Acoustic echo cancellation method and processing terminal

Abstract

The invention discloses an acoustic echo cancellation method and a processing terminal, and the method comprises the following steps: 1, carrying out the time alignment of an input far-end signal and a near-end signal, and obtaining a near-end input signal; 2, framing the near-end input signal according to the time domain signal, and performing Fourier transform on each frame to obtain the near-end input signal of the frequency domain corresponding to each frame; 3, superposing the near-end input signal and the far-end signal in the frequency domain on the frequency dimension to obtain a combined signal, extracting a real part of the combined signal, filtering, and taking a logarithm to obtain a filtered logarithm filtering signal; and step 4, inputting the logarithmic filtering signal and the combined signal into a trained deep neural network, outputting a preliminary audio signal in a frequency domain by the deep neural network, and performing inverse Fourier transform on the preliminary audio signal to obtain a final audio signal in a time domain. According to the method, the echo component is effectively suppressed, and the key characteristics of the near-end voice signal are reserved.

Description

Technical Field

[0001] This invention relates to the field of acoustic echo cancellation technology, specifically an acoustic echo cancellation method and processing terminal. Background Technology

[0002] Acoustic echoes can exist in voice communication systems, significantly reducing the clarity of speech signals and thus lowering call quality. For example, in two-way hands-free communication scenarios, acoustic echoes can severely impact the communication experience. Therefore, acoustic echo cancellation is necessary.

[0003] For acoustic echo cancellation, existing methods often employ adaptive filtering algorithms. However, adaptive filtering algorithms have limitations in processing nonlinear echoes, failing to effectively suppress acoustic echo components and often resulting in the loss of key features of the near-end speech signal, meaning that the key features of the near-end speech signal cannot be preserved. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide an acoustic echo cancellation method and processing terminal that can solve the problems described in the background art.

[0005] The technical solution for achieving the objective of this invention is: an acoustic echo cancellation method, comprising the following steps: Step 1: Input the remote signal and near-end signals Time alignment is performed to obtain the near-end input signal. To eliminate the time delay between far-end and near-end signals; Step 2: Input signal at the near end The signal is divided into frames based on the time domain signal, and a Fourier transform is performed on each frame to obtain the near-end input signal in the frequency domain corresponding to each frame. k represents time in the frequency domain, and f represents frequency. To form a time-frequency point, Near-end input signal The expression is as follows:

[0006] In the formula, Near-end signal In the frequency domain expression, For remote signals In the frequency domain expression, for In the frequency domain expression, This represents the room impulse response along the acoustic transmission path from the far end to the near end. For noise Expression in the frequency domain; Step 3: Convert the near-end input signal in the frequency domain and remote signals The two signals are superimposed along the frequency dimension to stitch them together, thus obtaining a combined signal. , Then, extract the combined signal. The real part of the extracted signal is obtained by filtering the extracted real part and then taking its logarithm to obtain the filtered logarithmic signal. ; Step 4: Logarithmically filter the signal Combined signals The input is fed into a trained deep neural network, which outputs a preliminary audio signal in the frequency domain. The initial audio signal Perform an inverse Fourier transform to obtain the final audio signal in the time domain. , The trained deep neural network optimizes by minimizing the loss function, which includes a spectral loss function, an echo perception loss function, and a speech activity detection loss function. The value of the loss function is the sum of the spectral loss function, the echo perception loss function, and the speech activity detection loss function.

[0007] Furthermore, in step 3, the extracted real part is filtered using a Mel filter to convert linear frequencies into nonlinear frequencies, thereby improving low-frequency resolution and reducing high-frequency resolution. noise It is additive noise.

[0008] Furthermore, the deep neural network includes an encoder, a decoder, and a deep filtering network module. The encoder and decoder are connected to each other to logarithmically filter the signal. Input to encoder, combined signal The input is fed into the depth filtering network module, and the encoder's output is also fed into the depth filtering network module.

[0009] Furthermore, the spectral loss function Including amplitude loss and phase loss Spectral loss function Calculated using the following formula:

[0010] In the formula, This represents the initial audio signal output by the deep neural network. At time and frequency points The range at that point, express At time and frequency points The range at that point, This indicates the calculation of mean square error. Indicates taking the amplitude of the complex spectrum Power of 1 It is a constant, and, , This represents the phase information of the complex spectrum of the deep neural network's predicted output. The imaginary unit, This indicates that the complex spectrum of the predicted output of the deep neural network is at the time and frequency points. The true phase angle at that point, Phase information representing the complex spectrum of the near-end signal. The complex spectrum of the near-end signal at time and frequency points The true phase angle at that point, This represents the phase angle of the signal.

[0011] Furthermore, the echo sensing loss function Calculated using the following formula:

[0012] In the formula, Echo weighting coefficients At time and frequency points The calculation formula for the location is as follows:

[0013] In the formula, Indicates remote signal power, Indicates the near-end input signal The power.

[0014] Furthermore, the speech activity detection loss function is calculated using the following formula. :

[0015] In the formula, K represents The total number of frames divided according to the time-domain signal. , This represents the probability predicted by the deep neural network for speech activity detection in the k-th frame. This represents the true VAD tag of the k-th frame of the near-end signal. The true VAD tag value is 0 or 1. When the true VAD tag value is 1, it represents speech. When the true VAD tag value is 0, it represents non-speech.

[0016] Furthermore, the optimization objective of a deep neural network is to minimize the sum L of these three loss functions, which is calculated using the following formula:

[0017] In the formula, a and b represent coefficients. , .

[0018] Furthermore, during the training process of the deep neural network, in each training iteration, the loss functions are summed to form a total loss function, and the echo weighting coefficients of the deep neural network are updated through the backpropagation algorithm. The goal is to minimize the optimization objective, that is, to minimize the sum of the loss functions, thereby obtaining the performance echo weighting coefficients. At this point, the corresponding deep neural network is a pre-trained deep neural network.

[0019] Furthermore, a deep neural network is trained, and the final audio signal is obtained through the trained deep neural network. Its specific implementation includes the following steps: Step 5: Apply a depth filtering algorithm to the near-end input signal. Enhancement is performed to obtain the spectrum after depth filtering enhancement. , Spectrum The calculation formula is as follows:

[0020] In the formula, These represent the complex coefficients of the i-th order filter in a deep neural network. , This represents the total order of the filters in the deep neural network. Indicates a point in time and the input spectrum at frequency f, For a constant, when At that time, it is permissible for non-causal taps to be included in linear combinations. Output signal Enhancement is performed to obtain the enhanced spectrogram. , Logarithmically filtered signal The output signal obtained after being processed by the encoder and decoder in sequence. Compare the enhanced spectrogram after deep filtering with the enhanced spectrogram. Fuse according to the following formula:

[0021] In the formula, This represents the final spectrum after echo cancellation, obtained through fusion. The result Performing an inverse Fourier transform yields the final audio signal in the time domain after echo cancellation. .

[0022] A processing terminal, comprising: Memory, used to store program instructions; A processor for running the program instructions to perform the steps of the acoustic echo cancellation method.

[0023] The beneficial effects of the present invention are as follows: 1. In remote conferencing systems, where loud speaker volume and external volume are common in voice communication scenarios, low signal-to-return ratios can lead to incomplete echo cancellation or near-end voice removal. To address this, the existing DeepFilterNe noise reduction algorithm framework is modified for echo cancellation. The algorithm employs a two-stage network structure: first, ERB scaling gain is used to enhance the spectral envelope, simulating human frequency perception; then, deep filtering is used to enhance the periodic components of the speech. 2. Use the gcc-phat algorithm to perform delay alignment between echo and near-end voice; 3. Use audio vad tags for training to distinguish between speech frames and non-speech frames, reducing oversuppression; 4. Use echo sensing loss to sense the magnitude of the echo and dynamically adjust the amount of echo suppression to alleviate the problem of over-suppression.

[0024] 5. The algorithm has the effect of reverberation suppression. If noise is added to the training data, echo cancellation, reverberation suppression, and noise suppression can be achieved.

[0025] After processing by this invention, the echo component is effectively suppressed while the key features of the near-end speech signal are fully preserved. Attached Figure Description

[0026] Figure 1 This is a schematic diagram illustrating voice communication in both near-end and far-end rooms, including the application of dual audio channels. Figure 2 This is a schematic diagram of the principle framework of the present invention; Figure 3 This is a schematic diagram illustrating the principle of deep neural network training and testing according to the present invention; Figure 4 A flowchart illustrating a preferred embodiment of the acoustic echo cancellation method; Figure 5 This is a schematic diagram of the processing terminal. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and specific embodiments: like Figures 1-4 As shown, an acoustic echo cancellation method includes the following steps: Step 1: Time-align the input far-end and near-end signals to eliminate the time delay between them.

[0028] In this step, the time delay between the far-end signal and the near-end signal can be calculated using GCC-PHAT (Generalized Cross-Correlation Phase Transform). Then, the calculated time delay is compensated to the far-end signal to make the far-end signal and the near-end signal time-aligned, that is, to make the far-end signal time-aligned with the echo component in the near-end input signal.

[0029] refer to Figure 1 In a voice communication system, in each audio channel, the near-end input signal The expression in the time domain is as follows:

[0030] In the formula, n represents discrete time. This refers to the near-end signal, that is, the desired clean near-end speech signal. Indicates a remote signal. The room impulse response (RIR) represents the acoustic transmission path from the far end to the near end, describing the acoustic path from the sound playback end (such as a speaker) to the sound acquisition end (such as a microphone). The length is represented by L. This represents a linear convolution operation, used to simulate the process of an echo being produced after a distant signal travels through an acoustic transmission path. This represents noise; in this example, the noise used is additive noise.

[0031] For voice communication systems, dual channels are generally used. Figure 1 The "near-end room" and "far-end room" illustrated can be two rooms in the real world or virtual rooms on the internet to simulate voice communication between two rooms. They can also be two locations, with an echo between them. Examples include online telephone conferences and video conferences, which involve the transmission of human voices. For dual-channel or multi-channel audio, the audio signals from each channel input to the microphone (i.e., the sound acquisition module) are acquired and processed in parallel. This parallel approach means that the audio signals from two channels are acquired and processed simultaneously.

[0032] The microphone in the near-end room includes two audio channels. Taking the first audio channel as an example, its acquisition end is the near-end signal in the near-end room (i.e., the near-end human voice in the figure). Noise was also collected at the same time. And the remote signals collected from the remote end via the voice communication system. The signal after the far-end human voice (i.e., the sound from the far end in the diagram) passes through the room impulse response (i.e., the echo path in the diagram), and this signal is the near-end input signal of the first audio channel (i.e., s1(n) in the diagram). The near-end input signal of the second audio channel (i.e., s2(n) in the diagram) is similar, so it will not be described in detail here.

[0033] The room impulse response is the change in the echo produced by the acoustic transmission path of the sound output from the speakers in the near room, which is reflected within the near room and finally input to the microphone.

[0034] It is understandable that a speaker and a microphone can be the same device. For example, a microphone has both sound acquisition and sound playback functions, which means it can be two functional modules of the same device (one is a sound acquisition module and the other is a sound playback module). Alternatively, they can be functional modules of two different devices, with one device having a sound acquisition module and the other having a sound playback module.

[0035] Step 2: Time-align the far-end signal and the near-end signal to form the near-end input signal. The signal is divided into frames based on the time domain signal, and a Fourier transform is performed on each frame to obtain the near-end input signal in the frequency domain corresponding to each frame. k represents time in the frequency domain, and f represents frequency. A time-frequency point.

[0036] Understandably, framing based on the time-domain signal can be achieved using window functions, typically short and overlapping ones. Then, performing a Fast Fourier Transform (FFT) on the signal segment (i.e., frame) within each window function yields the near-end input signal in the frequency domain. .

[0037] Near-end input signal The expression is as follows:

[0038] In the formula, That is The expression in the frequency domain, that is, the expression after STFT transformation, for The expression after STFT (Short Time Fourier Transform) transformation, for The expression after STFT transformation, for The expression after STFT transformation.

[0039] Step 3: Convert the near-end input signal in the frequency domain and remote signals The two signals are superimposed along the frequency dimension to stitch them together, thus obtaining a combined signal. . Figure 2 The term "Concat" in this context refers to the concatenation operation.

[0040] This splicing operation provides subsequent deep neural networks with the ability to simultaneously process mixed microphone inputs and far-end signals, which is crucial for the echoes generated by the acoustic transmission path of the deep neural network trajectory.

[0041] Then, extract the combined signal. The real part of the signal is extracted, and the extracted real part is filtered and its logarithm is taken to obtain the filtered logarithmic filtered signal. Logarithmic filtered signal That is Figure 2 In Combined signals That is Figure 2 In .

[0042] For example, this embodiment uses a Mel filter for filtering. The filtered signal obtained after filtering with the Mel filter is... It is a logarithmic spectral feature. The Mel filter is a nonlinear frequency scale that simulates human auditory perception. It converts linear frequencies into Mel frequencies, resulting in higher low-frequency resolution and lower high-frequency resolution, which is more in line with the human ear's perception of speech.

[0043] Step 4: Logarithmically filter the signal Combined signals The input is fed into a trained deep neural network (DNN), which outputs a preliminary audio signal in the frequency domain. The initial audio signal Perform an inverse Fourier transform to obtain the final audio signal in the time domain. .

[0044] Understandably, the inverse Fourier transform can be achieved using the short-time inverse Fourier transform (ISTFT).

[0045] The deep neural network includes an encoder, a decoder, and a deep filtering network module (DF-Net), with the encoder and decoder connected together. Logarithmically filtered signals. Input to encoder, combined signal The encoder output is also input to the depth filtering network module, meaning the encoder output is also connected to the depth filtering network module.

[0046] Digital filtering signal After being processed by the encoder and decoder in sequence, the output signal is obtained. That is Figure 2 In .

[0047] The deep neural network is trained to obtain a trained deep neural network. Training the deep neural network includes determining the loss function, which includes the spectral loss function, the echo perception loss function, and the speech activity detection loss function (also known as VAD loss). By combining multiple loss functions, the deep neural network can be trained. This multi-objective strategy ensures a balance between echo cancellation effect and near-end signal quality.

[0048] The spectral loss function primarily ensures audio fidelity, while the echo perception loss function dynamically adjusts the amount of echo suppression to prevent over-suppression. The speech activity detection loss function primarily distinguishes between speech and non-speech to reduce artifacts during silence.

[0049] Spectral loss function Including amplitude loss and phase loss The amplitude loss is obtained by calculating the mean square error of the amplitude between the output and the input. The initial audio signal is calculated by the output of the deep neural network. Complex spectrum and The mean square error between the complex spectra. And for each time frequency point. The calculation is performed on the above, and then the mean square error at all time points is summed, which is to calculate the amplitude loss. and phase loss The mean square error of each time point needs to be summed up, and the result of the summation is the mean square error of each time point.

[0050] Specifically, the spectral loss function Amplitude loss and phase loss Calculated using the following formula:

[0051] In the formula, This represents the initial audio signal output by the deep neural network. At time and frequency points The range at that point, express At time and frequency points The range at that point, This represents the mean square error calculation, which is the sum of the squares of the differences between two magnitudes. Indicates taking the amplitude of the complex spectrum Power of 1 It is a constant, and, This embodiment ,set up The purpose is to perceive the loudness of audio. This represents the phase information of the complex spectrum of the deep neural network's predicted output. The imaginary unit, This indicates that the complex spectrum of the predicted output of the deep neural network is at the time and frequency points. The true phase angle at that point, Phase information representing the complex spectrum of the near-end signal. The complex spectrum of the near-end signal at time and frequency points The true phase angle at that point, This represents the phase angle of the signal.

[0052] A strong echo, meaning a high echo intensity (amplitude), such as exceeding a preset amplitude suppression threshold, is considered a strong echo. Otherwise, it can be considered a weak echo, indicating that the near-end signal dominates. When the echo is strong, we want the deep neural network to actively eliminate it; however, when the echo is weak, we want the deep neural network to reduce suppression to avoid damaging the near-end signal. Therefore, an echo weighting coefficient needs to be set. In order to hope to use echo weighting coefficients The degree of echo suppression by the deep neural network can be adjusted so that the degree of suppression can be adaptively adjusted according to the intensity of the echo.

[0053] Specifically, echo weighting coefficients Based on each time frequency point The ratio of echo power to total signal power is used to determine the signal strength. Total power includes echo power and microphone signal power; the total microphone signal power is... power and The sum of the powers of the two sources, the echo power is also the sum of the powers of the two sources. The power, therefore, at the time frequency point Echo weighting coefficients at the location The calculation formula is as follows:

[0054] In the formula, This represents the echo power, which is the power of the signal at the far end. This indicates the microphone signal power, which is the power of the near-end input signal.

[0055] Based on echo weighting coefficients The echo sensing loss function is calculated using the following formula. :

[0056] The speech activity detection loss function is calculated using the following formula. :

[0057] In the formula, K represents The total number of frames divided according to the time-domain signal. , This represents the probability predicted by the deep neural network for speech activity detection in the k-th frame. The true VAD tag represents the k-th frame of the near-end signal. The true VAD tag value is 0 or 1. When the true VAD tag value is 1, it represents speech. When the true VAD tag value is 0, it represents non-speech. The value of the true VAD tag is determined by the near-end signal. For the true VAD tag of the near-end signal of the k-th frame, if its energy (i.e. amplitude) exceeds the preset threshold, the true VAD tag value of the near-end signal of the k-th frame is 1. Otherwise, its value is 0. Figure 3 In this context, "Vad label" refers to the VAD label. Figure 3 In this context, Vad predict means... .

[0058] After calculating the three types of loss functions mentioned above, the optimization objective of the deep neural network is to minimize the sum L of these three loss functions. Simultaneously, to ensure a balance between signal fidelity, echo suppression, and speech preservation, appropriate coefficients need to be added. Specifically, the sum L of the loss functions is calculated using the following formula:

[0059] In the formula, a and b represent coefficients. , In this embodiment, a=0.1, b=0.1.

[0060] During the training process of a deep neural network, in each training iteration, the output of the deep neural network is compared with the optimization objective. Based on the comparison result, the three types of loss functions mentioned above can be calculated. These three types of loss functions are then added together to form the sum of the loss functions. The echo weighting coefficients of the deep neural network are then updated using a backpropagation algorithm (such as using the Adam or SGD optimizer). The goal is to minimize the optimization objective, that is, to minimize the sum of the loss functions, thereby obtaining the echo weighting coefficients of the deep neural network with the best performance. At this point, the corresponding deep neural network is a pre-trained deep neural network.

[0061] For example, a deep neural network is trained, and the final audio signal is obtained through the trained deep neural network. Its specific implementation includes the following steps.

[0062] Step 5: Perform the enhancement spectrum network and the enhancement of the periodic components of the speech through deep filtering in sequence to enhance the speech.

[0063] Spectrum enhancement networks can use the scaling gain of the equivalent rectangular bandwidth (ERB) to enhance the spectral envelope, thus simulating human frequency perception. This is equivalent to enhancing the output signal in step 4. Enhancement is performed to obtain the enhanced spectrogram. . Figure 2 The term "Apply Gains" in this context refers to enhanced spectrum networks.

[0064] By applying depth filtering, the focus is placed on enhancing the periodic components of the speech; that is, a depth filtering algorithm is used on the near-end input signal. Enhancements are made to strengthen the periodic components of the speech.

[0065] Spectrum enhanced by deep filtering The definition in the TF domain (time-frequency domain) is as follows:

[0066] In the formula, This represents the spectrum after deep filtering enhancement. These represent the complex coefficients of the i-th order filter in a deep neural network. , This refers to the total order filter (i.e., the maximum order filter) of the deep neural network, which is also the order of the filter in the deep neural network. Indicates a point in time The input spectrum at frequency f. This is an optional parameter that takes the value of a constant. At that time, it is permissible for non-causal taps to be included in linear combinations. This determines whether the depth filtering algorithm can use future audio information to help with current echo cancellation.

[0067] To further ensure that deep filtering primarily affects the periodic components of the signal, existing deep filtering methods are improved by introducing a learnable weighting factor. Weighting factor This weighting factor can be predicted using a deep neural network. This is used to fuse the spectrum map enhanced by deep filtering with the spectrum map obtained by the enhanced spectrum network, thereby obtaining the final spectrum map after echo cancellation, and the fusion is performed according to the following formula:

[0068] In the formula, This represents the final spectrum after echo cancellation, obtained through fusion. This represents the spectrum obtained from the enhanced spectrum network.

[0069] Among them, combined signals The output result obtained after inputting into the deep filtering network module, including its output structure and complex coefficients. (Right now Figure 2 of Weighting factors Together with the input into the depth filtering module (i.e. Figure 2 (Deep Filter in the middle), and then get .

[0070] Step 6: After processing in the previous step, the obtained... Performing an inverse Fourier transform yields the echo-cancelled audio signal in the time domain. .

[0071] Understandably, the inverse Fourier transform can be performed using the inverse short-time Fourier transform (ISTFT), that is, using the inverse short-time Fourier transform (ISTFT) to... By performing the operation, the desired audio signal after echo cancellation can be obtained. .

[0072] like Figure 5 As shown, the present invention also provides a processing terminal 100, which includes: Memory 101 is used to store program instructions; Processor 102 is configured to run the program instructions to perform the steps of the acoustic echo cancellation method.

[0073] The embodiments disclosed in this specification are merely illustrative of one aspect of the invention, and the scope of protection of the invention is not limited to these embodiments. Any other functionally equivalent embodiments fall within the scope of protection of the invention. Those skilled in the art can make various other corresponding changes and modifications based on the technical solutions and concepts described above, and all such changes and modifications should fall within the scope of protection of the claims of this invention.

Claims

1. An acoustic echo cancellation method, characterized in that, Includes the following steps: Step 1: Input the remote signal and near-end signals Time alignment is performed to obtain the near-end input signal. To eliminate the time delay between far-end and near-end signals; Step 2: Input signal at the near end The signal is divided into frames based on the time domain signal, and a Fourier transform is performed on each frame to obtain the near-end input signal in the frequency domain corresponding to each frame. k represents time in the frequency domain, and f represents frequency. To form a time-frequency point, Near-end input signal The expression is as follows: In the formula, Near-end signal In the frequency domain expression, For remote signals In the frequency domain expression, for In the frequency domain expression, This represents the room impulse response along the acoustic transmission path from the far end to the near end. For noise Expression in the frequency domain; Step 3: Convert the near-end input signal in the frequency domain and remote signals The two signals are superimposed along the frequency dimension to stitch them together, thus obtaining a combined signal. , Then, extract the combined signal. The real part of the extracted signal is obtained by filtering the extracted real part and then taking its logarithm to obtain the filtered logarithmic signal. ; Step 4: Filter the logarithmic signal Combined signals The input is fed into a trained deep neural network, which outputs a preliminary audio signal in the frequency domain. The initial audio signal Perform an inverse Fourier transform to obtain the final audio signal in the time domain. , The trained deep neural network optimizes by minimizing the loss function, which includes a spectral loss function, an echo perception loss function, and a speech activity detection loss function. The value of the loss function is the sum of the spectral loss function, the echo perception loss function, and the speech activity detection loss function.

2. The acoustic echo cancellation method according to claim 1, characterized in that, In step 3, the extracted real part is filtered using a Mel filter to convert linear frequencies into nonlinear frequencies, thereby improving low-frequency resolution and reducing high-frequency resolution. noise It is additive noise.

3. The acoustic echo cancellation method according to claim 1, characterized in that, Deep neural networks consist of an encoder, a decoder, and a deep filtering network module. The encoder and decoder are connected to each other and logarithmically filter the signal. Input to encoder, combined signal The input is fed into the depth filtering network module, and the encoder's output is also fed into the depth filtering network module.

4. The acoustic echo cancellation method according to claim 3, characterized in that, Spectral loss function Including amplitude loss and phase loss Spectral loss function Calculated using the following formula: In the formula, This represents the initial audio signal output by the deep neural network. At time and frequency points The range at that point, express At time and frequency points The range at that point, This indicates the calculation of mean square error. Indicates taking the amplitude of the complex spectrum Power of 1 It is a constant, and, , This represents the phase information of the complex spectrum of the deep neural network's predicted output. The imaginary unit, This indicates that the complex spectrum of the predicted output of the deep neural network is at the time and frequency points. The true phase angle at that point, Phase information representing the complex spectrum of the near-end signal. The complex spectrum of the near-end signal at time and frequency points The true phase angle at that point, This represents the phase angle of the signal.

5. The acoustic echo cancellation method according to claim 4, characterized in that, The echo sensing loss function Calculated using the following formula: In the formula, Echo weighting coefficients At time and frequency points The calculation formula for the location is as follows: In the formula, Indicates remote signal power, Indicates the near-end input signal The power.

6. The acoustic echo cancellation method according to claim 5, characterized in that, The speech activity detection loss function is calculated using the following formula. : In the formula, K represents The total number of frames divided according to the time-domain signal. , This represents the probability predicted by the deep neural network for speech activity detection in the k-th frame. This represents the true VAD tag of the k-th frame of the near-end signal. The true VAD tag value is 0 or 1. When the true VAD tag value is 1, it represents speech. When the true VAD tag value is 0, it represents non-speech.

7. The acoustic echo cancellation method according to claim 6, characterized in that, The optimization objective of a deep neural network is to minimize the sum L of these three loss functions. The sum L of the loss functions is calculated using the following formula: In the formula, a and b represent coefficients. , .

8. The acoustic echo cancellation method according to claim 7, characterized in that, During the training process of a deep neural network, in each training iteration, the loss functions are summed to form a total loss function, and the echo weighting coefficients of the deep neural network are updated through the backpropagation algorithm. The goal is to minimize the optimization objective, that is, to minimize the sum of the loss functions, thereby obtaining the performance echo weighting coefficients. At this point, the corresponding deep neural network is a pre-trained deep neural network.

9. The acoustic echo cancellation method according to claim 8, characterized in that, Train a deep neural network, and obtain the final audio signal through the trained deep neural network. Its specific implementation includes the following steps: Step 5: Apply a depth filtering algorithm to the near-end input signal. Enhancement is performed to obtain the spectrum after depth filtering enhancement. , Spectrum The calculation formula is as follows: In the formula, These represent the complex coefficients of the i-th order filter in a deep neural network. , This represents the total order of the filters in the deep neural network. Indicates a point in time and the input spectrum at frequency f, For a constant, when At that time, it is permissible for non-causal taps to be included in linear combinations. Output signal Enhancement is performed to obtain the enhanced spectrogram. , Logarithmically filtered signal The output signal obtained after being processed by the encoder and decoder in sequence. Compare the enhanced spectrogram after deep filtering with the enhanced spectrogram. Fuse according to the following formula: In the formula, This represents the final spectrum after echo cancellation, obtained through fusion. The result Performing an inverse Fourier transform yields the final audio signal in the time domain after echo cancellation. .

10. A processing terminal, characterized in that, It includes: Memory, used to store program instructions; A processor for running the program instructions to perform the steps of the acoustic echo cancellation method as described in any one of claims 1-9.

Images