly improves the accuracy and generalization ability of echo cancellation, ensures the echo cancellation effect in complex environments, improves the stability and adaptability of the model, and enhances the user's call experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119007737A_ABST
    Figure CN119007737A_ABST
Patent Text Reader

Abstract

The invention discloses an echo cancellation algorithm, which comprises the following steps: S1, introducing a DNSMOS tool to screen high-quality data, and carrying out noise reduction preprocessing on data with relatively low MOS score; s2, adopting a progressive training method; s3, an alignment module based on a cross attention mechanism is adopted, and local features are extracted through depth one-dimensional convolution to improve the performance of the alignment module; s4, generating an echo time delay label by adopting an automatic labeling method; s5, calculating relative delay probability distribution; s6, improving the echo cancellation capability in a complex environment by adopting a knowledge distillation algorithm; s7, performing joint processing through delay alignment and filter coefficient estimation, and compressing the feature dimension through affine transformation to obtain a filter coefficient; s8, using an SI-SNR loss function and a PLCPA-ASYM loss function to optimize the output signal quality of the noise reduction model; and S9, introducing an audio state detection function. According to the method, the traditional signal processing and deep learning technologies are combined, an improved echo cancellation algorithm is provided, and the method has high precision, strong generalization and complex environment adaptability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and signal processing, and in particular to an echo cancellation algorithm. Background Art

[0002] With the rapid development of modern communication technology, especially the widespread use of mobile smart terminals, Internet telephony has become an important means of communication in people's daily life and work. However, the common echo problem in Internet telephony has seriously affected the call quality and user experience. Echoes are mainly divided into acoustic echo and line echo: line echo is caused by matching coupling between lines, while acoustic echo is caused by the sound of the speaker being fed back to the microphone multiple times during an external voice call. Due to various reasons, there will be an unstable long delay between the echo and the reference signal, which brings great challenges to the echo cancellation algorithm.

[0003] Traditional echo cancellation algorithms are mainly based on signal processing technology. These algorithms usually divide the long-delay echo cancellation task into multiple modules for optimization, including delay estimation, double talk estimation, linear echo cancellation, and nonlinear residual suppression. However, this method has certain limitations. Under noisy and reverberant conditions, the delay estimation effect is poor and the performance is unstable, which affects the overall call quality. In addition, these traditional methods have limited echo cancellation capabilities in complex environments and cannot effectively handle the changing communication environment in practical applications.

[0004] In recent years, with the development of deep learning technology in the field of speech enhancement and signal processing, more and more end-to-end echo cancellation methods based on deep neural networks have been proposed. Compared with traditional methods, these deep learning methods not only improve speech enhancement performance, but also make the model structure more compact and simple. However, existing deep learning methods still face many challenges in dealing with long delays, complex noise and reverberation environments. Specifically, these methods have defects in the following aspects:

[0005] 1. Data quality issues: Existing deep learning models have high requirements for the quality of training data, but in practical applications, it is difficult to obtain high-quality training data. Existing methods usually fail to fully consider the subjective auditory quality of the data, resulting in a large number of low-quality samples in the training data, which affects the training effect and generalization ability of the model.

[0006] 2. Inaccurate delay estimation: Traditional delay estimation methods, such as the generalized cross-correlation method (GCC-PHAT), are unstable in low signal-to-noise ratio and complex environments, resulting in inaccurate delay estimation, which in turn affects the echo cancellation effect. Although existing deep learning methods have improved delay estimation performance under certain conditions, they still cannot achieve ideal results under long delay and non-steady-state conditions.

[0007] 3. Inadequate performance of the alignment module: In existing deep learning echo cancellation methods, the alignment module is usually based on a simple feature extraction method and fails to fully utilize the local feature information of the signal, resulting in insufficient alignment accuracy and affecting the overall echo cancellation effect.

[0008] 4. Limited echo cancellation capability in complex environments: Existing echo cancellation methods have insufficient generalization capabilities when faced with complex real-world scene data. Especially in complex environments such as long delays, noise interference, and reverberation, the echo cancellation effect of existing methods is far from ideal.

[0009] 5. Lack of effective joint training mechanism: Existing methods usually process each module of echo cancellation independently and fail to make full use of various auxiliary information for joint optimization, resulting in the model's less than ideal effect when dealing with echo cancellation tasks in complex environments.

[0010] Therefore, how to provide an echo cancellation algorithm is a problem that those skilled in the art need to solve urgently. Summary of the invention

[0011] One purpose of the present invention is to propose an echo cancellation algorithm. The present invention proposes an improved echo cancellation algorithm, which combines traditional signal processing technology with deep learning technology, introduces the DNSMOS tool to screen high-quality data, adopts a progressive training method, uses a self-attention mechanism model and deep one-dimensional convolution to extract features, automatically labels and distills knowledge, and other technical means to significantly improve the echo cancellation effect, and has high precision, strong generalization and adaptability to complex environments.

[0012] An echo cancellation algorithm according to an embodiment of the present invention includes the following steps:

[0013] S1. In the data preprocessing stage, the DNSMOS tool is introduced to screen high-quality data, and the data with low MOS scores are preprocessed by using traditional signal processing and neural network models to reduce noise, and high-quality reference signals are screened out again;

[0014] S2. In the data generation stage, a progressive training method is adopted. Clean audio is used for pre-training in the initial stage, and then data from easy to difficult is gradually generated for denoising model training;

[0015] S3, adopting an alignment module based on a cross attention mechanism and using deep one-dimensional convolution to extract local features to improve the performance of the alignment module;

[0016] S4. Generate an echo delay label using an automatic labeling method, and filter out low-quality reference signals and echo signals in combination with the DNSMOS tool;

[0017] S5. Calculate the relative delay probability distribution of the reference signal and the echo signal through an alignment module, and explicitly optimize the alignment module in combination with the delay label;

[0018] S6. A new teacher-student knowledge distillation algorithm combining multiple auxiliary information is used. The student model learns the audio output, frame-level voice probability and echo delay probability distribution of the teacher model, improving the echo cancellation capability in real environments.

[0019] S7, during real-time calls, filter coefficients are estimated by combining the self-attention model and the cross-attention alignment model to eliminate echo, noise, and late reverberation;

[0020] S8. The target loss of the noise reduction algorithm uses the SI-SNR loss function and the PLCPA-ASYM loss function. The PLCPA-ASYM loss function performs a second Fourier transform on the audio data after signal restoration, and then calculates the loss value to optimize the output signal quality of the noise reduction model.

[0021] S9. Introduce the voice state detection function, use the short-time energy threshold and neural network model to generate the voice state label, and improve the listening quality of the noise reduction model.

[0022] Optionally, the feature frames all contain local information; the calculation formula of the deep one-dimensional convolution is:

[0023]

[0024] Among them, y t represents the output feature, x t-i represents the output feature, w i represents the convolution kernel weight, k represents the convolution kernel size;.

[0025] Optionally, the S4 specifically includes:

[0026] S41. Use the DNSMOS tool to filter out reference signals and echo signal data with extremely poor quality;

[0027] S42, using the traditional signal processing method GCC-PHAT to generate a delay label of the echo signal relative to the reference signal, the specific formula is as follows:

[0028]

[0029] in, represents the Fourier transform, Represents X i The complex conjugate of (f), represents the inverse Fourier transform, x1(t) represents the microphone time domain signal, and x2(t) represents the reference time domain signal.

[0030] Optionally, the input of the relative delay probability distribution is the local feature of S3, and its calculation formula is

[0031] for:

[0032]

[0033] Among them, corr i,j represents the probability that the echo signal of the i-th frame is delayed by j frames relative to the reference signal, q i,m,1 represents the characteristic value of the echo signal at the i-th frame m, k i,m,j represents the m-bit eigenvalue of the reference signal in the ij frame, p represents the dimension of the eigenvector, and D represents the preset maximum number of delayed frames.

[0034] Optionally, the teacher model uses a large-scale denoising model based on learnable features, alignment modules and global self-attention, and the student model uses a small-scale model based on traditional spectral features, alignment modules and block streaming self-attention to learn the final output audio and estimated frame-level voice probability distribution of the teacher model, as well as the relative delay probability distribution of the echo reference signal in the alignment module, so as to be close to the estimation result of the teacher model.

[0035] The beneficial effects of the present invention are:

[0036] (1) Improved data quality: By introducing the DNSMOS tool to screen high-quality data and performing noise reduction preprocessing on data with low MOS scores, high-quality reference signals are screened out again, ensuring the high quality of training data and significantly improving the training effect and generalization ability of the model.

[0037] (2) Progressive training enhances model applicability: A progressive training method is adopted. Clean audio is used for pre-training in the initial stage, and data from easy to difficult is gradually generated for noise reduction model training. This helps the model better adapt to environments of different complexities and improves the model's echo cancellation capabilities in various scenarios.

[0038] (3) Accurate signal alignment: The end-to-end self-attention mechanism model adopts an alignment module based on the cross-attention mechanism and uses deep one-dimensional convolution to extract local features, which significantly improves the performance of the alignment module, ensures the accuracy and stability of signal alignment, and enhances the echo cancellation effect.

[0039] (4) Reliable delay estimation: An automatic labeling method is used to generate echo delay labels. The high-quality reference data and echo signals screened by the DNSMOS tool are combined to optimize the time alignment module, ensuring the accuracy and stability of delay estimation. It can work effectively even in low signal-to-noise ratio and complex environments.

[0040] (5) Echo cancellation capability in complex environments: The new knowledge distillation algorithm that combines multiple auxiliary information improves the echo cancellation capability in complex environments. Specifically, it combines the teacher model and the student model to make full use of multiple auxiliary information for optimization, ensuring the generalization and robustness of the model in complex environments.

[0041] (6) Optimizing the loss function: The target loss of the noise reduction algorithm uses the SI-SNR loss function and the PLCPA-ASYM loss function. A second Fourier transform is performed on the audio data after signal restoration, and then the loss value is calculated. This optimizes the output signal quality of the noise reduction model, ensures the naturalness and clarity of the signal, and improves the user's call experience.

[0042] (7) Voiced state detection function: The voiced state detection function is introduced to generate voiced state labels through short-time energy thresholds and neural network models, which effectively improves the accuracy of the noise reduction model in detecting voiced segments, further improving the noise reduction effect and the user's listening quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0044] Figure 1 An overall flow chart of an echo cancellation algorithm proposed by the present invention;

[0045] Figure 2 The generation of a speech signal in a room and the echo cancellation processing flow of an echo cancellation algorithm according to an embodiment of the present invention;

[0046] Figure 3 A flow chart of real-time data generation of an echo cancellation algorithm according to an embodiment of the present invention;

[0047] Figure 4 A denoising model training framework for an echo cancellation algorithm according to an embodiment of the present invention;

[0048] Figure 5 A schematic diagram of an alignment model based on an attention mechanism of an echo cancellation algorithm according to an embodiment of the present invention;

[0049] Figure 6 A delay distribution heat map of an echo cancellation algorithm in a real scenario according to an embodiment of the present invention;

[0050] Figure 7 A schematic diagram of a learnable one-dimensional convolution denoising framework of an echo cancellation algorithm according to an embodiment of the present invention;

[0051] Figure 8A schematic diagram of a Fourier transform-based denoising framework of an echo cancellation algorithm according to an embodiment of the present invention;

[0052] Fig. 9 A schematic diagram of a filter estimation model based on a self-attention mechanism of an echo cancellation algorithm according to an embodiment of the present invention;

[0053] Fig.10 The present invention is a flowchart of a denoising training method based on knowledge distillation for an echo cancellation algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0055] The present invention proposes an improved denoising model and data processing method. The DNSMOS tool is introduced to screen high-quality data, and low MOS score data is pre-processed for denoising. A progressive method is used to generate data from easy to difficult to improve the model quality. The end-to-end self-attention model removes the traditional signal processing module, and uses an alignment module based on a cross-attention mechanism and an explicit time alignment training method to enhance the signal alignment and denoising effects. At the same time, the GCC-PHAT method is used to generate high-confidence echo delay duration labels. Finally, through the knowledge distillation method, the student model learns the output audio, voice probability, and signal correlation weights of the teacher model, further improving the model performance and data processing efficiency.

[0056] During a real-time call, the voice signal is composed of the near-end signal, the far-end signal played by the speaker, and the background noise. In order to eliminate the echo signal, it is usually necessary to use the loopback signal (reference signal) to perform linear adaptive echo cancellation. In order to further suppress the residual echo, the post-processing part will also perform non-linear residual echo cancellation based on the correlation between the loopback signal and the residual signal. The generation of voice signals in a room and the echo cancellation process are as follows: Figure 2 shown.

[0057] In a room with background noise, considering only a single microphone and a single speaker, the signal at the microphone is expressed as:

[0058] y(t)=x(t)*h x (t)+n(t)*h n (t)+d(t)*h d (t);

[0059] Where y(t) represents the microphone signal at time t, x(t) represents the near-end clean speech signal at time t, and h x(t) represents the room impulse response at the near-end sound source at time t, n(t) represents the noise signal at time t, h n (t) represents the room impulse response at the noise source at time t, d(t) represents the far-end echo signal, and h d (t) represents the room impulse response at the echo source.

[0060] Taking data quality into consideration, we used the neural network-based DNSMOS model to filter out data with different subjective opinion scores (MOS) levels. When the subjective opinion score is lower than 4.0, we used the full neural network model and the noise reduction module distribution preprocessing of traditional signal processing for further screening. This way, the filtered data can avoid the model being polluted by data of too low quality, and can also obtain enough high-quality data to improve model performance.

[0061] During the data preparation stage, background noise, echo, and late and early room impulse responses are taken into account, and on-the-fly simulation generation is used to enrich the diversity of the data. Considering the need for progressive learning of the model, we also gradually increase the signal-to-noise ratio, signal-to-echo ratio, and room reverberation difficulty. It should also be noted that considering that our acoustic model needs to cover the reverberation suppression function, our target speech retains the complete room early reflection pulse signal (0-20ms), and exponentially decays the room late reflection pulse signal (20-200ms). Our data production process is as follows: Figure 3 shown.

[0062] The denoising algorithm we proposed is an end-to-end method based on deep learning, which consists of three parts: feature extraction, filter estimation, and signal restoration. In addition, considering the delay of the echo signal relative to the reference signal, we also added an alignment module v based on deep learning. We use the relative delay probability distribution calculated by the attention mechanism and the loss function between the echo delay duration labels. Through this explicit time alignment training method, the reference signal and the echo signal can be accurately aligned, thereby improving the denoising effect of the overall model. Among them, the estimation of the echo delay duration label adopts the traditional signal processing method GCC-PHAT to generate, avoiding manual labeling. This method of automatically generating delay labels is still not accurate enough under extremely low signal-to-noise ratios. In order to further ensure the credibility of the label information, we further screened the reference and echo signals (MOS>3.2) using the DNSMOS tool.

[0063] In order to make the output signal quality close to the target signal and improve the listening experience while reducing the problem of over-suppression, our target loss uses both the scale-invariant signal-to-noise ratio SI-SNR loss function and the power-law compressed phase-aware non-parallel PLCPA-ASYM loss function. It should be noted that when calculating the PLCPA-ASYM loss value, we perform a secondary Fourier transform on the audio data after signal restoration and then calculate the loss value.

[0064] In order to further improve the performance of the model and to cope with the need to amplify the digital gain of the voiced segment to improve the listening experience in practical applications, we also introduced the voiced state detection (VAD) function. The voiced state label data is generated based on the short-term energy threshold of the target signal when the quality is good (high MOS score). We use a neural network model with better performance to generate real data. The denoising model training framework is as follows: Figure 4 shown.

[0065] The alignment model we use is based on the cross-attention mechanism, which can greatly improve the echo cancellation performance while simplifying the processing flow. First, considering the time-varying nature of single-frame features, we use deep one-dimensional convolution to abstract the feature information of historical frames to the current feature frame (such as the previous 15 frames) so that each frame feature contains local information. This is conducive to calculating the similarity between the microphone features and the reference features, and then using affine transformation to further compress the current frame features. Finally, the cross-correlation is calculated using the vector dot product method using the reference abstract features and the historical microphone abstract features (such as the previous 100 frames, covering 1s delay), and then normalized (Softmax) to obtain the relative delay probability distribution of each frame. The aligned reference features are obtained by weighted summing on the time axis and the corresponding delay probability distribution. Alignment models based on the attention mechanism are as follows: Figure 5 shown.

[0066] By observation Figure 6 The delay distribution heat map in the real scene shows that the correlation of the alignment model has a more obvious horizontal line (larger value) in a certain delay interval. By using the weighted summation of the delay distribution, the aligned reference signal can be obtained to greatly improve the echo cancellation performance.

[0067] The first row is the input signal, the second row is the reference signal, the third row is the output signal, and the fourth row is the delay distribution heat map.

[0068] The echo cancellation method we constructed combines delay alignment and filter coefficient estimation to eliminate echo, noise, and late reverberation at the same time. For feature extraction, we used a learnable one-dimensional convolution. By increasing the output dimension of the one-dimensional convolution, the listening experience can be greatly improved. We retained the traditional Fourier transform frequency domain feature as an option.

[0069] Based on the learnable one-dimensional convolution front end, we use a frame length (kernel size) of 20ms, a frame shift (stridesize) of 10ms, and a microphone signal output channel number (channels) of 2048 (can also be 512, 1024, etc.), and the reference signal output channel number is 128. Considering the difficulty of training learnable one-dimensional convolution and inverse one-dimensional convolution, we do not add noise and simulated room response signals in the initial stage of training, and try to pre-train the entire network and microphone reference signal to restore itself. The features of the reference signal are obtained by the alignment model to obtain the aligned reference features, which are compressed to the specified dimension by feature concatenation and affine transformation. The merged features are input into the Conformer model based on the attention mechanism to further abstract the features. The filter coefficients consistent with the input feature dimension are obtained by affine transformation and Sigmoid nonlinear function. The output features are obtained by multiplying the microphone signal with the filter coefficients. Finally, the final output signal is obtained by inverse one-dimensional convolution. Figure 7 shown.

[0070] In addition, unlike the learnable one-dimensional convolution front end, the method based on the Fourier transform front end retains the real number calculation inverse method of the alignment model and the filter estimation model. Our input feature is the amplitude spectrum with a compression factor of 0.3, and the filter coefficient output is a complex number estimate of the same dimension as the Fourier transform. The frequency domain coefficients after filtering are obtained by complex point multiplication, and finally the final output signal is obtained by inverse Fourier transform. Figure 8 shown.

[0071] The self-attention mechanism model mentioned above is composed of multiple Conformer models based on the self-attention mechanism. We used a simplified version of the Conformer unit including the convolution module, the multi-query self-attention module, and the feedforward network. Fig. 9 shown.

[0072] Among them, the convolution module is used to abstract the information within the local window length. Another important module is the self-attention module, which is a brain signal processing mechanism that simulates human vision to automatically learn and calculate the contribution of input data to output data. In the field of speech recognition applications, the attention mechanism considers the contribution of the acoustic features of all time frame sequences of a single sentence to the output features of a certain frame. Real-time speech recognition systems usually use a chunk-wise online attention mechanism, that is, local dependency information is modeled by limiting the scope of the self-attention mechanism. In the field of speech recognition, the Conformer model has more efficient parallel computing capabilities and better recognition performance than recursive neural networks. We introduced such a network structure into speech enhancement and achieved good results.

[0073] Since the data simulation method can only cover limited scenarios, and because the computing power of the embedded platform is limited, only smaller streaming models can be used for inference, the noise reduction effect in actual scenarios will be greatly reduced. Specifically, in real scenarios, there will be differences in echo delay, speaker properties, channels during calls, and accent and speaker coverage issues. In addition, even if there is audio from real scenes for training, the model cannot be improved according to the regular training process because there is no target voice.

[0074] In the face of the above problems, we proposed a knowledge distillation method that combines multiple auxiliary information to transfer the capabilities of the large model to the small model, while allowing the small model to cover the data of real scenes. Specifically, during knowledge distillation, we first use a strict data preprocessing method to obtain data with sufficient data quality, and train a large model with strong generalization ability, which we call the teacher model. Specifically, our teacher model uses a larger-scale denoising model based on learnable features and offline self-attention. Then, the teacher model is used to denoise the real data to obtain the output audio, and the output audio is used as the target audio of the small model. Here, the small model is called the student model. In this way, distillation training can directly allow the student model to learn the generalization ability of the teacher model while covering more real scenes. The student model not only learns the final output audio of the teacher model, but also learns the frame-level voice probability distribution estimated by the teacher model and the relative delay probability distribution of the echo reference signal in the alignment module. In actual audio, there is a problem of delay jitter of the echo signal relative to the reference signal. In addition, due to the large difference between the echo and the reference signal and the complex environment, the accuracy of the delay calculated by the traditional signal estimation method will be greatly reduced. Here we can still use the delay estimation model of the large model as the soft label. Specifically, we use the correlation matrix of the large model as the target result and use the mean squared error (MSE) loss function to make the delay estimation result of the small model close to that of the large model. Similarly, we also let the small model learn the frame-level voice probability distribution estimated by the teacher model. Our knowledge distillation framework is as follows Fig.10 shown.

[0075] In order to verify the final effect, we selected some blind test data sets from AEC Challenge–ICASSP 2021. The average subjective value MOS (the higher the score, the better the subjective listening experience) evaluation used the latest test interface of the AEC Challenge organizer. The benchmark comparison model we selected was the result of Percepnet. ST NE means near-end single talk, ST FE means far-end single talk, DT ECHO means dual talk echo, and DT OTHER means dual talk other.

[0076] The 7M model was adopted. After introducing the explicit optimization method of the alignment module and the deep one-dimensional convolutional model, the MOS evaluation was improved, as shown in Table 1.

[0077] Table 1 Test results of the explicit optimization method of the alignment module and the deep one-dimensional convolutional model

[0078] - Depthwise1d Depthwise1d+AlignLoss STNE 4.00 4.08 4.10 STFEEcho 3.76 3.64 4.15 DTEcho 4.00 4.20 4.30 DTOther 3.40 3.70 4.01 average 3.79 3.91 4.14

[0079] After introducing the data preprocessing process we proposed, the MOS evaluation also improved, as shown in Table 2.

[0080] Table 2 Test results of the model introducing data preprocessing process

[0081]

[0082] After combining various previous optimization methods, we also added a variety of auxiliary information distillation learning methods, which further improved the denoising effect. Our model outperforms the baseline model in all indicators, as shown in Table 3.

[0083] Table 37M model test results

[0084]

[0085]

[0086] In addition, we introduced the above optimization method to a smaller model and selected a 400KB model. From Table 4, we can see that the overall MOS score is better than the 8M PercepeNet benchmark model.

[0087] Table 4400KB model test results

[0088] frame Ours PercepNet STNE 4.23 3.85 STFEEcho 4.5 4.19 DTEcho 4.52 4.34 DTOther 4.04 4.07 average 4.33 4.11

[0089] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An echo cancellation algorithm, characterized in that: The steps include: S1. In the data preprocessing stage, the DNSMOS tool is introduced to screen high-quality data, and the data with low MOS scores are preprocessed by using traditional signal processing and neural network models to reduce noise, and high-quality reference signals are screened out again; S2. In the data generation stage, a progressive training method is adopted. Clean audio is used for pre-training in the initial stage, and then data from easy to difficult is gradually generated for denoising model training; S3, adopting an alignment module based on a cross attention mechanism and using deep one-dimensional convolution to extract local features to improve the performance of the alignment module; S4. Generate an echo delay label using an automatic labeling method, and filter out low-quality reference signals and echo signals in combination with the DNSMOS tool; S5. Calculate the relative delay probability distribution of the reference signal and the echo signal through an alignment module, and explicitly optimize the alignment module in combination with the delay label; S6. A new teacher-student knowledge distillation algorithm combining multiple auxiliary information is used. The student model learns the audio output, frame-level voice probability and echo delay probability distribution of the teacher model, improving the echo cancellation capability in real environments. S7, during real-time calls, filter coefficients are estimated by combining the self-attention model and the cross-attention alignment model to eliminate echo, noise, and late reverberation; S8. The target loss of the noise reduction algorithm uses the SI-SNR loss function and the PLCPA-ASYM loss function. The PLCPA-ASYM loss function performs a second Fourier transform on the audio data after signal restoration, and then calculates the loss value to optimize the output signal quality of the noise reduction model. S9. Introduce the voice state detection function, use the short-time energy threshold and neural network model to generate the voice state label, and improve the listening quality of the noise reduction model.

2. According to the echo cancellation algorithm of claim 1, the alignment module is characterized by using deep one-dimensional convolution to abstract the feature information of the historical frame to the current feature frame, and the feature frames all contain local information; the calculation formula of the deep one-dimensional convolution is: in, y t represents the output feature, x t-i represents the output feature, w i represents the convolution kernel weight, k represents the convolution kernel size;.

3. An echo cancellation algorithm according to claim 1, characterized in that: The S4 specifically includes: S41. Use the DNSMOS tool to filter out reference signals and echo signal data with extremely poor quality; S42, using the traditional signal processing method GCC-PHAT to generate a delay label of the echo signal relative to the reference signal, the specific formula is as follows: in, represents the Fourier transform, Represents X i The complex conjugate of (f), represents the inverse Fourier transform, x1(t) represents the microphone time domain signal, and x2(t) represents the reference time domain signal.

4. The echo cancellation algorithm according to claim 1, characterized in that: The input of the relative delay probability distribution is the local feature of S3, and its calculation formula is: Among them, corr i,j represents the probability that the echo signal of the i-th frame is delayed by j frames relative to the reference signal, q i,m,1 represents the characteristic value of the echo signal at the i-th frame m, k i,m,j represents the m-bit eigenvalue of the reference signal in the ij frame, p represents the dimension of the eigenvector, and D represents the preset maximum number of delayed frames.

5. The echo cancellation algorithm according to claim 1, characterized in that: The teacher model uses a large-scale denoising model based on learnable features, alignment modules and global self-attention, and the student model uses a small-scale model based on traditional spectral features, alignment modules and block streaming self-attention to learn the final output audio and estimated frame-level voice probability distribution of the teacher model, as well as the relative delay probability distribution of the echo reference signal in the alignment module, so as to be close to the estimation result of the teacher model.

Citation Information

Cited By

  • Dynamic spectrum access method based on attention mechanism and echo state network

    CN120263321A

  • Voice processing method and electronic device

    CN122695983A

  • Voice processing methods and electronic devices

    CN122695983B