Dynamic howling suppression optimization design method based on FPGA

Abstract

The invention relates to the field of dynamic howling suppression, in particular to a dynamic howling suppression optimization design method based on an FPGA (Field Programmable Gate Array), which comprises the steps of signal acquisition and preprocessing, gain processing and howling detection, noise injection and path estimation and howling suppression processing. According to the dynamic howling suppression optimization design method based on the FPGA, howling signals can be detected in a short time, suppression is carried out at the same time, the maximum stable gain of a system is improved, the tone quality of output audio signals is guaranteed, the effect is good, and compared with other methods, the dynamic howling suppression optimization design method based on the FPGA has the advantages that the method is easy to implement. The real-time dynamic howling suppression algorithm can accurately estimate the sound feedback and eliminate the sound feedback from the original sound, and has a good effect and high sound fidelity.

Description

Technical Field

[0001] The present invention relates to the field of dynamic howling suppression, and particularly to an optimized design method for dynamic howling suppression based on FPGA. Technical Background

[0002] With the rapid development of digital signal processing technology, people's requirements for speech signals have been gradually improved; in on-site meetings, howling is a common phenomenon. The generation of howling will not only reduce the gain of the sound reinforcement system and fail to reach the minimum sound pressure level required by the meeting, but more seriously, it will damage the sound reinforcement equipment in the meeting and cause certain harm to the human ear; for the howling problem in on-site sound reinforcement meetings, the most commonly used methods include using special acoustic rooms, equipment, etc., frequency shifting method, amplitude compression method, notch method. However, these methods are expensive to implement and have poor effects.

[0003] Currently, adaptive howling suppression technology has become the mainstream algorithm for howling suppression. This algorithm first estimates the feedback path, convolves the estimated feedback path with the speaker signal to obtain the estimated feedback signal, and then subtracts it from the input signal to obtain the signal after feedback suppression; the most commonly used algorithms include Normalized Least Mean Squares (NLMS) algorithm, Recursive Least Square (RLS) algorithm, Normalized Sub-band Adaptive Filter (NSAF) algorithm, etc. However, these algorithms have the disadvantages of requiring a large number of filter taps for feedback path estimation, large computational complexity, large howling suppression delay, and difficulty in real-time calculation.

[0004] Therefore, those skilled in the art have provided an optimized design method for dynamic howling suppression based on FPGA to solve the problems raised in the above background technology. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides an optimized design method for dynamic howling suppression based on FPGA, including the following steps: S1. Signal acquisition and preprocessing: Collect audio signals through a microphone, convert them into digital signals by an audio acquisition module, and then input them into the FPGA chip; S2. Gain processing and howling detection: Adjust the gain of the input signal, and then perform howling detection; when howling is detected, switch the output to Gaussian white noise through the audio output control module; when no howling is detected, output the received audio signal; use an improved howling detection algorithm for real-time detection; S3. Noise injection and path estimation: When howling is detected, switch the output to Gaussian white noise and estimate the acoustic feedback path parameters; When the loudspeaker outputs Gaussian white noise, the microphone collects the signal of the white noise after passing through the acoustic feedback path. At this time, acoustic feedback path estimation calculation is performed to obtain the acoustic feedback path parameters, and the output is switched to the normal acoustic feedback audio signal through the audio output control module; S4. Howling suppression processing: Filter the original signal according to the estimated acoustic feedback path parameters and output the suppressed audio signal.

[0006] Preferably, the process of howling detection includes: SS1. Frame and window the input signal, and obtain the signal power spectrum through short-time Fourier transform; SS2. Determine the candidate howling frequency points. Since the howling component is a single-frequency narrowband signal with continuously increasing energy, there is relatively high energy at the howling frequency points. Therefore, usually several frequency points with relatively large signal power amplitudes in each frame are selected as the candidate howling frequency points; SS3. Then perform detection and decision based on the set detection threshold T.

[0007] Preferably, the method for estimating the acoustic feedback path parameters: When howling is detected, the loudspeaker outputs a Gaussian white noise signal with a short-time mean of zero, and the feedback signal is used to estimate the acoustic feedback path parameters; When the loudspeaker outputs a Gaussian white noise signal with a short-time zero mean, it can be considered that the output Gaussian white noise and the pure speech signal input by the mic are independent of each other. At this time, the signal collected by the microphone can be expressed as: Equation 1 In the formula represents the FIR model coefficients of the acoustic feedback path; vector , vector represents the acoustic feedback signal received by the microphone after the loudspeaker output signal passes through the feedback path with P taps; P is the number of paths of the acoustic feedback path; is the pure speech signal; At this time, is the acoustic feedback Gaussian white noise signal received by the microphone after the Gaussian white noise passes through the acoustic feedback path, that is ; Therefore, and are independent of each other; The actual acoustic feedback signal received by the microphone is related to f(n) for calculation, and can be obtained: Equation 2 The value of can be obtained from Equation 2 Formula 3 wherein, =0, is autocorrelation; is cross-correlation with f(n). Preferably, it further includes a howling detection algorithm based on the frame-interval spectral flatness deviation feature, and its specific calculation process is as follows: First, calculate the signal power spectrum and select the frequency points with larger power spectra as candidate howling frequency points; Secondly, calculate the geometric mean and the average of the power spectra of the past R frames of the signal at the candidate howling frequency points ; Formula 4 Formula 5 Then, calculate the LSFM feature parameter of the signal at the frequency according to and ; Formula 6 Finally, calculate the variance of the past Q frames parameters to obtain the ISFD feature parameter ; Formula 7 Through the above analysis, the speech signal is a non-stationary signal, and the power spectrum at the frequency changes continuously with time, so the LSFM feature parameter at the frequency has large fluctuations; therefore, the value of the ISFD feature parameter of the speech signal is large; while the power spectrum of the howling signal has the same transformation characteristics at the howling frequency point, has the same spectral flatness characteristics, and the LSFM feature parameter of the howling signal at the howling frequency point has small fluctuations; therefore, the value of the ISFD feature parameter of the howling signal is small; thus, it can be seen that the ISFD feature parameter can be used for howling detection to distinguish speech from howling.

[0008] Preferably, the present solution adopts two detection methods, namely the peak harmonic energy ratio (PHPR) and the peak adjacent power ratio (PNPR), and combines the howling detection algorithm based on the frame-interval spectral flatness deviation feature as an improved howling detection optimization algorithm.

[0009] Preferably, the peak harmonic power ratio (PHPR) detection algorithm has the following expression: Formula 8 Formula 9 In the formula, represents the frequency the power at an integer multiple (harmonic) frequency, is the frequency the power at, when the power at frequency the power at and the frequency the ratio of the power at an integer multiple (harmonic) frequency exceeds the detection threshold (TH), then it is determined that is the howling frequency.

[0010] Preferably, the peak adjacent power ratio (PNPR) detection algorithm has the following expression: Formula 10 Formula 11 In the formula, represents the frequency the power of the adjacent frequency component, is the frequency the power at, when the power at frequency the power at and the frequency the ratio of the power at the adjacent frequency exceeds the detection threshold (TN), then it is determined that is the howling frequency. Preferably, the implementation steps of the howling detection optimization algorithm are as follows: Step 1: Perform VAD detection on the input signal. When the input signal is a silent signal, it is determined that no howling occurs. Otherwise, perform howling detection processing on the signal; Step 2: Frame and window the input signal, and obtain the signal power spectrum through discrete Fourier transform; Step 3: Determine the candidate howling frequency points; Since the energy at the howling frequency points is relatively large when howling occurs, generally select the frequency points with relatively large power as candidate howling frequency points; Detect the frequency point with the maximum power as the basic candidate howling frequency point for PHPR and PNPR, and then detect M larger power frequency points as the candidate howling frequency points for ISFD; Step 4: Calculate the ISFD eigenvalue, calculate the spectral flatness at each candidate howling frequency point, and obtain the ISFD eigenvalue of the current frame by calculating the variance based on the spectral flatness parameters of the past continuous Q frames at each candidate howling frequency point; Step 5: Calculate the peak harmonic energy ratio (PHPR) and the peak adjacent power ratio (PNPR); Step 6: Perform threshold decision respectively. When the peak harmonic energy ratio (PHPR) is greater than or equal to the threshold TH, it is determined that howling occurs; When the peak adjacent power ratio (PNPR) is greater than or equal to the threshold TN, it is determined that howling occurs; When the ISFD eigenvalue is less than the detection threshold TS, it is determined that howling occurs; Step 7: Finally, analyze the howling judgment result. Combine the howling results generated by the above three judgments. When the Peak Near Power Ratio (PNPR) generates howling or the Peak Harmonic Power Ratio (PHPR) determines howling, and when the ISFD eigenvalue determines howling, then it is finally determined that howling occurs; otherwise, it is determined that no howling occurs.

[0011] Preferably, the optimization design method for acoustic feedback path estimation is as follows: According to the principle of acoustic feedback path estimation, the actually received acoustic feedback signal by the microphone and The related calculation is represented by Formula 2, where =0, and Formula 2 can be simplified to Formula 12 Formula 12 When the speaker output is Gaussian white noise, is and The related calculation; Then, Formula 12 can be expanded as follows: = Formula 13 The left side of Formula 13 is a Toeplitz matrix. This kind of equation is called the Yule - Walker equation and is solved using the Levinson - Durbin recursive algorithm.

[0012] The technical effects and advantages of the present invention: Based on the common howling detection methods, a new comprehensive howling detection method is constructed by combining the howling detection algorithm based on the frame - to - frame spectral flatness deviation feature. At the same time, during the process of implementing the acoustic feedback path parameters based on FPGA, the cross - correlation calculation method is optimized using the FFT algorithm. Through a large number of simulation experiments, it is verified that the dynamic howling suppression optimization design method based on FPGA has a faster convergence speed, a smaller steady - state error, and better howling suppression performance. The experimental results show that the dynamic howling suppression optimization design method based on FPGA designed in this paper can detect howling signals in a shorter time and suppress them simultaneously, improving the maximum stable gain of the system and ensuring the sound quality of the output audio signal, with good results.

[0013] This paper adopts a real-time dynamic howling suppression algorithm, which is an adaptive short-time noise injection technology. Short-time white noise is injected into the speaker to effectively calculate the feedback path to obtain the optimal FIR filter, thereby effectively eliminating acoustic feedback (AF); this algorithm uses an adaptive FIR filter for howling suppression, automatically adjusting its own impulse response, that is, the coefficients of the digital filter, to adapt to the characteristics of signal changes, so as to achieve optimal filtering; compared with other methods, the real-time dynamic howling suppression algorithm can more accurately estimate the acoustic feedback and eliminate it from the original sound, with good effects and high sound fidelity. Description of the Drawings

[0014] Figure 1 It is the real-time howling suppression system architecture diagram of an FPGA-based dynamic howling suppression optimization design method provided by an embodiment of the present application; Figure 2 It is the basic principle diagram of howling detection of an FPGA-based dynamic howling suppression optimization design method provided by an embodiment of the present application; Figure 3 It is the schematic diagram of the optimized howling detection algorithm of an FPGA-based dynamic howling suppression optimization design method provided by an embodiment of the present application; Figure 4 It is the schematic diagram of the acoustic feedback path parameter estimation based on FPGA of an FPGA-based dynamic howling suppression optimization design method provided by an embodiment of the present application; Detailed Embodiments

[0015] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments. The embodiments of the present invention are given for the purpose of illustration and description, and are not exhaustive or limit the present invention to the disclosed form. Many modifications and variations are obvious to those of ordinary skill in the art. The embodiments are selected and described to better illustrate the principles and practical applications of the present invention, and enable those of ordinary skill in the art to understand the present invention and thus design various embodiments with various modifications suitable for specific purposes. Embodiment

[0016] Please refer to Figures 1 to 4 In this embodiment, an FPGA-based dynamic howling suppression optimization design method is provided, including The principle of the real-time dynamic howling suppression algorithm is divided into two stages. In the first stage, it detects whether there is a howling phenomenon in the acoustic feedback. When howling is detected, the speaker output signal is replaced with a short-time Gaussian white noise signal. In the second stage, in the case of playing Gaussian white noise, the parameters of the acoustic feedback signal model are estimated, and the information of the speaker output sound signal after passing through the acoustic feedback channel parameters is subtracted from the signal collected by the microphone, obtaining a relatively pure original sound signal collected by the microphone, which plays the role of eliminating howling.

[0017] It can be seen from Figure 1 that this solution includes the following steps: S1. The microphone collects the audio signal, and the analog audio is converted into a digital audio signal through the audio acquisition module; the digital audio signal is input into the FPGA chip, and the FPGA realizes the functions of digital audio signal processing and real-time dynamic howling suppression algorithm. S2. The input audio signal is subjected to gain processing, and then howling detection is performed. When howling is detected, Gaussian white noise is switched and output through the audio output control module. Otherwise, the received audio signal is output.

[0018] S3. When the speaker outputs Gaussian white noise, the microphone collects the signal after the white noise passes through the acoustic feedback path. At this time, the acoustic feedback path estimation calculation is performed to obtain the acoustic feedback path parameters, and the output is switched to the normal acoustic feedback audio signal through the audio output control module. S4. Howling suppression processing is performed according to the estimated acoustic feedback path parameters to obtain a relatively clean original audio signal, which has a good howling suppression effect.

[0019] Howling detection is the key to the dynamic howling suppression algorithm, which determines the howling suppression effect of the algorithm. The missed detection of howling will cause the howling not to be suppressed in a timely and effective manner, while the false detection will filter out the frequency components of the actual speech due to incorrect suppression, resulting in the distortion of the speech signal and affecting the performance of the sound reinforcement system. The basic principle block diagram of the commonly used howling detection is as shown in Figure 2 shown; It can be seen from Figure 2 that the basic implementation process of howling detection is as follows: First, the input signal is framed and windowed, and the signal power spectrum is obtained through the short-time Fourier transform. Secondly, the candidate howling frequency points are determined. Since the howling component is a single-frequency narrowband signal with continuously increasing energy, there is relatively high energy at the howling frequency points. Therefore, usually several frequency points with larger signal power amplitudes in each frame are selected as the candidate howling frequency points. Then, the detection decision is made according to the set detection threshold T.

[0020] Common howling detection uses methods such as the Peak-to-Harmonic Power Ratio (PHPR) detection algorithm and the Peak-to-Neighboring Power Ratio (PNPR) detection. However, in an environment with a relatively complex actual sound field or when the target signal is audio, if a single detection feature is used, the effect is not ideal enough.

[0021] In order to accurately estimate the acoustic feedback channel model parameters, the real-time dynamic howling suppression algorithm assumes that the acoustic feedback path model is a function with the FIR filtering characteristic; when howling is detected, the output signal of the speaker is a Gaussian white noise signal with a short-time mean of zero, and the feedback signal is used to estimate the acoustic feedback path parameters.

[0022] When the speaker outputs a Gaussian white noise signal with a short-time zero mean, it can be considered that the output Gaussian white noise and the pure speech signal input by the mic are independent of each other. At this time, the signal collected by the microphone can be expressed as: Equation 1 In the formula represents the FIR model coefficients of the acoustic feedback path, vector , vector represents the acoustic feedback signal received by the microphone after the output signal of the speaker passes through the feedback path with P taps; P is the number of paths of the acoustic feedback path; is the pure speech signal; At this time, is the acoustic feedback Gaussian white noise signal received by the microphone after the Gaussian white noise passes through the acoustic feedback path, that is ; Therefore, and are independent of each other; The actual acoustic feedback signal received by the microphone is related to f(n) for calculation, and we can get: Equation 2 For the settings and operation methods, unless otherwise specifically stated and limited, they are all implemented according to the conventional means in this field; From Equation 2, we can obtain the value of Equation 3 Among them, =0. is autocorrelation. is cross-correlation with f(n).

[0023] Optimized Design of Dynamic Squeal Suppression Algorithm Based on FPGA - Squeal Detection Optimized Design Scheme The performance of the squeal detection algorithm determines the ultimate acoustic feedback suppression effect of the squeal suppression algorithm. Commonly used feature - based squeal detection algorithms generally have a high false alarm rate, which easily causes distortion of the output speech. To effectively control the false alarm rate, this paper proposes an improved squeal detection algorithm, which combines the squeal detection algorithm based on the inter - frame spectral flatness deviation with the peak harmonic power ratio (PHPR) and peak adjacent power ratio (PNPR) for squeal detection optimization.

[0024] Analysis shows that the squeal component is a single - frequency narrow - band signal with energy increasing over time at the squeal frequency point, and the power spectrum at the squeal frequency point has the same variation characteristics. Usually, the power spectrum of the speech signal has greater non - stationary characteristics. Therefore, this paper introduces the feature method based on the inter - frame spectral flatness deviation (Interframe Spectral Flatness Deviation, ISFD) into the squeal detection algorithm, and uses the difference in the power spectrum transformation characteristics between speech and squeal to distinguish speech and squeal and improve the performance of squeal detection.

[0025] Feature Method of Inter - frame Spectral Flatness Deviation The inter - frame spectral flatness deviation, also known as the long - term spectral flatness measure (LSFM) feature, was first proposed by Yanna et al. This feature is a measure of the power spectrum flatness characteristics of a long - time signal. For the squeal component being a narrow - band signal, it can be optimized into a squeal detection algorithm based on the inter - frame spectral flatness deviation feature. The specific calculation process is as follows: First, calculate the signal power spectrum and select the frequency points with larger power spectra as candidate squeal frequency points.

[0026] Secondly, calculate the geometric mean and the average of the power spectra of the past R frames of the signal at the candidate squeal frequency points .

[0027] Equation 4 Equation 5 Then, calculate the LSFM feature parameter at the frequency according to and ; Equation 6 Finally, calculate the past Q frames The variance of the parameters to obtain the ISFD feature parameters .

[0028] Equation 7 Through the above analysis, the speech signal is a non-stationary signal, and the power spectrum at the frequency changes continuously with time. Then the LSFM feature parameters at the frequency have large fluctuations. Therefore, the ISFD feature parameters of the speech signal take large values; while the power spectrum of the howling signal has the same transformation characteristics at the howling frequency point and has the same spectral flatness characteristics. The LSFM feature parameters of the howling signal at the howling frequency point have small fluctuations. Therefore, the ISFD feature parameters of the howling signal take small values; thus, it can be seen that the ISFD feature parameters can be used for howling detection to distinguish speech and howling.

[0029] Howling detection optimization algorithm In the actual application of the on-site sound reinforcement system, the sound field is relatively complex and the target signal is an audio signal. If a single detection feature is used, the effect is not ideal; in order to improve the detection accuracy as much as possible and minimize the damage to the sound quality, this paper adopts two detection methods, the peak harmonic power ratio (PHPR) and the peak adjacent power ratio (PNPR), and combines the howling detection algorithm based on the frame-to-frame spectral flatness deviation feature as an improved howling detection optimization algorithm; The peak harmonic power ratio (PHPR) detection algorithm, and its expression is as follows: Equation 8 Equation 9 In the formula, represents the power at the integer multiple (harmonic) frequency of the frequency , is the power at the frequency . When the ratio of the power at the frequency to the power at the integer multiple (harmonic) frequency of the frequency exceeds the detection threshold (TH), then it is judged that is the howling frequency; The peak adjacent power ratio (PNPR) detection algorithm, and its expression is as follows: Equation 10 Equation 11 In the formula, represents the power of the adjacent frequency component of the frequency , is the power at the frequency . When the power at the frequency to the power at the frequency When the ratio of the power at adjacent frequencies exceeds the detection threshold (TN), it is determined that the howling frequency; The basic original diagram of its algorithm is as Figure 3 shown.

[0030] Implementation steps of the howling detection optimization algorithm: Perform VAD detection on the input signal. When the input signal is a silent signal, it is determined that no howling has occurred. Otherwise, perform howling detection processing on the signal.

[0031] Frame and window the input signal, and obtain the signal power spectrum through discrete Fourier transform.

[0032] Determine the candidate howling frequency points. Since the energy at the howling frequency points is relatively large when howling occurs, generally select the frequency points with larger power as the candidate howling frequency points. Detect the frequency point with the maximum power as the basic candidate howling frequency point for PHPR and PNPR, and then detect M frequency points with larger power as the candidate howling frequency points for ISFD.

[0033] Calculate the ISFD eigenvalue, calculate the spectral flatness at each candidate howling frequency point, and obtain the ISFD eigenvalue of the current frame by calculating the variance based on the spectral flatness parameters of the past continuous Q frames at each candidate howling frequency point.

[0034] Calculate the peak harmonic energy ratio (PHPR) and the peak adjacent power ratio (PNPR).

[0035] Perform threshold decision respectively. When the peak harmonic energy ratio (PHPR) is greater than or equal to the threshold TH, it is determined that howling has occurred; when the peak adjacent power ratio (PNPR) is greater than or equal to the threshold TN, it is determined that howling has occurred; when the ISFD eigenvalue is less than the detection threshold TS, it is determined that howling has occurred.

[0036] Finally, analyze the howling decision result. Combining the results of howling generation from the above three decisions, when the peak adjacent power ratio (PNPR) generates howling or the peak harmonic energy ratio (PHPR) determines howling generation, and when the ISFD eigenvalue determines howling generation, it is finally determined that howling has occurred, otherwise it is determined that no howling has occurred.

[0037] Optimized design method for acoustic feedback path estimation According to the principle of acoustic feedback path estimation, the acoustic feedback signal actually received by the microphone and The relevant calculation is represented by formula 2, where =0, formula 2 can be simplified to formula 12 Formula 12 When the speaker output is Gaussian white noise, is Related to Related calculations; Then, Equation 12 can be expanded as follows: = Equation 13 The left side of Equation 13 is a matrix of related functions, symmetric about the diagonal, and all elements on the main diagonal and any diagonal line parallel to the main diagonal are equal; such a matrix is called a Toeplitz matrix, and such an equation is called a Yule-Walker equation; this matrix equation does not require a large amount of calculations like solving a general matrix equation. Using the properties of the Toeplitz matrix, an efficient method for solving this equation can be obtained, and the Levinson-Durbin recurrence algorithm is used for solving; The Levinson-Durbin recurrence algorithm needs to calculate autocorrelation and cross-correlation operations during the calculation process. The computational complexity of time-domain cross-correlation operations is extremely large. In this paper, the FFT algorithm is used to accelerate cross-correlation calculations, and the Levinson-Durbin recurrence algorithm is implemented in parallel based on FPGA; The schematic diagram of acoustic feedback path parameter estimation based on FPGA is as Figure 4 shown.

[0038] In this paper, by utilizing the highly parallel and reconfigurable characteristics of FPGA, the efficient acceleration of the calculation of the Levinson-Durbin recurrence algorithm is realized; The algorithm calculation is implemented with a 100MHz clock, the 2048-point FFT algorithm is used to parallelize and optimize the cross-correlation calculation process, which takes about 3ms, and the recurrence algorithm is implemented with a parallel architecture, and the implementation time of the algorithm is about 3.3ms.

[0039] Due to the particularity and complexity of the acoustic feedback path in the on-site sound reinforcement system, extremely high requirements are imposed on the performance of the adaptive algorithm and the implementation device. In actual applications, the volume switch is often adjusted, and the received signal power is variable, requiring a high convergence speed.

[0040] In view of the characteristics of high real-time performance and fast convergence rate of howling suppression in on-site meetings, this paper proposes a real-time dynamic howling suppression algorithm. This algorithm can achieve real-time processing, significantly improve the system gain, bring less sound distortion, and has a low hardware cost. This paper proposes an optimized howling detection method that combines a howling detection algorithm based on the deviation of spectral flatness between frames, a Peak-to-Harmonic Power Ratio (PHPR) detection algorithm, and a Peak-to-Neighboring Power Ratio (PNPR) detection method, effectively controlling the false alarm rate and thus reducing the speech distortion rate. At the same time, this paper uses the Levinson-Durbin recursive algorithm to calculate the acoustic feedback path parameters and implements the FFT algorithm in parallel based on FPGA to accelerate the cross-correlation calculation. While ensuring the accuracy of the calculated data, it reduces resources and optimizes the calculation difficulty. Through a large number of simulation tests and actual sound field tests, it is verified that the howling suppression scheme in this paper can process signals in real time, has a good howling suppression effect, and can obtain good sound quality. This algorithm can be applied to public sound reinforcement systems, such as conference rooms, multimedia classrooms, and lecture halls, to design a howling suppressor that can achieve real-time processing.

[0041] Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art and related fields without creative efforts shall fall within the scope of protection of the present invention. The structures, devices, and operation methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art without special instructions and limitations.

Claims

1. An optimized design method for dynamic howling suppression based on FPGA, characterized in that, It includes the following steps: S1. Signal acquisition and preprocessing: Collect the audio signal through a microphone, convert it into a digital signal by the audio acquisition module, and then input it into the FPGA chip; S2. Gain processing and howling detection: Adjust the gain of the input signal, and then perform howling detection; when howling is detected, switch the output to Gaussian white noise through the audio output control module; When no howling is detected, output the received audio signal; Adopt an improved howling detection algorithm for real-time detection; S3. Noise injection and path estimation: When howling is detected, switch the output to Gaussian white noise and estimate the acoustic feedback path parameters; When the speaker outputs Gaussian white noise, the microphone collects the signal after the white noise passes through the acoustic feedback path. At this time, perform acoustic feedback path estimation calculation to obtain the acoustic feedback path parameters, and switch the output to the normal acoustic feedback audio signal through the audio output control module; S4. Howling suppression processing: Filter the original signal according to the estimated acoustic feedback path parameters and output the suppressed audio signal.

2. The optimized design method for dynamic howling suppression based on FPGA according to claim 1, characterized in that The process of the howling detection includes: SS1. Frame the input signal and apply a window, and obtain the signal power spectrum through short-time Fourier transform; SS2. Determine the candidate howling frequency points. Since the howling component is a single-frequency narrowband signal with continuously increasing energy, there is higher energy at the howling frequency points. Therefore, usually select several frequency points with larger signal power amplitudes per frame as the candidate howling frequency points; SS3. Then perform detection and decision according to the set detection threshold T.

3. A dynamic howling suppression optimization design method based on FPGA according to claim 1, characterized in that, The estimation method of the acoustic feedback path parameters: When howling is detected, the signal output by the speaker is Gaussian white noise with a short-time mean of zero, and use the feedback signal to estimate the acoustic feedback path parameters; When the speaker outputs Gaussian white noise with a short-time zero mean, it can be considered that the output Gaussian white noise is independent of the pure speech signal input by the mic. At this time, the signal collected by the microphone can be expressed as: Formula 1 where represents the FIR model coefficients of the acoustic feedback path; vector , the vector represents the acoustic feedback signal received by the microphone after the speaker output signal passes through a feedback path with P taps; P is the number of paths of the acoustic feedback path; is a pure voice signal; At this time, is the acoustic feedback Gaussian white noise signal received by the microphone after the Gaussian white noise passes through the acoustic feedback path, that is, ; Therefore, and are independent of each other; The actual sound feedback signal received by the microphone Through calculations related to f(n), we can obtain: Formula 2 Can be obtained from Equation 2 value Formula 3 Among them, = 0, is autocorrelation; is cross-correlated with f(n).

4. A method for optimizing the design of dynamic howling suppression based on FPGA according to claim 1, characterized in that It also includes a howling detection algorithm based on the frame-to-frame spectral flatness deviation feature, and its specific calculation process is as follows: First, calculate the signal power spectrum and select the frequency points with larger power spectra as candidate whistling frequency points; Secondly, calculate the geometric mean and the average of the power spectra of the past R-frame signals at the candidate whistling frequency points ; and the average ; Formula 4 Formula 5 Then, based on and calculate the LSFM feature parameters of the signal at the frequency ; ; Formula 6 Finally, calculate the variance of the parameters of the past Q frames to obtain the ISFD feature parameters ; Formula 7 Through the above analysis, the speech signal is a non-stationary signal, and the power spectrum at the frequency changes continuously with time. Therefore, the LSFM characteristic parameters at the frequency have large fluctuations; thus, the ISFD characteristic parameters of the speech signal have large values; while the power spectrum of the howling signal has the same transformation characteristics at the howling frequency point and has the same spectral flatness characteristics, and the LSFM characteristic parameters of the howling signal at the howling frequency point have small fluctuations; therefore, the ISFD characteristic parameters of the howling signal have small values; it can be seen that the ISFD characteristic parameters can be used for howling detection to distinguish speech from howling.

5. The optimized design method for dynamic howling suppression based on FPGA according to claim 4, characterized in that This solution adopts two detection methods: peak harmonic power ratio (PHPR) and peak adjacent power ratio (PNPR), and combines the howling detection algorithm based on the frame-to-frame spectral flatness deviation feature as an improved howling detection optimization algorithm.

6. The optimized design method for dynamic howling suppression based on FPGA according to claim 5, characterized in that The expression of the peak harmonic power ratio (PHPR) detection algorithm is as follows: Formula 8 Formula 9 In the formula, represents the frequency the power at an integer multiple (harmonic) frequency, is the power at frequency When the ratio of the power at frequency to the power at an integer multiple (harmonic) frequency of frequency exceeds the detection threshold (TH), then it is determined that is the howling frequency.

7. A method for optimizing the design of dynamic howling suppression based on FPGA according to claim 6, characterized in that, The expression of the peak adjacent power ratio (PNPR) detection algorithm is as follows: Formula 10 Formula 11 In the formula, represents the frequency the power of the adjacent frequency component, is the power at the frequency When the ratio of the power at the frequency to the power at the adjacent frequency exceeds the detection threshold (TN), then it is judged that is the howling frequency.

8. A method for optimizing the design of dynamic howling suppression based on FPGA according to claim 7, characterized in that, The implementation steps of the howling detection optimization algorithm: Step 1: Perform VAD detection on the input signal. When the input signal is a silent signal, it is determined that no howling has occurred. Otherwise, perform howling detection processing on the signal; Step 2: Frame and window the input signal, and obtain the signal power spectrum through discrete Fourier transform; Step 3: Determine the candidate howling frequency points; since the energy at the howling frequency points is large when howling occurs, generally select the frequency points with larger power as the candidate howling frequency points; detect the frequency point with the maximum power as the basic candidate howling frequency point for PHPR and PNPR, and then detect M larger power frequency points as the candidate howling frequency points for ISFD; Step 4: Calculate the ISFD eigenvalue, calculate the spectral flatness at each candidate whistling frequency point, and obtain the ISFD eigenvalue of the current frame by calculating the variance based on the spectral flatness parameters of the past consecutive Q frames at each candidate whistling frequency point; Step 5: Calculate the peak harmonic power ratio (PHPR) and the peak adjacent power ratio (PNPR); Step 6: Conduct threshold judgments respectively. When the peak harmonic power ratio (PHPR) is greater than or equal to the threshold TH, it is determined that whistling occurs; when the peak adjacent power ratio (PNPR) is greater than or equal to the threshold TN, it is determined that whistling occurs; when the ISFD eigenvalue is less than the detection threshold TS, it is determined that whistling occurs; Step 7: Finally, analyze the whistling judgment result. Combining the results of the above three judgments of whistling occurrence, when the peak adjacent power ratio (PNPR) determines whistling or the peak harmonic power ratio (PHPR) determines whistling, and when the ISFD eigenvalue determines whistling, it is finally determined that whistling occurs; otherwise, it is determined that no whistling occurs.

9. A method for optimizing the design of dynamic howling suppression based on FPGA according to claim 3, characterized in that, The optimization design method for acoustic feedback path estimation is as follows: According to the principle of acoustic feedback path estimation, the acoustic feedback signal actually received by the microphone and The related calculation is represented by Formula 2, where = 0, Formula 2 can be simplified to Formula 12 Formula 12 When the speaker output is Gaussian white noise, For Related to Related calculations; Then, Equation 12 can be expanded as follows: = Formula 13 The left side of Equation 13 is a Toeplitz matrix. This kind of equation is called the Yule-Walker equation and is solved using the Levinson-Durbin recursive algorithm.

Images