Feedback suppression method, device, medium and product in sound reinforcement scenario

Abstract

The present invention provides a feedback suppression method, device, medium and product in a sound reinforcement scenario. The method includes: obtaining a microphone signal of the current frame; performing frequency-domain adaptive filtering processing on the microphone to be processed, and outputting a predicted target signal to a speaker. Among them, the frequency-domain adaptive filtering processing includes: multiplying an estimated acoustic transfer function of the previous frame by the speaker signal of the previous frame in the frequency domain to obtain a feedback estimated signal, and then subtracting the feedback estimated signal from the microphone signal to obtain a residual signal; the residual signal is output to the speaker as a predicted target signal after inverse transformation. At the same time, the estimated acoustic transfer function of the previous frame is iterated to obtain the estimated acoustic transfer function of the current frame, and the step size of the iteration is controlled by a recursive neural network model. The present invention can improve the accuracy of acoustic feedback estimation in a variety of actual scenarios, improve the distortion degree in the sound reinforcement scenario and increase the maximum gain.

Description

Technical Field

[0001] The present invention relates to the technical field of audio processing, and specifically to a feedback suppression method, device, medium and product in a sound reinforcement scenario. Background Art

[0002] The howling phenomenon is a common problem in audio systems, commonly found in scenarios such as conference systems, classroom sound reinforcement, speeches, KTVs, and hearing aids. The root cause of howling is that the sound output by the speaker is picked up and amplified by the microphone again, forming a positive feedback loop. In this loop, when certain specific frequency signals accumulate and amplify rapidly in the feedback loop, a harsh howling sound will be generated. The existing howling suppression technologies mainly include four categories: frequency shifting, notch filtering, deep voice enhancement, and adaptive filtering.

[0003] The frequency shifting technology emerged earlier and is relatively simple to implement. Frequency shifting can shift the peak of the feedback signal wave, causing the signal to be superimposed out of phase in the feedback loop, so as to slow down the speed of feedback amplification. Frequency shifting can slightly increase the gain when howling occurs, usually by 1dB - 2dB. Frequency shifting will cause significant distortion, and at the same time, the gain that can be increased is limited.

[0004] The notch filtering technology includes static notch filtering and adaptive notch filtering. Static notch filtering requires measuring the specific acoustic environment and designing a series of notch filters to filter the frequency bands where howling is most likely to occur. Adaptive notch filtering automatically estimates the howling frequency through howling detection and automatically designs filters to suppress the howling frequency. The detection methods of howling frequency points can be divided into two categories: fixed rules based on spectral characteristics and deep learning methods. However, the use of notch filters will cause signal distortion. In complex acoustic environments, such as scenarios with multiple speakers, multiple microphones, and moving microphones, there may be too many frequency points where howling may occur. The use of a large number of notch filters will cause serious distortion and make the speech unintelligible. When the howling detection is relatively reliable, the gain when howling occurs can usually be increased by 6dB - 8dB through notch filtering.

[0005] The deep voice enhancement technology realizes howling suppression through methods of deep learning for voice enhancement. These methods do not rely on known speaker signals, nor on the prediction of acoustic transfer functions. They directly remove howling sounds from the microphone signal in the way of noise cancellation through a neural network model. Because they do not rely on the estimation of acoustic transfer functions, they are very robust to changes in the acoustic environment. However, since they do not use known speaker signals as a reference, it is difficult for the model to accurately estimate the feedback signal and stably restore the original sound source. In particular, when the feedback is strong, the signal-to-noise ratio of the microphone signal is low, and the signal after voice enhancement is often severely distorted. In addition, sounds similar to howling in characteristics, such as some musical instruments, are easily mis-cancelled.

[0006] Adaptive filtering technology is a signal processing technology that can continuously iterate its parameters through the real-time input and output of the system to approach the predefined error minimization. It estimates the feedback signal and eliminates it according to the estimated feedback signal to achieve the purpose of restoring the target voice. Since the speaker signal is known, when the acoustic transfer function is completely known, not only can the occurrence of howling be completely avoided, but also the feedback can be perfectly eliminated to truly restore the target sound source. Therefore, there is no upper limit to the improvement of the maximum gain by perfect adaptive filtering. The difficulty of adaptive filtering lies in the estimation of the acoustic transfer function. Adaptive filtering iteratively adjusts the estimation of the acoustic transfer function through the result of feedback cancellation to gradually approach the true acoustic transfer function. Adaptive filtering is divided into time-domain adaptive filtering and frequency-domain adaptive filtering. According to the scope of action of the estimated acoustic transfer function, time-domain adaptive filtering is achieved through time-domain convolution. Usually, the delay is small, but the fitting of long acoustic transfer functions is insufficient, and it is widely used in the field of hearing aids; frequency-domain adaptive filtering processes the signal in the frequency domain, replacing the convolution operation in time-domain adaptive filtering with multiplication. The time complexity increases slowly with the increase of the filter length. Therefore, a longer filter can be used in time, and the fitting ability for long acoustic transfer functions is stronger, which is suitable for the sound reinforcement scenario.

[0007] In traditional frequency-domain adaptive filtering, the control of the acoustic transfer function iteration can be classified as Wiener filtering, Kalman filtering, etc. based on different theoretical assumptions. The assumptions include that the acoustic transfer function does not change suddenly, the feedback signal is not correlated with the current target signal, etc. In practical applications, these assumptions are often not met. For example, when the direction and position of the microphone may change suddenly, or when the feedback signal is highly correlated with the target signal during a long note in singing. At this time, there will be a huge error in the estimation of the acoustic transfer function and even the feedback, resulting in serious distortion of the sound reinforcement or no howling suppression effect. An existing solution proposes to predict the mutation of the feedback path through a neural network to overcome the problem of locking in Kalman filtering, which specifically solves one problem of statistical adaptive filtering, but cannot systematically solve the defects of statistical adaptive filtering. Another existing solution uses a neural network to assist adaptive filtering, but it still uses statistical methods for adaptive filtering. It only judges a situation of step-size locking through a neural network and resets it specifically, which cannot meet the requirements in multi-voice scenarios. Summary of the Invention

[0008] The first object of the present invention is to provide a feedback suppression method in a sound reinforcement scenario, which can improve the accuracy of acoustic feedback estimation in various actual scenarios, improve the distortion degree in the sound reinforcement scenario, and increase the maximum gain.

[0009] The second object of the present invention is to provide a computer device for implementing the above-mentioned feedback suppression method in a sound reinforcement scenario.

[0010] The third object of the present invention is to provide a computer-readable storage medium for implementing the feedback suppression method in the above sound reinforcement scenario.

[0011] The fourth object of the present invention is to provide a computer program product for implementing the feedback suppression method in the above sound reinforcement scenario.

[0012] To achieve the above first object, the present invention provides a feedback suppression method in a sound reinforcement scenario, which includes the following steps: obtaining a microphone signal of the current frame; performing frequency-domain adaptive filtering processing on the signal of the current frame and outputting a predicted target signal to a speaker: wherein, the frequency-domain adaptive filtering processing includes the following steps: in the frequency domain, multiplying the estimated acoustic transfer function of the previous frame by the speaker signal of the previous frame to obtain a feedback estimated signal; in the frequency domain, subtracting the feedback estimated signal from the microphone signal of the current frame to obtain a residual signal; the residual signal is used as the predicted target signal after inverse transformation and is output to the speaker through a feedforward path. At the same time, the estimated acoustic transfer function of the previous frame is iterated in a direction where the cross-correlation between the residual signal of the current frame and the speaker signal of the previous frame approaches a smaller value to obtain the estimated acoustic transfer function of the current frame, and the step size of the iteration is controlled by a pre-trained recursive neural network model.

[0013] As can be seen from the above solution, the present invention suppresses howling through frequency-domain adaptive filtering. Among them, the step size is controlled by a pre-trained recursive neural network model to iterate the estimated acoustic transfer function, replacing the method of controlling the adaptive process based on statistical assumptions in traditional adaptive filtering, which can cover the accuracy of acoustic feedback estimation in more actual scenarios, and significantly improve the distortion degree and the maximum gain in the sound reinforcement scenario.

[0014] A further solution is that during the training process of the recursive neural network model, it includes the following steps: simulating impulse responses changing on different moving trajectories in rooms of different sizes; simulating the feedback path by using the collected voice and music data through the generated impulse responses and processing them using adaptive filtering, wherein the iteration step size of the adaptive filtering is controlled by the recursive neural network model, and the loss function is the mean square error between the simulated feedback signal and the amplitude spectrum of the feedback signal estimated in the adaptive filtering during the overall simulation process.

[0015] Thus, the collected voice and music data are used as training data to cover diversified scenarios such as sudden changes in acoustic functions, long drawn-out sounds, and strong background noises. During the adaptive process, the iteration step size of the estimated acoustic transfer function depends on the minimization regression of the estimation error of the feedback signal in the training data, ensuring the lowest feedback estimation and the most accurate sound restoration in various real scenarios. It solves the problem that traditional adaptive filtering does not model real audio data and is not optimized enough in many real scenarios.

[0016] A further solution is that during the training of the recurrent neural network model, when simulating the feedback path by using the collected speech and music data through the generated impulse response, the feedback gain corresponding to the feedback path is random within a set range. Among them, at the beginning of training, the feedback gain needs to be set below the critical gain.

[0017] It can be seen that during the initial training process, premature divergence of the estimation of the acoustic transfer function is avoided, enabling the model to learn better.

[0018] A further solution is that the input of the recurrent neural network model is a vector formed by combining the amplitude spectrum of the microphone signal and the amplitude spectrum of the residual signal for each frame, and the output is a vector of the same length as the input as the hidden layer. The hidden layer is activated through a linear layer and uses the logistic function, and a single scalar is output as the common step size for all subbands in the current frame during iteration.

[0019] A further solution is that the input of the recurrent neural network model is a vector formed by combining the amplitude spectrum of the microphone signal and the amplitude spectrum of the residual signal for each frame, and the output is a vector of the same length as the input as the hidden layer. The hidden layer is activated through a linear layer and uses the logistic function, and the step sizes used for iteration of each subband in the current frame are output respectively.

[0020] It can be seen that the recurrent neural network model can also output the step sizes used for iteration of each subband in the current frame, thereby improving the iterative effect of estimating the acoustic transfer function.

[0021] A further solution is that the model structure of the recurrent neural network model is LSTM or GRUs.

[0022] A further solution is that the estimated acoustic transfer function of the previous frame is iterated in the direction where the cross-correlation between the residual signal of the current frame and the speaker signal of the previous frame approaches a smaller value. The expression is:

[0023] , is the gradient of the m-th subband during the iteration of the estimated acoustic transfer function of the (k - 1)-th frame, represents the m-th subband of the residual signal of the k-th frame, is the m-th subband of the speaker signal of the k-th frame, is the conjugate complex number of, is the recursive smoothing of the speaker spectral energy of the m-th subband of the k-th frame, ,where, is the smoothing coefficient; the estimated acoustic transfer function of the current frame is obtained:

[0024] ,where, is the m-th sub-band of the estimated acoustic transfer function for the k-th frame, is the m-th sub-band of the estimated acoustic transfer function for the (k - 1)-th frame, is the step size of the iteration.

[0025] To achieve the above-mentioned second object, a computer device provided by the present invention includes a processor and a memory, wherein: a computer program is stored on the memory, and when the computer program is executed by the processor, the above-mentioned feedback suppression method based on neural network controlled adaptive filtering is implemented.

[0026] To achieve the above-mentioned third object, a computer-readable storage medium provided by the present invention has a computer program stored thereon, wherein: when the computer program is executed by a processor, the above-mentioned feedback suppression method in a sound reinforcement scenario is implemented.

[0027] To achieve the above-mentioned fourth object, a computer program product provided by the present invention includes computer instructions, wherein: when the computer instructions are executed by a processor, the above-mentioned feedback suppression method in a sound reinforcement scenario is implemented. Description of the Drawings

[0028] Figure 1 is a flowchart of an embodiment of the feedback suppression method in a sound reinforcement scenario of the present invention.

[0029] Figure 2 is a principle block diagram of an embodiment of the feedback suppression method in a sound reinforcement scenario of the present invention.

[0030] The present invention will be further described below in conjunction with the drawings and embodiments. Detailed Embodiment

[0031] The feedback suppression method in a sound reinforcement scenario of the present invention controls the iteration of the step size through a recurrent neural network trained on a large amount of real data, improves the prediction accuracy of the estimated acoustic transfer function in various actual scenarios, and improves the maximum gain and distortion that can be achieved in a sound reinforcement scenario. The present invention also provides a computer device, a computer-readable storage medium, and a computer program product for implementing the above-mentioned howling suppression method.

[0032] Embodiment of the feedback suppression method in a sound reinforcement scenario:

[0033] This embodiment will be described with an indoor sound reinforcement scenario as an example. A microphone and a speaker are installed indoors. The signal collected by the microphone is processed by an adaptive filtering system and then output to the speaker. In this scenario, the target sound source is collected by the microphone and then output to the speaker. The sound played by the speaker is reflected by the surrounding environment and then collected by the microphone. The collected feedback signal is analyzed and eliminated by the adaptive filtering system in real time, so that no howling phenomenon occurs during the sound reinforcement process of the target sound source, achieving the purpose of howling suppression.

[0034] Assume that the speaker signal is played, the sound wave is reflected in the room and then picked up again by the microphone, and then played by the speaker again to form a closed-loop system. The acoustic transfer function is usually simplified to a linear time-invariant system (Linear Time-Invariant, LTI), based on the assumption that the actual physical environment is stable and the positions of the sound source and the receiver are fixed. The acoustic transfer function (Acoustic Transfer Function, ATF) is a mathematical representation that describes the change of the audio signal during the process of sound propagation from the sound source to the receiving point. It basically describes how the acoustic system "transfers" or "transforms" the input signal.

[0035] Thus, in this closed-loop system with feedback, the microphone signal to be processed is expressed as , where represents the parameter of a linear system with L points at time point n, represents the last L sampling points of the speaker signal at time point n, represents the signal of the target sound source, which is the target that the adaptive filtering system needs to restore from

[0036] The adaptive filtering system realizes howling suppression by executing the feedback suppression method in the sound reinforcement scenario of this embodiment. Specifically, the adaptive filtering system is implemented by a computer program. Refer to Figure 1 , when the computer program is executed, it includes the following steps:

[0037] S11: Obtain the microphone signal of the current frame.

[0038] S12: Perform frequency-domain adaptive filtering processing on the microphone signal of the current frame.

[0039] S13: Output the predicted target signal to the speaker, and at the same time iterate the estimated acoustic transfer function of the previous frame to obtain the estimated acoustic transfer function of the current frame.

[0040] Refer to Figure 2 , in the above step S11, the microphone signal of the current frame in the frequency domain collected by the microphone 20 includes the signal of the target sound source , and the speaker signal of the previous frame After entering the speaker 30, the speaker 30 outputs the actual feedback signal obtained through the actual acoustic transfer function H.

[0041] The duration of each frame of the microphone signal is the same as that of each frame of the speaker signal, and after short-time Fourier transform, they are divided into the same number of subbands in the frequency domain. The m-th subband of the microphone signal of the current frame (the k-th frame) is expressed as​ , the m-th subband of the speaker signal of the previous frame (the (k - 1)-th frame) is denoted as . The m-th subband of the estimated acoustic transfer function of the previous frame (the (k - 1)-th frame) is denoted as . In the above step S12, in the frequency domain, the feedback estimation signal is obtained by multiplying the estimated acoustic transfer function of the previous frame by the speaker signal of the previous frame, specifically including: in the frequency domain, through the m-th subband of the estimated acoustic transfer function of the previous frame and the m-th subband of the speaker signal of the previous frame are multiplied to obtain the m-th subband of the feedback estimation signal of the k-th frame . Then, in the frequency domain, the microphone signal of the current frame is subtracted from the feedback estimation signal to obtain the residual signal, specifically including: each subband of the microphone signal of the current frame is subtracted from the corresponding subband of the feedback estimation signal, that is, the residual signal corresponding to the subband is obtained, and the m-th subband of the k-th frame residual signal is denoted as , .

[0042] In the above step S13, after the residual signal undergoes an inverse short-time Fourier transform, the predicted target signal in the time domain is obtained, and the predicted target signal is output to the speaker 30 through the feedforward path G. The feedforward path G refers to the channel through which the signal is transmitted from the system input to the output, usually including post-processing such as a limiter. At the same time, the estimated acoustic transfer function of the previous frame is iterated in the direction where the cross-correlation between the residual signal of the current frame and the speaker signal of the previous frame approaches a smaller value to obtain the estimated acoustic transfer function of the current frame, and the step size of the iteration is controlled by the pre-trained recurrent neural network model 10.

[0043] During the iteration of the estimated acoustic transfer function of the previous frame, each subband is independently processed. Specifically, while determining the speaker output, the estimated acoustic transfer function of the previous frame is iterated in the direction where the cross-correlation with the residual signal of the current frame and the speaker signal of the previous frame approaches a smaller value, and this process is expressed as: , is the gradient of the m-th subband during the iteration of the estimated acoustic transfer function of the k-th frame, represents the m-th subband of the residual signal of the k-th frame, is the m-th subband of the speaker signal of the k-th frame, is 's conjugate complex number, is the recursive smoothing of the speaker spectral energy of the m-th subband of the k-th frame, , where is the smoothing coefficient, which is empirically set between 0.8 and 0.9, so as to obtain the m-th subband of the estimated acoustic transfer function for the k-th frame , where is the m-th subband of the estimated acoustic transfer function for the k-th frame, is the m-th subband of the estimated acoustic transfer function for the (k - 1)-th frame, is the step size of the iteration, that is, during the iteration of the estimated acoustic transfer function for the (k - 1)-th frame, the step size for each subband iteration is . The step size for each frame iteration is controlled by a pre-trained recurrent neural network model to find the most suitable step size relative to the actual feedback signal.

[0044] The input of the recurrent neural network model in this embodiment is a vector formed by combining the amplitude spectrum of the microphone signal and the amplitude spectrum of the residual signal for each frame, and the output is a vector of the same length as the input as the hidden layer. The hidden layer is activated by a linear layer and a logistic function, and the output is a single scalar as the common step size for all subbands in the current frame iteration, as follows: , where RNN is the above-mentioned recurrent neural network model, is the hidden state of the recurrent neural network, which is used to remember the previous information in the time series and pass it backward.

[0045] In different embodiments, the output of the pre-trained recurrent neural network model is a vector , and this vector represents the step size used for each subband iteration in the current frame, that is . Thus, during the process of iterating the estimated acoustic transfer function of the previous frame to obtain the estimated acoustic transfer function of the current frame, the iteration is performed with the step size corresponding to each subband in the current frame. For example, the m-th subband of the estimated acoustic transfer function for the k-th frame , the (m - 1)-th subband of the estimated acoustic transfer function for the k-th frame , the (m - 2)-th subband of the estimated acoustic transfer function for the k-th frame .

[0046] Optionally, the model structure of the recurrent neural network model is Long Short-Term Memory (LSTM) or Gated Recurrent Units (GRUs).

[0047] The input and output of the recurrent neural network model during the training process are generated in real-time simulation through a simulation feedback environment. The training process includes the following steps:

[0048] Simulate the impulse responses changing on different movement trajectories in rooms of different sizes, which can be specifically implemented by the existing IMAGE method. Simulate the feedback path of the collected voice and music data through the generated impulse responses, and use adaptive filtering for processing. Among them, the iteration step size of the adaptive filtering is controlled by a recurrent neural network model, and the loss function is the mean square error between the simulated feedback signal and the amplitude spectrum of the feedback signal estimated in the adaptive filtering during the overall simulation process. The loss function comes from the entire simulation process to avoid overfitting of the feedback prediction for a certain frame. The feedback path refers to the path where the signal output by the system is transmitted from the output end to the input end. In the sound reinforcement scenario, the feedback path is the above-mentioned actual acoustic transfer function.

[0049] When training the recurrent neural network model, during the process of simulating the feedback path of the collected voice and music data through the generated impulse responses, the feedback gain corresponding to the feedback path is random within a set range. Among them, in the initial stage of training, the feedback gain needs to be set below the critical gain to avoid premature divergence of the estimated acoustic function, so that the model cannot learn. As the accuracy of the model effect improves, after being able to avoid howling at a higher gain, the feedback gain range in the simulation can be continuously increased to enable the model to gradually have the ability to handle higher feedback gains. In the context of howling suppression, the critical gain refers to the maximum gain value that the acoustic system can withstand before starting to generate howling, usually expressed in decibels (dB). The critical gain is also called the Maximum Stable Gain (MSG), that is, the upper limit of the gain at which the system can operate safely in actual applications.

[0050] In summary, the present invention suppresses howling by means of frequency-domain adaptive filtering. Among them, a pre-trained recurrent neural network model is used to control the step size for iterative estimation of the acoustic transfer function, replacing the traditional way of controlling the adaptive process based on statistical assumptions in adaptive filtering. Its advantages are that there is no theoretical upper limit to the improvement of the Maximum Stable Gain (ASG), and it does not adopt a sound quality-damaging processing method, ensuring high sound restoration after processing. The howling suppressor or feedback suppressor implemented based on the present invention can cover the accuracy of acoustic feedback estimation in more actual scenarios, significantly improving the distortion degree and maximum gain in the sound reinforcement scenario.

[0051] Embodiment of a computer device:

[0052] The computer device of this embodiment includes a processor and a memory. The memory stores a computer program, and when the processor executes the computer program, it implements the above-mentioned embodiment of the feedback suppression method in the sound reinforcement scenario.

[0053] A computer device may include, but is not limited to, a processor and a memory. Those skilled in the art can understand that a computer device may include more or fewer components, or combine certain components, or different components. For example, a computer device may also include input / output devices, network access devices, buses, etc.

[0054] For example, the processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microcontroller or the processor may also be any conventional processor, etc. The processor is the control center of the computer device, connecting various parts of the entire computer device through various interfaces and lines.

[0055] The memory can be used to store computer programs and / or modules. The controller realizes various functions of the computer device by running or executing the computer programs and / or modules stored in the memory, and by calling the data stored in the memory. For example, the memory may mainly include a program storage area and a data storage area. Among them, the program storage area may store an operating system, application programs required for at least one function (such as a voice reception function, a voice-to-text conversion function, etc.); the data storage area may store data created according to the use of the mobile phone (such as audio data, text data, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0056] Examples of computer-readable storage media:

[0057] If the modules integrated in the computer device of the above embodiments are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes of the feedback suppression method embodiments in the sound reinforcement scenario, it can also be completed by a computer program instructing relevant hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a controller, the steps of the feedback suppression method embodiments in the above sound reinforcement scenario can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The storage medium can include: any entity or device capable of carrying computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0058] Embodiment of computer program product:

[0059] The computer program product of this embodiment includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes each step of the feedback suppression method embodiments in the above sound reinforcement scenario.

[0060] Finally, it should be emphasized that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A feedback suppression method in a sound reinforcement scenario, characterized in that Including the following steps: Obtain the microphone signal of the current frame; Perform frequency-domain adaptive filtering on the microphone signal of the current frame and output the predicted target signal to the speaker: Wherein, the frequency-domain adaptive filtering includes the following steps: In the frequency domain, multiply the estimated acoustic transfer function of the previous frame by the speaker signal of the previous frame to obtain a feedback estimated signal; In the frequency domain, subtract the feedback estimated signal from the microphone signal of the current frame to obtain a residual signal; The residual signal is inversely transformed to be used as the predicted target signal and output to the speaker through a feedforward path. Meanwhile, the estimated acoustic transfer function of the previous frame is iterated in the direction where the cross-correlation between the residual signal of the current frame and the speaker signal of the previous frame approaches a smaller value, to obtain the estimated acoustic transfer function of the current frame, and the step size of the iteration is controlled by a pre-trained recurrent neural network model; During the training process of the recurrent neural network model, it includes the following steps: Simulate impulse responses that vary on different moving trajectories in rooms of different sizes; Simulate the feedback path for the collected speech and music data through the generated impulse response and perform processing using adaptive filtering, wherein the iteration step size of the adaptive filtering is controlled by the recurrent neural network model, and the loss function is the mean square error between the simulated feedback signal and the amplitude spectrum of the feedback signal estimated in the adaptive filtering during the overall simulation process.

2. The feedback suppression method in an amplified sound scenario according to claim 1, wherein: When the recurrent neural network model is trained, during the process of simulating the feedback path for the collected speech and music data through the generated impulse response, the feedback gain corresponding to the feedback path is random within a set range, and at the beginning of the training, the feedback gain is set below the critical gain.

3. The feedback suppression method in an amplified sound scenario according to claim 1, wherein: The model structure of the recurrent neural network model is LSTM or GRUs.

4. The feedback suppression method in an amplified sound scenario according to claim 1, wherein: The input of the recurrent neural network model is a vector formed by merging the amplitude spectrum of the microphone signal of each frame and the amplitude spectrum of the residual signal, and the output is a vector of the same length as the input as a hidden layer. The hidden layer is activated by a linear layer and a logistic function, and outputs a single scalar, and the single scalar is used as the common step size for all subbands of the current frame during iteration.

5. The feedback suppression method in an amplified sound scenario according to claim 1, wherein: The input of the recurrent neural network model is a vector formed by merging the amplitude spectrum of the microphone signal of each frame and the amplitude spectrum of the residual signal, and the output is a vector of the same length as the input as a hidden layer. The hidden layer is activated by a linear layer and a logistic function, and respectively outputs the step sizes used for iteration of each subband of the current frame.

6. The feedback suppression method in an amplified sound scenario according to any one of claims 1 to 4, wherein: The estimated acoustic transfer function of the previous frame is iterated in the direction where the cross-correlation between the residual signal of the current frame and the loudspeaker signal of the previous frame approaches a smaller value, and the expression is: , is the gradient of the m-th subband when iterating the estimated acoustic transfer function of the (k - 1)-th frame, represents the m-th subband of the residual signal of the k-th frame, is the m-th subband of the loudspeaker signal of the k-th frame, is the conjugate complex number of, is the recursive smoothing of the loudspeaker spectral energy of the m-th subband of the k-th frame, , where, is the smoothing coefficient; Obtain the estimated acoustic transfer function of the current frame: , where is the m-th subband of the estimated acoustic transfer function for the k-th frame, is the m-th subband of the estimated acoustic transfer function for the (k - 1)-th frame, is the step size of the iteration.

7. A computer device, comprising a processor and a memory, characterized in that: A computer program is stored on the memory, and when the computer program is executed by the processor, it implements the feedback suppression method in the sound reinforcement scenario described in any one of claims 1 to 6 above.

8. A computer-readable storage medium, on which a computer program is stored, characterized in that: When the computer program is executed by the processor, it implements the feedback suppression method in the sound reinforcement scenario described in any one of claims 1 to 6 above.

9. A computer program product, comprising computer instructions, characterized in that: When the computer instructions are executed by the processor, they implement the feedback suppression method in the sound reinforcement scenario described in any one of claims 1 to 6 above.

Images