Video conference echo suppression method based on low-delay adaptive learning model

Abstract

The invention relates to a video conference echo suppression method based on a low-delay adaptive learning model. The method comprises the following steps: acquiring acoustic scene noise features, and matching a lightweight LSTM neural network model by using the acoustic scene noise features; the method comprises the following steps: collecting real-time sound signals of a video conference, preprocessing and extracting multi-modal features of the real-time sound signals; performing adaptive filtering by using an NLMS algorithm to estimate an echo path between a near-end microphone signal and a far-end reference signal of the real-time sound signal, performing convolution on the echo path and the far-end reference signal to generate an echo estimation signal, and subtracting the echo estimation signal from the near-end microphone signal to obtain a preliminary residual signal; learning a gain by using the lightweight LSTM neural network model, and further suppressing the residual signal by using the gain to obtain a time domain enhanced signal; comfortable noise is introduced, and incremental learning is adaptively triggered. And efficient linear and nonlinear echo cancellation is supported.

Description

Technical Field

[0001] This invention relates to the field of real-time audio signal processing technology, and in particular to a video conferencing echo suppression method based on a low-latency adaptive learning model. Background Technology

[0002] In video conferencing, echo occurs when distant audio played from the local speaker is picked up by the local microphone and transmitted back to the distant end, causing remote participants to hear their own voice. In existing technologies, adaptive filters are commonly used for echo cancellation. Specifically, the adaptive filter estimates the echo based on the distant audio and then subtracts the echo signal from the mixed audio signal. However, in practice, this can introduce nonlinear distortion into the acoustic echo, resulting in a significant amount of residual echo. Performance degrades significantly in complex acoustic scenarios (such as highly reflective glass conference rooms), and issues such as slow convergence and speech distortion during two-way communication exist. Furthermore, while existing deep learning solutions offer good performance, their high computational complexity makes it difficult to meet the low latency requirements of video conferencing (typically requiring end-to-end latency <20ms). Summary of the Invention

[0003] To address, or at least partially address, the aforementioned technical problems, this invention provides a video conferencing echo suppression method based on a low-latency adaptive learning model.

[0004] In a first aspect, the present invention provides a video conferencing echo suppression method based on a low-latency adaptive learning model, comprising: When the system is started or a sudden environmental change is detected, or when the user actively triggers the process, an acoustic scene detection process will be performed to obtain the acoustic scene noise characteristics. The acoustic scene noise features are used as the current acoustic scene description vector to match the acoustic scene description label of the lightweight LSTM neural network model in order to match the trained lightweight LSTM neural network model for the same acoustic scene from the model library. Acquire real-time audio signals from video conferences, including near-end microphone signals and far-end reference signals, preprocess and extract multimodal features of the real-time audio signals; The NLMS algorithm uses adaptive filtering to estimate the echo path between the near-end microphone signal and the far-end reference signal. The echo path is convolved with the far-end reference signal to generate an echo estimation signal. The echo estimation signal is then subtracted from the near-end microphone signal to obtain a preliminary residual signal. A gain is learned using a lightweight LSTM network, and the gain is used to further suppress the residual signal to obtain a time-domain enhanced signal; Use artificial noise that matches the spectral characteristics of the acoustic scene noise to fill the silent segments of the time-domain enhanced signal; After a set number of frames, the difference between the current echo return loss enhancement and the average echo return loss enhancement over a set period of time is calculated. If the difference is less than a set threshold, incremental learning is triggered.

[0005] Furthermore, upon startup or detection of a sudden environmental change or user-initiated triggering, an acoustic scene detection process will be executed to acquire environmental noise characteristics, including: When the device is started, the acoustic scene changes, or the user actively triggers it, the acoustic scene noise is collected for a set duration at a set sampling frequency. The reverberation time, power spectrum estimation, and spectral flatness of acoustic scene noise are calculated as acoustic scene noise characteristics.

[0006] Furthermore, the reverberation time acquisition process includes: bandpass filtering and direct current removal of the acquired acoustic scene noise signal according to the set frequency band; calculating the square of the filtered and direct current removed signal to obtain the energy envelope; and performing reverse integration or accumulation of the energy envelope from the end to the beginning. Where N is the total number of samples of acoustic scene noise. k This is a sample index for acoustic scene noise. for k The energy envelope at the point, where n is the sample index of the initial acoustic scene noise signal; the integral or summation result of the energy envelope. Take the logarithm to convert to decibels and normalize; find the sample points corresponding to -5dB and -35dB on the normalized curve. and , for [ , Mapping to a time interval, and normalizing the curve within the corresponding time interval. Performing linear regression and fitting the data yields: , where a is the slope of the fit, representing the decay rate, t is the time, and the reverberation time is: 60 / |a|; The power spectrum estimation process includes: The acoustic scene noise signal is divided into frames and windowed; Fourier transform is performed on the windowed acoustic scene noise signal s(n) of each frame to obtain the corresponding acoustic scene noise frequency domain signal; the square of the modulus of the acoustic scene noise frequency domain signal of each frame is calculated to obtain the power spectrum of the acoustic scene noise of each frame; the power spectrum of all acoustic scene noise frames is averaged to obtain a smooth power spectrum estimate. The process of obtaining spectral flatness includes: Given the power spectrum of a frame of acoustic scene noise signal Calculate its flatness SF as follows: ; in, k = 0, 1, ..., N / 2(Taking the positive frequency part), the numerator is the geometric mean of the power spectrum, and the denominator is the arithmetic mean of the power spectrum. In the calculation, logarithmic transformation is used to avoid numerical underflow. Calculate the median of the flatness of the acoustic scene noise signal across all frames, and use the median flatness as the spectral flatness of the entire acoustic scene noise.

[0007] Furthermore, real-time audio signals from video conferences are acquired, including near-end microphone signals and far-end reference signals. Preprocessing and extraction of multimodal features from the real-time audio signals include: According to the set sampling frequency, the near-end microphone signal d(n) and the far-end reference signal x(n) are synchronously acquired; the real-time audio signal of each frame of the video conference is preprocessed and multimodal features are extracted in a pipeline manner. The preprocessing includes: pre-emphasis, framing and windowing to convert the continuous signal into a short-time stationary frame sequence; multimodal features are extracted from the preprocessed real-time audio signal. The multimodal features include: Mel-spectral difference, frame-level energy ratio of the near-end microphone signal and the far-end reference signal, flatness of the current near-end microphone signal frame, normalized cross-correlation function of the near-end microphone signal d(n) and the far-end reference signal x(n), fundamental frequency features and LPC coefficients.

[0008] Furthermore, the process of obtaining the 8-dimensional Mel spectrum difference includes: passing the near-end microphone signal d(n) and the far-end reference signal x(n) through a Mel filter bank with a set number of channels to obtain their respective Mel spectra; calculating the logarithmic domain difference of the Mel spectra of the near-end microphone signal d(n) and the far-end reference signal x(n) in each channel; performing a discrete cosine transform on the logarithmic domain difference of the Mel filter bank in each channel, and retaining the first 8 low-order coefficients; Obtain the frame-level energy ratio of the near-end microphone signal d(n) and the far-end reference signal x(n): ; Calculate the normalized cross-correlation function between the near-end microphone signal d(n) and the far-end reference signal x(n). The cross-correlation function is a measure of the similarity or correlation between two signals at different time offsets. Find the maximum peak value of the cross-correlation function. and their corresponding delays ; Calculate the fundamental frequency characteristics of the near-end microphone signal d(n) and the far-end reference signal x(n), including: The fundamental frequency was estimated from the near-end microphone signal d(n) and the far-end reference signal x(n) using the autocorrelation method. and According to the base frequency and Calculate the fundamental frequency characteristics: The fundamental frequency characteristics include: the first fundamental frequency characteristic and the second fundamental frequency characteristic; if the fundamental frequency... and Both are valid; the first fundamental frequency characteristic is taken as the fundamental frequency ratio: Otherwise, it is 0; the second fundamental frequency feature uses the near-end voiced tone marker.

[0009] Furthermore, filling the silent segment with artificial noise that matches the spectral characteristics of the acoustic scene noise includes: in the frequency domain, generating a random phase, the random phase being uniformly distributed in [0, 2π], and constructing an amplitude using the power spectrum of the acoustic scene noise. , To obtain the power spectrum of the acoustic scene noise signal, frequency domain noise is constructed using amplitude and frequency; inverse Fourier transform is performed on the frequency domain noise to obtain time domain comfort noise; comfort noise is mixed into the silent section of the enhanced time domain signal according to a certain proportion.

[0010] Furthermore, after a set number of frames, the difference between the current echo return loss enhancement and the average echo return loss enhancement over a set period of time is calculated. If the difference is less than a set threshold, incremental learning is triggered, including: after a set number of frames, taking the logarithm of the ratio of the sum of the squares of the near-end microphone signal to the sum of the squares of the residual signal and multiplying it by 10 as the echo return loss enhancement.

[0011] Where d(n) is the near-end microphone signal used to calculate the echo return loss enhancement, and e(n) is the residual signal used to calculate the echo return loss enhancement. The echo return loss enhancement measures the instantaneous performance of the adaptive filter. The higher the echo return loss enhancement, the better the echo cancellation. Calculate the difference ΔERLE between the current echo return loss enhancement and the average echo return loss enhancement over a set time period in the past. If the difference is less than a set threshold, trigger incremental learning of the lightweight LSTM neural network model.

[0012] Furthermore, the incremental learning includes: A circular buffer is set up to continuously store real-time audio data within a set time period, including: far-end reference signal, near-end microphone signal, and adaptively filtered residual signal; multi-dimensional feature vectors of the real-time audio data in the circular buffer are extracted; when learning is triggered, the real-time audio data in the circular buffer is used as the fine-tuning dataset for model fine-tuning: during fine-tuning, the first layer of the lightweight LSTM network is frozen, and only the second layer of the network and the subsequent fully connected layers are unlocked for fine-tuning; scale-invariant signal-to-noise ratio is used as the loss function, and a lightweight optimizer such as momentum-driven SGD or AdamW is used for a set number of iterations. The fine-tuned model is forward-inferred on another small segment of cached data to calculate the echo return loss enhancement of the processed signal. If the echo return loss enhancement of the fine-tuned model exceeds the set optimization threshold compared to the old model, the fine-tuning is considered effective, and the parameters of the old model currently running in memory are atomically replaced with the parameters of the fine-tuned new model; if the validation fails, the fine-tuned model is discarded, the original model is used again, and the failure is recorded, and learning is not triggered again for a short period of time.

[0013] Secondly, the present invention provides a video conferencing echo suppression device based on a low-latency adaptive learning model, comprising: at least one processing unit, wherein the processing unit is connected to a storage unit, a near-end acquisition unit and a far-end acquisition unit via a bus unit, the storage unit stores a computer program, and the processing unit implements the video conferencing echo suppression method based on a low-latency adaptive learning model by running the computer program stored in the storage unit.

[0014] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed, implements the video conferencing echo suppression method based on a low-latency adaptive learning model.

[0015] The technical solutions provided in the embodiments of the present invention have the following advantages compared with the prior art: This invention provides a complete, precise, and adaptive echo cancellation system design. Starting from environmental perception, it achieves efficient linear and nonlinear echo cancellation through the collaborative work of classic adaptive filtering and a lightweight LSTM neural network model, demonstrating strong real-time processing capabilities. An online incremental learning mechanism is introduced, making this application not static but capable of continuous evolution and self-optimization, able to cope with the endless acoustic challenges of the real world, ultimately providing users with a clear, natural, and coherent voice communication experience. It is applicable to video conferencing audio processing in various acoustic scenarios. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of a video conferencing echo suppression method based on a low-latency adaptive learning model provided in an embodiment of the present invention; Figure 2 A flowchart for obtaining the reverberation time of acoustic scene noise provided in an embodiment of the present invention; Figure 3 A flowchart for estimating the background noise spectrum provided in an embodiment of the present invention; Figure 4 A flowchart for obtaining Mel spectrum differences is provided as an embodiment of the present invention; Figure 5 A flowchart for triggering reinforcement learning provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of a video conferencing echo suppression device based on a low-latency adaptive learning model, provided as an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0021] Example 1 See Figure 1As shown, the video conferencing echo suppression method based on a low-latency adaptive learning model provided in this application includes: The goal of this solution is to eliminate acoustic echoes in communication equipment and suppress background noise, thereby ensuring the clarity and naturalness of voice communication. The system employs a dual-path architecture combining classic adaptive filtering with advanced deep learning to handle complex and ever-changing acoustic scenarios. Adaptive filtering handles coarse cancellation of linear echoes; the neural network handles fine suppression of nonlinear residual echoes and noise. Furthermore, this application possesses environmental awareness and online incremental learning capabilities, enabling it to dynamically adapt to new acoustic scenarios and achieve a "gets smarter with use" effect. The video conferencing echo suppression method based on a low-latency adaptive learning model provided in this application possesses a three-in-one echo suppression capability of low latency, high fidelity, and strong adaptability, mainly including: S100, acoustic scene detection and model matching based on the acoustic scene detection results, including: S110: Upon startup or detection of a sudden change in the acoustic scene or user-initiated triggering, an acoustic scene detection operation is performed to extract acoustic scene features, selecting the optimal initial configuration for subsequent real-time processing. Specifically, step S110 includes: S111: When the device is started, the acoustic scene changes, or the user actively triggers the process, the system prompts "Calibration in progress" and collects acoustic scene noise for a set duration according to the set sampling frequency. The total number of acoustic scene noise samples is the sampling frequency multiplied by the set duration.

[0022] S112, after obtaining the acoustic scene noise, calculate the reverberation time, power spectrum estimation, and spectral flatness of the acoustic scene noise as acoustic scene noise features.

[0023] The reverberation time of acoustic scene noise represents the time required for the sound pressure level to decrease by 60 dB. A longer reverberation time implies more severe reverberation, increasing the difficulty of echo cancellation. The inverse integration method is used to estimate the reverberation time from the acoustic scene noise; specifically, such as... Figure 2 As shown, the acquired acoustic scene noise signal zs(n) is bandpass filtered and de-scanned according to a set frequency band. The set frequency band for bandpass filtering is 100Hz-8kHz, which corresponds to the human speech audio range to better cancel echoes in human speech. The de-scanning operation subtracts the average value of the overall acoustic scene noise signal from each acoustic scene noise signal; the square of the filtered and de-scanned signal is calculated to obtain the energy envelope; the energy envelope is then integrated backwards from the end or accumulated in a discrete form. Where N is the total number of samples of acoustic scene noise. k This is a sample index for acoustic scene noise. for kThe energy envelope at the point, where n is the sample index of the initial acoustic scene noise signal; the integral or accumulation result Take the logarithm and convert it to decibels: and normalize: ;exist Find the sample points corresponding to -5dB and -35dB on the curve. and ; for [ , Mapping to time intervals, and for the corresponding time intervals... Linear regression was performed on the curve to obtain the following result: Where a is the slope, representing the decay rate (in dB / s), and t is the time, then the reverberation time is: 60 / |a|.

[0024] like Figure 3 As shown, the estimated background noise spectrum includes: The acoustic scene noise signal is divided into frames and windowed (Hanning window). Specifically, the acoustic scene noise signal is divided into short frames of length L, and a frame shift is set, which is usually half the length of the short frame. A Hanning window is applied to each frame to reduce spectral leakage.

[0025] Perform a Fourier transform on the windowed acoustic scene noise signal zs(n) of each frame to obtain the corresponding acoustic scene noise frequency domain signal D(k).

[0026] The power spectrum of the acoustic scene noise for each frame is obtained by calculating the square of the modulus of the frequency domain signal D(k) of the acoustic scene noise for each frame. .

[0027] The flatness of the acoustic scene noise signal in all frames is calculated using the power spectrum. The median of the flatness is then used as the spectral flatness of the entire acoustic scene noise. The power spectrum of all acoustic scene noise frames is averaged to obtain a flat spectrum estimate; spectral flatness is a measure of the similarity between the signal spectrum and the flat spectrum. Noise spectra are typically flat, while the spectra of harmonic signals such as speech have obvious peaks and valleys, resulting in lower flatness. The process of obtaining spectral flatness includes: Given the power spectrum of a frame of acoustic scene noise signal ,in, k = 0, 1, ..., N / 2 (Taking the positive frequency portion), its flatness SF is: ; The numerator is the geometric mean of the power spectrum, and the denominator is the arithmetic mean of the power spectrum. Logarithmic transformation is used in the calculation to avoid numerical underflow.

[0028] Calculate the median of the flatness of the acoustic scene noise signal across all frames and use it as the spectral flatness of the entire acoustic scene noise.

[0029] S200 uses the acoustic scene noise features as the acoustic scene description vector and matches the acoustic scene description label of the lightweight LSTM neural network model to match the trained lightweight LSTM neural network model for the same acoustic scene from the model library.

[0030] In the specific implementation process, a pre-set model library is maintained, storing pre-trained lightweight LSTM neural network models for different acoustic scenarios, such as glass conference rooms, carpeted conference rooms, open offices, in-vehicle environments, and family living rooms. Each lightweight LSTM neural network model corresponds to an acoustic scene description label, which is the acoustic scene description vector of the acoustic scene for which the lightweight LSTM neural network model is obtained. The Euclidean distance or cosine similarity between the current acoustic scene description vector and each acoustic scene description label in the model library is calculated. The pre-set model with the closest distance or the highest similarity is selected and loaded into memory as the initial neural network model for real-time processing. This provides a high-performance model configuration for subsequent echo processing, avoiding adaptation from scratch.

[0031] The S300 synchronously acquires real-time audio signals from video conferences, including near-end microphone signals and far-end reference signals, preprocesses them, and extracts multimodal features of the real-time audio signals.

[0032] Specifically, step S300 includes the following processes: The near-end microphone signal d(n) and the far-end reference signal x(n) are simultaneously acquired at a set sampling frequency of 16kHz.

[0033] The real-time audio signal of each frame of the video conference is preprocessed and extracted in a pipeline manner to ensure real-time performance.

[0034] The preprocessing includes: pre-emphasis to boost high-frequency components, compensating for high-frequency attenuation caused by lip radiation during speech production, and making the spectrum flatter for easier subsequent analysis. , , Where α is the pre-weighting coefficient, which is usually taken as 0.95 to 0.97, and in one example the pre-weighting coefficient α is taken as 0.97.

[0035] Preprocessing also includes: framing and windowing to convert continuous signals into short-time stationary frame sequences; segmenting the pre-emphasized signal according to the set frame length and frame shift; and using a Hanning window, which can effectively reduce spectral sidelobes caused by frame truncation and improve the accuracy of spectral analysis. The preprocessed data is assigned to d(n) and x(n).

[0036] Multimodal features are extracted from the preprocessed real-time audio signal. These multimodal features are a mixture of characteristics that comprehensively reflect the dual-talk, echo, and noise states. The multimodal features include: Mel-spectral differences, frame-level energy ratio of the near-end microphone signal and the far-end reference signal, flatness of the current near-end microphone signal frame, normalized cross-correlation function of the near-end microphone signal d(n) and the far-end reference signal x(n), fundamental frequency characteristics, and LPC coefficients.

[0037] like Figure 4 As shown, the process of obtaining the Mel spectrum difference between the near-end microphone signal d(n) and the far-end reference signal x(n) includes: passing the near-end microphone signal d(n) and the far-end reference signal x(n) through a 40-channel Mel filter bank to obtain their respective Mel spectra: and Where m is the Mel channel index. Calculate the logarithmic domain difference of the Mel spectra of the near-end microphone signal d(n) and the far-end reference signal x(n) on each channel: ε is a minimum value to prevent division by zero. A discrete cosine transform is performed on the logarithmic domain differences of the 40-channel Mel filter bank, and the first 8 low-order coefficients are retained as Mel spectral differences. These 8 retained low-order coefficients contain the main difference information in the spectral shape of the near and far-end signals and are highly sensitive to dual-talk and residual echo. Specifically, when the near and far ends speak simultaneously, the Mel spectrum of the near-end microphone signal d(n) is... It will be a mixture of the near-end speech spectrum and the echo spectrum. The Mel spectrum of the far-end reference signal x(n) will be... It is a pure far-end speech spectrum. At this time, It will exhibit a very unique shape, neither all zeros (pure echo) nor a random noise shape. This unique shape information is precisely preserved in the low-order coefficients of the discrete cosine transform. By learning these coefficients, the neural network can learn to determine whether the spectral differences of the current frame conform to the dual-talk mode. After the adaptive filter, most of the linear echo is eliminated, but nonlinear residual echoes still exist. At this point... It still contains some related to Related components. This residual correlation will make It exhibits a specific pattern (e.g., small differences in some frequency bands and large differences in others). The lower-order coefficients of the discrete cosine transform can also capture this specific pattern, allowing the neural network to recognize that although the current frame has undergone linear filtering, there is still echo residue, thus triggering stronger suppression. It is obtained by compressing 40-dimensional, highly correlated Mel spectrum difference information into an 8-dimensional, decorrelated feature vector that best represents the macroscopic changes in the spectrum shape. This provides the neural network with crucial and efficient input information, enabling it to accurately perceive double talk and residual echo events.

[0038] Frame-level energy ratio of the near-end microphone signal d(n) and the far-end reference signal x(n): The energy ratio directly measures the degree of attenuation of the echo path. In a pure echo frame, the energy ratio is a stable value; if double-talk (near-end speech) occurs, the energy ratio will decrease sharply; if the ambient noise increases, the energy ratio will also change.

[0039] Calculate the flatness SF_current of the current near-end microphone signal frame. Compare it with the initial acoustic scene noise flatness SF. If the flatness SF_current of the current near-end microphone signal frame is close to the acoustic scene noise SF, then the current near-end microphone signal frame is more likely to be noise or residual echo; if SF_current is much smaller than the acoustic scene noise SF, then the current near-end microphone signal frame is likely to contain near-end speech rich in harmonics.

[0040] Calculate the normalized cross-correlation function between the near-end microphone signal d(n) and the far-end reference signal x(n). The cross-correlation function is a measure of the similarity or correlation between two signals at different time offsets (time delays τ). Find the maximum peak value of the cross-correlation function. and their corresponding delays High peak values ​​indicate a strong linear correlation between the current near-end microphone signal d(n) and the far-end reference signal x(n), meaning it's primarily echo. Low peak values ​​may indicate dual-talk or nonlinear distortion. (Time delay) Used for coarse calibration of system delay. Specifically, in a real-time system, this application utilizes the equivalence between time-domain cross-correlation and frequency-domain conjugate multiplication to calculate the normalized cross-correlation function between the near-end microphone signal d(n) and the far-end reference signal x(n), including: padding the near-end microphone signal d(n) and the far-end reference signal x(n) with zeros to a set length, and performing Fourier transforms to obtain the spectrum D(l) of the near-end microphone signal d(n) and the spectrum X(l) of the far-end reference signal x(n), where l is a frequency domain variable. The cross-power spectrum is then calculated in the frequency domain: Where conj represents the complex conjugate. An inverse Fourier transform is performed on the cross-power spectrum to obtain the cross-correlation function. The cross-correlation function is normalized using the Pearson correlation coefficient formula. The range of the cross-correlation function is [-1, 1], where 1 indicates perfect positive correlation (d(n) is a perfect scaled and delayed version of x(n); 0 indicates no correlation; and -1 indicates perfect negative correlation (d(n) is a delayed and inverted version of x(n)). The maximum peak value... and their corresponding delays After calculating the normalized cross-correlation function sequence, we find the maximum value (peak) and its location on it.

[0041] Calculate the fundamental frequency characteristics of the near-end microphone signal d(n) and the far-end reference signal x(n), including: estimating the fundamental frequency from the near-end microphone signal d(n) and the far-end reference signal x(n) using the autocorrelation method, respectively. and The near-end microphone signal d(n) or the far-end reference signal x(n) has the highest similarity to its own signal after a delay of one or more periods. The autocorrelation function will peak at a delay equal to the signal period T. The fundamental frequency is the ratio of the sampling rate to the period T. Based on the fundamental frequency... and Calculate the fundamental frequency characteristics: The fundamental frequency characteristics include: the first fundamental frequency characteristic and the second fundamental frequency characteristic; if the fundamental frequency... and Both are valid; the first fundamental frequency characteristic is taken as the fundamental frequency ratio: Otherwise, it is 0; when A value close to 1 indicates that the pitch of the speaker in the near-end microphone signal d(n) is very close to the pitch of the speaker in the far-end reference signal x(n), suggesting they may be the same person, i.e., an echo exists. When A deviation from 1.0, such as <0.8 or >1.2, indicates that the proximal and distal ends are two speakers with different pitches, i.e., double speech has occurred. When A value of 0 indicates that at least one side (near-end or far-end) is unvoiced or silent in the current frame. In this case, the first fundamental frequency feature is invalid and provides no information. The second fundamental frequency feature uses the near-end voiced indicator. , It is a binary feature that indicates whether the near-end microphone signal is voiced, with 1 indicating yes and 0 indicating no.

[0042] The process of obtaining LPC coefficients involves performing linear predictive analysis on the near-end microphone signal frames, such as the Levinson-Durbin recursion, to solve for the 4th-order LPC coefficients: a1, a2, a3, a4. The LPC coefficients reflect the formant structure (vocal tract shape) of the speech signal. Different near-end and far-end speakers exhibit different formant structures, providing information for neural networks to distinguish between different speakers. A speech sample can be predicted using a linear combination of its past samples. For a discrete-time near-end microphone signal d(n), its predicted value... It can be represented as: Where p is the order of the linear predictor, These are the LPC coefficients to be calculated. The error e(n) between the predicted and actual values ​​is: The goal of LPC analysis is to find a set of LPC coefficients that minimizes the mean square error of the prediction error e(n) over the entire frame of signal. First, the short-time autocorrelation function R(l) of the near-end microphone signal d(n) from 0 to 4 is calculated, resulting in the autocorrelation sequence [R(0), R(1), R(2), R(3), R(4)]. Based on the autocorrelation sequence, a Levinson-Durbin recursion is performed, starting from the first order and gradually progressing to the fourth order to obtain the fourth-order LPC coefficients. The LPC coefficients define a synthesized filter with all poles. The frequency response of this synthesized filter simulates the frequency response of the human vocal tract. The resonance characteristics of the vocal tract are manifested as several formants in the frequency response of the synthesized filter, which are key to distinguishing different vowels. The LPC coefficients directly determine the pole positions of this synthesized filter, thus determining the position and bandwidth of the formants. Therefore, the LPC coefficients a1, a2, a3, a4 are essentially a highly compressed mathematical description of the formant structure (i.e., the instantaneous shape of the vocal tract) of the current frame of speech signal. Different people (adult men, women, and children) have different formant frequencies when pronouncing the same vowel due to differences in their vocal tract physiological structures (length and shape). Therefore, the LPC coefficients of the far-end speaker and the near-end speaker will exhibit different patterns. During double-talk, the near-end signal d(n) is a mixture of near-end speech and far-end echo. Its LPC coefficients attempt to fit the spectrum of this mixed signal. However, neural networks can learn to identify which patterns in these coefficients are more biased towards the "near-end speaker" and which are more biased towards the "far-end speaker," thus helping to determine whether double-talk exists and who is speaking.

[0043] Fundamental frequency features and LPC coefficients provide information from two different dimensions of speech generation, and they are complementary. Even if two people happen to have the same pitch (F0), their vocal tract shapes will almost certainly be different, resulting in different LPC coefficients. This provides neural networks with more powerful discriminative capabilities. Using only four LPC coefficients, rich spectral envelope information is contained, making it a highly efficient feature representation, well-suited as input to lightweight neural networks such as lightweight LSTMs.

[0044] S400, NLMS algorithm adaptive filtering to estimate the echo path between the near-end microphone signal d(n) and the far-end reference signal x(n). echo path The echo estimation signal y(n) is generated by convolving the far-end reference signal x(n), and then subtracted from the near-end microphone signal d(n) to obtain the preliminary residual signal. .

[0045] The NLMS algorithm iteratively adjusts the filter weights to minimize the mean square error between the filter output and the desired signal. In this application, the received near-end microphone signal d(n) and far-end reference signal x(n) frames are stacked into a block. Once a block is accumulated (e.g., 4 frames), the filter is updated again to further reduce computational complexity. The filter weights are initialized to zero vectors. The NLMS algorithm iteratively executes the following steps: Obtain the frequency domain response of the current filter weights and the frequency domain response of the current near-end microphone signal block. Convolve the frequency domain response of the current filter weights and the frequency domain response of the current near-end microphone signal block, and then perform an inverse Fourier transform to obtain the estimated value of the current echo. Subtract the current echo estimate obtained in step one from the current near-end microphone signal acquired by the microphone to calculate the residual signal; the residual signal is the output signal after echo cancellation, which includes near-end speech, nonlinear echo residue, and background noise. Normalize the power of the current near-end microphone signal to obtain the normalization step size; use the normalization step size, the residual signal, and the current near-end microphone signal to update the filter weights to prepare for processing the next sample, where: f is the iteration step. The filter weights are for the (f+1)th iteration step. Let f be the filter weights for the f-th iteration step. For the normalization compensation in the f-th iteration step, Let f be the residual signal at the f-th iteration step. is the far-end reference signal of the input filter in the f-th iteration step.

[0046] Even after adaptive filtering, nonlinear distortion and residual echoes in the double-talk case still need to be addressed. This application utilizes a lightweight LSTM network to learn a gain for the residual signal. Further suppression is needed.

[0047] In specific implementation, the lightweight LSTM network includes: receiving multi-dimensional feature vectors and residual signals. The input layer is a logarithmic power spectrum, followed by two LSTM layers, each containing 32 or 64 neurons. The lightweight LSTM has memory capabilities, effectively modeling the temporal dynamics of speech and echo. A fully connected layer connects the lightweight LSTM layers, using the sigmoid activation function. The output layer, connected to the fully connected layers, outputs an ideal gain vector. The sigmoid function constrains the output between [0, 1].

[0048] residual signal Transform to the frequency domain and calculate its power spectrum. And take the logarithm. Multidimensional eigenvectors and residual signals The logarithmic power spectra are concatenated to form an extended feature vector, which is then input into a lightweight LSTM network. The network outputs the gain coefficients G of each frame to form a gain vector for multiple frames. The gain is applied to the spectrum of the residual signal, followed by an inverse Fourier transform, and then the OLA algorithm is applied to synthesize the final enhanced time-domain signal.

[0049] The S600 uses artificial noise that matches the spectral characteristics of the acoustic scene noise to fill silent segments, achieving comfortable noise injection. After suppressing echoes and noise, if the signal becomes completely silent, it will provide a very unnatural auditory experience for the user. Comfortable noise injection aims to fill silent segments with artificial noise that matches the spectral characteristics of the acoustic scene noise, maintaining auditory coherence. Noise spectrum estimation: The background noise spectrum is updated using the power spectrum of the acoustic scene noise estimated during the initialization phase and frames identified as silent through voice activity detection in real-time processing. In the frequency domain, random phases are generated, uniformly distributed in [0, 2π]. An amplitude is constructed using the power spectrum of the acoustic scene noise. First, construct frequency domain noise. Then, perform an inverse Fourier transform on the frequency domain noise to obtain time-domain comfort noise. In the silent section of the enhanced time-domain signal, mix in comfort noise at a certain proportion to ensure that the noise floor of the output signal is consistent with the perceived noise of the original acoustic scene.

[0050] S700, an online learning triggering mechanism for lightweight LSTM neural network models. For example... Figure 5 As shown, it includes: After a set number of frames, the logarithm of the ratio of the sum of the squared near-end microphone signal to the sum of the squared residual signal is multiplied by 10 to enhance the echo return loss. Echo return loss enhancement measures the instantaneous performance of an adaptive filter. A higher echo return loss enhancement indicates better echo cancellation.

[0051] Calculate the difference ΔERLE between the current echo return loss enhancement and the average echo return loss enhancement over a set period of time in the past.

[0052] If the difference is less than the set threshold, for example, ΔERLE<-3 dB, meaning the performance drops significantly by more than 3 dB, then it is determined that an unknown change has occurred in the current acoustic scene, a performance degradation has been detected, and an incremental learning process is triggered to fine-tune the neural network model to adapt to the new environment.

[0053] The incremental learning process includes: setting up a circular buffer to continuously store real-time audio data within a set time period, including: a far-end reference signal x(n), a near-end microphone signal d(n), and an adaptively filtered residual signal e(n); extracting multi-dimensional feature vectors from the real-time audio data in the circular buffer; when learning is triggered, the real-time audio data in the circular buffer is used as the fine-tuning dataset for model fine-tuning: during fine-tuning, the first layer of the lightweight LSTM network is frozen, and only the second layer of the network and subsequent fully connected layers are unlocked for fine-tuning. These layers are responsible for higher-level and more specialized decisions and are key to adapting to new environments. Scale-invariant signal-to-noise ratio is used as the loss function, and a lightweight optimizer such as momentum-driven SGD or AdamW is used to iterate the lightweight LSTM network a set number of times (e.g., 10-20 times) to avoid overfitting and destroying existing knowledge. The fine-tuned model is used for forward inference on another small segment of buffered data to calculate the echo return loss enhancement of the processed signal. If the improved echo loss of the fine-tuned model exceeds the set optimization threshold by more than 2dB compared to the old model, the fine-tuning is considered effective, and the parameters of the old model currently running in memory are atomically replaced with the parameters of the fine-tuned new model. This process takes milliseconds and is imperceptible to real-time processing. If the validation fails, the fine-tuned model is discarded, the original model is used again, and the failure is recorded, preventing further learning for a set period.

[0054] This solution provides a complete, precise, and adaptive echo cancellation system design. Starting from environmental perception, it achieves efficient linear and nonlinear echo cancellation through the collaborative work of classic adaptive filtering and a lightweight LSTM neural network model. An online incremental learning mechanism is introduced, making this application not static but capable of continuous evolution and self-optimization, able to cope with the endless acoustic challenges of real-world scenarios, ultimately providing users with a clear, natural, and coherent voice communication experience.

[0055] Example 2 See Figure 6 As shown, this embodiment of the invention provides a video conferencing echo suppression device based on a low-latency adaptive learning model, comprising: at least one processing unit, the processing unit being connected to a storage unit, a near-end acquisition unit, and a far-end acquisition unit via a bus unit, the storage unit serving as a computer-readable storage medium for storing software programs, computer-executable programs, and modules, such as the software program, computer-executable program, and module corresponding to the video conferencing echo suppression method based on a low-latency adaptive learning model in this embodiment of the invention. The processing unit implements the aforementioned video conferencing echo suppression method based on a low-latency adaptive learning model by running the software program, computer-executable program, and module stored in the storage unit, including: When the system is started or a sudden environmental change is detected, or when the user actively triggers the process, an acoustic scene detection process will be performed to obtain the acoustic scene noise characteristics. The acoustic scene noise features are used as the current acoustic scene description vector to match the acoustic scene description label of the lightweight LSTM neural network model in order to match the trained lightweight LSTM neural network model for the same acoustic scene from the model library. Acquire real-time audio signals from video conferences, including near-end microphone signals and far-end reference signals, preprocess and extract multimodal features of the real-time audio signals; The NLMS algorithm uses adaptive filtering to estimate the echo path between the near-end microphone signal and the far-end reference signal. The echo path is convolved with the far-end reference signal to generate an echo estimation signal. The echo estimation signal is then subtracted from the near-end microphone signal to obtain a preliminary residual signal. A gain is learned using a lightweight LSTM neural network model, and the gain is used to further suppress the residual signal to obtain a time-domain enhanced signal; Fill silent segments with artificial noise that matches the noise spectrum characteristics of the acoustic scene. After a set number of frames, the difference between the current echo return loss enhancement and the average echo return loss enhancement over a set period of time is calculated. If the difference is less than a set threshold, incremental learning is triggered.

[0056] Of course, the computer program stored in the storage unit of the video conferencing echo suppression device based on a low-latency adaptive learning model provided in the embodiments of the present invention is not limited to the method operation described above, but can also execute related operations in the video conferencing echo suppression method based on a low-latency adaptive learning model provided in any embodiment of the present invention.

[0057] Example 3 This invention provides a computer-readable storage medium storing a computer program. When executed, the computer program implements a video conferencing echo suppression method based on a low-latency adaptive learning model, comprising: When the system is started or a sudden environmental change is detected, or when the user actively triggers the process, an acoustic scene detection process will be performed to obtain the acoustic scene noise characteristics. The acoustic scene noise features are used as the current acoustic scene description vector to match the acoustic scene description label of the lightweight LSTM neural network model in order to match the trained lightweight LSTM neural network model for the same acoustic scene from the model library. Acquire real-time audio signals from video conferences, including near-end microphone signals and far-end reference signals, preprocess and extract multimodal features of the real-time audio signals; The NLMS algorithm uses adaptive filtering to estimate the echo path between the near-end microphone signal and the far-end reference signal. The echo path is convolved with the far-end reference signal to generate an echo estimation signal. The echo estimation signal is then subtracted from the near-end microphone signal to obtain a preliminary residual signal. A gain is learned using a lightweight LSTM neural network model, and the gain is used to further suppress the residual signal to obtain a time-domain enhanced signal; Fill silent segments with artificial noise that matches the noise spectrum characteristics of the acoustic scene. After a set number of frames, the difference between the current echo return loss enhancement and the average echo return loss enhancement over a set period of time is calculated. If the difference is less than a set threshold, incremental learning is triggered.

[0058] The computer-readable storage medium provided in this embodiment of the invention stores a computer program that is not limited to the method operation described above, but can also execute related operations in the video conferencing echo suppression method based on a low-latency adaptive learning model provided in any embodiment of the invention.

[0059] In the embodiments provided by this invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, structures, or units, and may be electrical, mechanical, or other forms.

[0060] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0061] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0062] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A video conferencing echo suppression method based on a low-latency adaptive learning model, characterized in that, include: When the system is started or a sudden environmental change is detected, or when the user actively triggers the process, an acoustic scene detection process will be performed to obtain the acoustic scene noise characteristics. The acoustic scene noise features are used as the current acoustic scene description vector to match the acoustic scene description label of the lightweight LSTM neural network model in order to match the trained lightweight LSTM neural network model for the same acoustic scene from the model library. Acquire real-time audio signals from video conferences, including near-end microphone signals and far-end reference signals, preprocess and extract multimodal features of the real-time audio signals; The NLMS algorithm uses adaptive filtering to estimate the echo path between the near-end microphone signal and the far-end reference signal. The echo path is convolved with the far-end reference signal to generate an echo estimation signal. The echo estimation signal is then subtracted from the near-end microphone signal to obtain a preliminary residual signal. A gain is learned using a lightweight LSTM neural network model, and the gain is used to further suppress the residual signal to obtain a time-domain enhanced signal; Use artificial noise that matches the spectral characteristics of the acoustic scene noise to fill the silent segments of the time-domain augmented signal; After a set number of frames, the difference between the current echo return loss enhancement and the average echo return loss enhancement over a set period of time is calculated. If the difference is less than a set threshold, incremental learning is triggered.

2. The video conferencing echo suppression method based on a low-latency adaptive learning model according to claim 1, characterized in that, Upon startup, detection of sudden environmental changes, or user-initiated triggering, an acoustic scene detection process is performed to acquire environmental noise characteristics, including: When the device is started, the acoustic scene changes, or the user actively triggers it, the acoustic scene noise is collected for a set duration at a set sampling frequency. The reverberation time, power spectrum estimation, and spectral flatness of acoustic scene noise are calculated as acoustic scene noise characteristics.

3. The video conferencing echo suppression method based on a low-latency adaptive learning model according to claim 2, characterized in that, The reverberation time acquisition process includes: bandpass filtering and direct current removal of the acquired acoustic scene noise signal according to the set frequency band; calculating the square of the filtered and direct current removed signal to obtain the energy envelope; and performing inverse integration or accumulation of the energy envelope from the end to the beginning. Where N is the total number of samples of acoustic scene noise. k This is a sample index for acoustic scene noise. for k The energy envelope at the point, where n is the sample index of the initial acoustic scene noise signal; the integral or summation result of the energy envelope. Take the logarithm to convert to decibels and normalize; find the sample points corresponding to -5dB and -35dB on the normalized curve. and , for [ , Mapping to a time interval, and normalizing the curve within the corresponding time interval. Performing linear regression and fitting the data yields: , where a is the slope of the fit, representing the decay rate, t is the time, and the reverberation time is: 60 / |a|; The power spectrum estimation process includes: The acoustic scene noise signal is divided into frames and windowed; Fourier transform is performed on the windowed acoustic scene noise signal s(n) of each frame to obtain the corresponding acoustic scene noise frequency domain signal; the square of the modulus of the acoustic scene noise frequency domain signal of each frame is calculated to obtain the power spectrum of the acoustic scene noise of each frame; the power spectrum of all acoustic scene noise frames is averaged to obtain a smooth power spectrum estimate. The process of obtaining spectral flatness includes: Given the power spectrum of a frame of acoustic scene noise signal Calculate its flatness SF as follows: ； in, k = 0, 1, ..., N / 2 (Taking the positive frequency part), the numerator is the geometric mean of the power spectrum, and the denominator is the arithmetic mean of the power spectrum. In the calculation, logarithmic transformation is used to avoid numerical underflow. Calculate the median of the flatness of the acoustic scene noise signal across all frames, and use the median flatness as the spectral flatness of the entire acoustic scene noise.

4. The video conferencing echo suppression method based on a low-latency adaptive learning model according to claim 1, characterized in that, Real-time audio signals from video conferences are acquired, including near-end microphone signals and far-end reference signals. Preprocessing and extraction of multimodal features from the real-time audio signals include: According to the set sampling frequency, the near-end microphone signal d(n) and the far-end reference signal x(n) are synchronously acquired; the real-time audio signal of each frame of the video conference is preprocessed and multimodal features are extracted in a pipeline manner. The preprocessing includes: pre-emphasis, framing and windowing to convert the continuous signal into a short-time stationary frame sequence; multimodal features are extracted from the preprocessed real-time audio signal. The multimodal features include: Mel-spectral difference, frame-level energy ratio of the near-end microphone signal and the far-end reference signal, flatness of the current near-end microphone signal frame, normalized cross-correlation function of the near-end microphone signal d(n) and the far-end reference signal x(n), fundamental frequency features and LPC coefficients.

5. The video conferencing echo suppression method based on a low-latency adaptive learning model according to claim 1, characterized in that, The process of obtaining the 8-dimensional Mel spectrum difference includes: passing the near-end microphone signal d(n) and the far-end reference signal x(n) through a Mel filter bank with a set number of channels to obtain their respective Mel spectra; calculating the logarithmic domain difference of the Mel spectra of the near-end microphone signal d(n) and the far-end reference signal x(n) in each channel; performing a discrete cosine transform on the logarithmic domain difference of the Mel filter bank in each channel, and retaining the first 8 low-order coefficients; Obtain the frame-level energy ratio of the near-end microphone signal d(n) and the far-end reference signal x(n): ; Calculate the normalized cross-correlation function between the near-end microphone signal d(n) and the far-end reference signal x(n). The cross-correlation function is a measure of the similarity or correlation between two signals at different time offsets. Find the maximum peak value of the cross-correlation function. and their corresponding delays ; Calculate the fundamental frequency characteristics of the near-end microphone signal d(n) and the far-end reference signal x(n), including: The fundamental frequency was estimated from the near-end microphone signal d(n) and the far-end reference signal x(n) using the autocorrelation method. and According to the base frequency and Calculate the fundamental frequency characteristics: The fundamental frequency characteristics include: the first fundamental frequency characteristic and the second fundamental frequency characteristic; if the fundamental frequency... and Both are valid; the first fundamental frequency characteristic is taken as the fundamental frequency ratio: Otherwise, it is 0; the second fundamental frequency feature uses the near-end voiced tone marker.

6. The video conferencing echo suppression method based on a low-latency adaptive learning model according to claim 1, characterized in that, The method of filling silent segments with artificial noise that matches the spectral characteristics of the acoustic scene noise includes: in the frequency domain, generating random phases uniformly distributed in [0, 2π], and constructing amplitudes using the power spectrum of the acoustic scene noise. , To obtain the power spectrum of the acoustic scene noise signal, frequency domain noise is constructed using amplitude and frequency; inverse Fourier transform is performed on the frequency domain noise to obtain time domain comfort noise; comfort noise is mixed into the silent section of the enhanced time domain signal according to a certain proportion.

7. The video conferencing echo suppression method based on a low-latency adaptive learning model according to claim 1, characterized in that, After a set number of frames, the difference between the current echo return loss enhancement and the average echo return loss enhancement over a set period of time is calculated. If the difference is less than a set threshold, incremental learning is triggered, including: after a set number of frames, the logarithm of the ratio of the sum of the squares of the near-end microphone signal to the sum of the squares of the residual signal is multiplied by 10 as the echo return loss enhancement. Where d(n) is the near-end microphone signal used to calculate the echo return loss enhancement, and e(n) is the residual signal used to calculate the echo return loss enhancement. The echo return loss enhancement measures the instantaneous performance of the adaptive filter. The higher the echo return loss enhancement, the better the echo cancellation. Calculate the difference ΔERLE between the current echo return loss enhancement and the average echo return loss enhancement over a set time period in the past. If the difference is less than a set threshold, trigger incremental learning of the lightweight LSTM neural network model.

8. The video conferencing echo suppression method based on a low-latency adaptive learning model according to claim 1, characterized in that, The incremental learning includes: A circular buffer is set up to continuously store real-time audio data within a set time period, including: far-end reference signal, near-end microphone signal, and adaptively filtered residual signal; multi-dimensional feature vectors of the real-time audio data in the circular buffer are extracted; when learning is triggered, the real-time audio data in the circular buffer is used as the fine-tuning dataset for model fine-tuning: during fine-tuning, the first layer of the lightweight LSTM network is frozen, and only the second layer of the network and the subsequent fully connected layers are unlocked for fine-tuning; scale-invariant signal-to-noise ratio is used as the loss function, and a lightweight optimizer such as momentum-driven SGD or AdamW is used for a set number of iterations. The fine-tuned model is forward-inferred on another small segment of cached data to calculate the echo return loss enhancement of the processed signal. If the echo return loss enhancement of the fine-tuned model exceeds the set optimization threshold compared to the old model, the fine-tuning is considered effective, and the parameters of the old model currently running in memory are atomically replaced with the parameters of the fine-tuned new model; if the validation fails, the fine-tuned model is discarded, the original model is used again, and the failure is recorded, and learning is not triggered again for a short period of time.

9. A video conferencing echo suppression device based on a low-latency adaptive learning model, comprising: At least one processing unit, wherein the processing unit is connected to a storage unit, a near-end acquisition unit and a far-end acquisition unit via a bus unit, wherein the storage unit stores a computer program, and the processing unit implements the video conferencing echo suppression method based on a low-latency adaptive learning model as described in any one of claims 1-8 by running the computer program stored in the storage unit.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed, implements the video conferencing echo suppression method based on a low-latency adaptive learning model as described in any one of claims 1-8.

Images