A microphone anti-howling system

Abstract

The application relates to the technical field of electroacoustic processing, and discloses a microphone anti-howling system. The system continuously collects an audio input stream containing time-domain waveform data and background acoustic parameters through an audio capture module; a dynamic range compression module processes the audio stream to generate a standardized audio frame sequence; a feature extraction module extracts sound pressure space-time variation features including a sound pressure gradient field, a sound energy density distribution and a phase coherence atlas from the sequence; a howling identification network in a howling identification module outputs a howling occurrence probability and hotspot area coordinates based on the features; a probability correction module performs real-time correction on the probability according to environmental acoustic parameters; and an intervention scheme construction module generates a targeted acoustic intervention scheme according to the corrected probability and the hotspot area coordinates. The application realizes accurate identification and effective suppression of the howling phenomenon, and significantly improves the stability and sound quality performance of an audio system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electro-acoustic processing, in particular to a microphone anti-howling system. BACKGROUND

[0002] The current microphone anti-howling mainly adopts the fixed frequency notch or global gain control method. The existing technology is based on single dimension analysis, and the time domain waveform and frequency domain characteristics cannot be effectively combined. The feature extraction method is simple, and the sound pressure space-time variation characteristics are not fully explored. The howling recognition mechanism is solidified, and it is mainly dependent on threshold judgment, which cannot adapt to complex acoustic environment. The probability correction is missing, and the influence of environmental factors on howling is not quantitatively considered. The intervention scheme generation is empirical, and lacks targeted spatial positioning processing. The existing method needs to solve the key technical problems of multi-feature fusion, intelligent recognition, environmental adaptation and precise intervention.

[0003] The traditional anti-howling system has obvious shortcomings in accuracy and adaptability. The dynamic range of audio acquisition is limited, and loud sound pressure level signals are prone to distortion. The compression algorithm parameters are fixed, which cannot adapt to the acoustic characteristics of different scenes. The time-space feature extraction dimension is single, and the sound field variation law is not complete. The recognition network training data is insufficient, and the generalization ability in complex scenes is limited. The environmental parameter monitoring is not comprehensive, and the probability correction accuracy is low. The hotspot area positioning is rough, and the spatial resolution is insufficient. The intervention scheme library size is limited, and the matching degree of the scheme with the actual demand is low. SUMMARY

[0004] The purpose of the present application is to provide a microphone anti-howling system to solve the problems raised in the background art.

[0005] To achieve the above purpose, the present application provides a microphone anti-howling system, which comprises:

[0006] An audio capture module for continuously capturing the audio input stream of the target microphone in the operating scene, the audio input stream containing time domain waveform data and background acoustic parameters;

[0007] A dynamic range compression module for performing dynamic range compression processing on the audio input stream to generate a standardized audio frame sequence;

[0008] A feature extraction module for extracting sound pressure space-time variation characteristics from the standardized audio frame sequence, the sound pressure space-time variation characteristics including sound pressure gradient field, sound energy density distribution and phase coherence spectrum;

[0009] A howling recognition module including a howling recognition network for feeding the sound pressure space-time variation characteristics into the howling recognition network to output the howling occurrence probability and the hotspot area coordinates;

[0010] a probability correction module, configured to correct the howling occurrence probability in real time according to an environmental acoustic parameter, and generate a corrected howling occurrence probability;

[0011] an intervention scheme construction module, configured to construct an acoustic intervention scheme according to the corrected howling occurrence probability and the hotspot region coordinates.

[0012] Preferably, the dynamic range compression process is specifically implemented as follows:

[0013] monitoring a peak level and an average level of the audio input stream, and calculating a transient dynamic range coefficient;

[0014] adjusting a gain control curve based on the transient dynamic range coefficient, the gain control curve being defined by a logarithmic amplifier and a limiter together;

[0015] applying the gain control curve to the audio input stream for amplitude shaping, so as to eliminate overload distortion components;

[0016] performing frequency band equalization on the amplitude-shaped signal, so as to improve mid-frequency intelligibility; and outputting a standardized audio frame sequence with uniform energy distribution.

[0017] Preferably, the process of extracting the sound pressure spatiotemporal variation feature includes:

[0018] dividing the standardized audio frame sequence into overlapping time windows, each time window corresponding to a sound pressure sampling period; and for each time window, calculating a sound pressure gradient field, the sound pressure gradient field being obtained through three-dimensional spatial difference operation and reflecting a rate of change of sound pressure with position;

[0019] deriving a sound energy density distribution, the sound energy density distribution being generated by convolution of sound pressure square and medium characteristics;

[0020] measuring a phase coherence map, the phase coherence map being obtained through cross-correlation analysis of phase differences of different microphone units;

[0021] aggregating the sound pressure gradient field, the sound energy density distribution and the phase coherence map of consecutive time windows to form a spatiotemporal fusion feature vector.

[0022] Preferably, the howling recognition network operates in the following manner:

[0023] inputting the sound pressure spatiotemporal variation feature vector into a convolutional neural network layer, the convolutional neural network layer learning local patterns of the sound pressure gradient field; outputting the feature to a recurrent neural network layer, the recurrent neural network layer tracking temporal evolution of the sound energy density distribution; and transferring a hidden state of the recurrent neural network layer to a fully connected layer, the fully connected layer calculating an anomaly score of the phase coherence map;

[0024] The local mode, time evolution and anomaly score are fused to generate a howling occurrence probability through a softmax function; meanwhile, a clustering algorithm is used to identify high-density areas in the feature map to output hotspot area coordinates.

[0025] Preferably, the step of identifying high-density areas in the feature map using a clustering algorithm to output hotspot area coordinates comprises:

[0026] extracting spatial coordinates and feature intensity values of all feature points from the feature map;

[0027] calculating the distance between each feature point and adjacent feature points to generate a distance matrix;

[0028] applying a density clustering algorithm based on the distance matrix to divide the feature points into multiple clusters;

[0029] identifying the feature point density of each cluster, and selecting clusters with feature point density exceeding a preset density threshold as high-density areas;

[0030] calculating the geometric center coordinates of each high-density area to output the hotspot area coordinates.

[0031] Preferably, the step of real-time correcting the howling occurrence probability according to environmental acoustic parameters comprises:

[0032] collecting environmental acoustic parameters, including environmental noise spectrum and reverberation time;

[0033] calculating the deviation of the environmental noise spectrum from the reference noise spectrum to generate a noise correction factor;

[0034] adjusting the sensitivity threshold of the howling occurrence probability according to the reverberation time, wherein the sensitivity threshold is inversely proportional to the reverberation time;

[0035] weighting and averaging the howling occurrence probability using the noise correction factor and the sensitivity threshold to output the corrected howling occurrence probability.

[0036] Preferably, the step of adjusting the sensitivity threshold of the howling occurrence probability according to the reverberation time comprises:

[0037] measuring the reverberation time value of the current environment to obtain a mapping relationship table of predefined reverberation time and sensitivity threshold;

[0038] querying the corresponding reference sensitivity threshold in the mapping relationship table according to the reverberation time value;

[0039] calculating the deviation proportion of the reverberation time value from the standard reverberation time, and linearly scaling the reference sensitivity threshold based on the deviation proportion;

[0040] scaling the value as the adjusted sensitivity threshold.

[0041] Preferably, the method of constructing the acoustic intervention scheme comprises:

[0042] Resolving the spatial relationship of the hotspot area coordinates to determine the boundary of the acoustic interference area;

[0043] Selecting an intervention intensity level based on the corrected howling occurrence probability, the intervention intensity level being divided into three grades of low, medium and high;

[0044] Designing a frequency response adjustment strategy for each intervention intensity level, the frequency response adjustment strategy involving notch filter settings and gain attenuation curves;

[0045] Generating beamforming parameters for the acoustic interference area, the beamforming parameters including steering vectors and null depth;

[0046] Combining the intervention intensity level, the frequency response adjustment strategy and the beamforming parameters to form a complete acoustic intervention scheme.

[0047] Preferably, the implementation of the frequency response adjustment strategy comprises:

[0048] Determining the center frequency and bandwidth of the notch filter according to the intervention intensity level, the center frequency of the notch filter corresponding to the dominant frequency of the howling; setting the starting point and slope of the gain attenuation curve, the gain attenuation curve ensuring smooth transition to avoid auditory abruptness;

[0049] Verifying the compatibility of the notch filter with the system phase response and adjusting the group delay parameter;

[0050] Applying an adaptive algorithm to dynamically update the notch filter coefficients and output an optimized frequency response configuration file.

[0051] Preferably, the determination of the center frequency and bandwidth of the notch filter according to the intervention intensity level comprises:

[0052] Predefining a correspondence between the intervention intensity level and the notch filter parameters, wherein each intervention intensity level is associated with a set of center frequency and bandwidth values;

[0053] Retrieving the corresponding center frequency and bandwidth values from the correspondence according to the currently selected intervention intensity level;

[0054] Setting the resonance point of the notch filter based on the retrieved center frequency;

[0055] Setting the attenuation range of the notch filter based on the retrieved bandwidth value.

[0056] Compared with the prior art, the present application has the following beneficial effects:

[0057] A standardized audio frame sequence is generated by continuously capturing a microphone audio input stream and performing dynamic range compression processing. The dynamic range compression uses an adaptive algorithm to process different sound pressure level signals, and the standardization processing unifies the data format. Sound pressure spatiotemporal variation features are extracted from the standardized audio frame sequence, including a sound pressure gradient field, a sound energy density distribution, and a phase coherence atlas. The sound pressure gradient field represents the sound pressure spatial variation rate, the sound energy density distribution reflects the energy aggregation characteristics, and the phase coherence atlas detects the sound wave interference phenomenon. Multi-dimensional feature extraction comprehensively characterizes the sound field dynamic variation law.

[0058] The sound pressure spatiotemporal variation features are input into a howling identification network, and the howling occurrence probability and hotspot area coordinates are output. The identification network is based on a deep learning architecture and obtains identification ability through a large number of samples. The probability calculation considers the nonlinear relationship between the features, and the hotspot coordinates are accurately positioned through a sound field reconstruction algorithm. The howling occurrence probability is corrected in real time according to environmental acoustic parameters, including reverberation time, background noise level, and other indicators. The correction algorithm establishes a mapping relationship between the environment and the probability, and uses a Bayesian update method to fuse real-time data.

[0059] An acoustic intervention scheme is constructed according to the corrected howling occurrence probability and hotspot area coordinates. The intervention scheme is designed for the hotspot area and includes measures such as directional equalization and beamforming. The scheme parameters are set according to the probability level, balancing the processing effect and signal quality. Through accurate intervention, the howling is effectively suppressed while maintaining the quality of the voice. The scheme execution process is monitored in real time, and the parameters are dynamically adjusted according to the effect to achieve closed-loop optimization.

[0060] The system realizes effective prevention and control of howling through the synergistic effect of dynamic range compression, multi-dimensional feature extraction, intelligent identification, environmental correction, and accurate intervention. Dynamic compression ensures signal quality, feature extraction provides comprehensive sound field information, intelligent identification realizes early warning, environmental correction enhances adaptability, and accurate intervention ensures suppression effect. The system significantly improves the microphone's ability to resist howling and the quality of the voice. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 A working principle diagram of the microphone anti-howling system described in the present application;

[0062] Figure 2 A principle diagram for dynamic range compression processing;

[0063] Figure 3 A principle diagram for the operation of the howling identification network;

[0064] Figure 4 A schematic diagram of the hotspot area clustering distribution of the howling identification module;

[0065] Figure 5 A comparison diagram of the environmental correction effect of the howling occurrence probability in different acoustic environments. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the present application.

[0067] Please refer to Figure 1 The present application provides a microphone anti-howling system, which comprises an audio capture module, a dynamic range compression module, a feature extraction module, a howling identification module, a probability correction module and an intervention scheme construction module. The specific implementation is as follows:

[0068] The audio capture module is used for continuously capturing an audio input stream of a target microphone in an operation scene, and the audio input stream contains time-domain waveform data and background acoustic parameters. The dynamic range compression module is used for performing dynamic range compression processing on the audio input stream to generate a standardized audio frame sequence. The feature extraction module is used for extracting sound pressure spatio-temporal variation features from the standardized audio frame sequence, and the sound pressure spatio-temporal variation features include sound pressure gradient field, sound energy density distribution and phase coherence atlas. The howling identification module comprises a howling identification network, which is used for feeding the sound pressure spatio-temporal variation features into the howling identification network to output howling occurrence probability and hotspot area coordinates. The probability correction module is used for correcting the howling occurrence probability in real time according to environmental acoustic parameters to generate corrected howling occurrence probability. The intervention scheme construction module is used for constructing an acoustic intervention scheme according to the corrected howling occurrence probability and the hotspot area coordinates.

[0069] Embodiment 1: Please refer to Figure 2, the specific implementation of the dynamic range compression processing is monitoring the peak level and the average level of the audio input stream, calculating the instantaneous dynamic range coefficient, adjusting the gain control curve based on the instantaneous dynamic range coefficient, the gain control curve being defined by a logarithmic amplifier and a limiter, applying the gain control curve to the audio input stream for amplitude shaping, eliminating overload distortion components, performing band equalization processing on the amplitude-shaped signal to improve mid-frequency intelligibility, and outputting a standardized audio frame sequence with uniform energy distribution. The process of extracting the sound pressure spatiotemporal variation feature includes dividing the standardized audio frame sequence into overlapping time windows, each time window corresponding to a sound pressure sampling period, for each time window, calculating a sound pressure gradient field, the sound pressure gradient field being obtained through a three-dimensional spatial difference operation, reflecting the rate of change of sound pressure with position, deriving a sound energy density distribution, the sound energy density distribution being generated by convolution of sound pressure square and medium characteristics, measuring a phase coherence map, the phase coherence map being obtained by cross-correlation analysis of the phase difference of different microphone units, and aggregating the sound pressure gradient field, the sound energy density distribution, and the phase coherence map of consecutive time windows to form a spatiotemporal fusion feature vector.

[0070] In specific implementations, the dynamic range compression processing is started by monitoring the peak level and the average level of the audio input stream, the peak level and the average level being provided in real time by the audio capture module, the process of calculating the instantaneous dynamic range coefficient involves normalizing the difference between the peak level and the average level, the gain control curve is adjusted based on the instantaneous dynamic range coefficient, the gain control curve being defined by a logarithmic amplifier and a limiter, the logarithmic amplifier being responsible for compressing the input signal, and the limiter preventing signal overshoot, the gain control curve is applied to the audio input stream for amplitude shaping, eliminating overload distortion components, amplitude shaping being achieved through dynamic gain adjustment, ensuring that the signal amplitude is within a preset range, band equalization processing is performed on the amplitude-shaped signal to improve mid-frequency intelligibility, the band equalization processing uses a digital filter bank to enhance specific frequency components, and a standardized audio frame sequence with uniform energy distribution is output, the standardized audio frame sequence providing stable input for subsequent feature extraction. The medium characteristics specifically refer to the air density p and the sound speed c that sound relies on when propagating in air, these parameters are used as known constants in the calculation of the sound energy density distribution. The band equalization processing is implemented using a set of parallel digital bandpass filters, the filter bank covering the main frequency range of human speech, and the mid-frequency intelligibility is enhanced by increasing the gain of the filters within the 300Hz to 3400Hz frequency band.

[0071] In some embodiments, the calculation of the instantaneous dynamic range coefficient can be implemented using the formula: wherein: represents the instantaneous dynamic range coefficient, represents the peak level of the audio input stream, represents the average level of the audio input stream, wherein R denotes a preset reference level value, and the formula ensures that the coefficient reflects the dynamic change of the signal. Optionally, the adjustment of the gain control curve can be realized by a lookup table manner, which maps the instantaneous dynamic range coefficient to the corresponding gain value.

[0072] The process of extracting the spatial-temporal variation features of sound pressure starts from dividing the sequence of normalized audio frames into overlapping time windows, each corresponding to a sound pressure sampling period. The overlapping time windows are set based on a fixed time step. For each time window, the sound pressure gradient field is calculated, which is obtained through a three-dimensional spatial difference operation involving the calculation of partial derivatives of sound pressure values in the spatial dimension, reflecting the rate of change of sound pressure with position. The sound energy density distribution is derived, which is generated by convolving the square of sound pressure with medium characteristics, including air density and sound speed parameters. The phase coherence map is measured, which is obtained by cross-correlation analysis of the phase difference between different microphone units. The specific steps include: performing Fourier transform on the synchronous audio signals from different microphone units to obtain their respective phase spectra; for each frequency component, calculating the phase difference between specified pairs of microphone units; by analyzing the consistency of the phase difference at multiple frequency components or the correlation of the phase difference between multiple pairs of microphone units, the phase coherence map is constructed to quantify the spatial coherence degree of the sound field. The cross-correlation analysis calculates the phase consistency between multi-channel signals, and the sound pressure gradient field, sound energy density distribution, and phase coherence map of consecutive time windows are aggregated to form a spatio-temporal fusion feature vector, which is used as the input of the howling recognition network.

[0073] In some embodiments, the division of overlapping time windows can adjust the window length and overlap ratio to adapt to different scenarios. It can be understood that the calculation of the sound pressure gradient field needs to ensure that the spatial resolution is high enough to capture subtle changes. Optionally, the measurement of the phase coherence map can use fast Fourier transform to optimize computational efficiency. It can be understood that the aggregation operation of the spatio-temporal fusion feature vector is realized by splicing or weighted averaging.

[0074] Embodiment 2: refer to Figure 3The operation mode of the howling identification network is that the sound pressure space-time change feature vector is input into a convolutional neural network layer, the convolutional neural network layer learns the local mode of the sound pressure gradient field, the output feature map is mapped to a recurrent neural network layer, the recurrent neural network layer tracks the time evolution of the sound energy density distribution, the hidden state of the recurrent neural network layer is transmitted to a fully connected layer, the fully connected layer calculates the anomaly score of the phase coherence map, the local mode, the time evolution and the anomaly score are fused, the howling occurrence probability is generated through a softmax function, at the same time, a clustering algorithm is used to identify the high-density area in the feature map, and the hotspot area coordinates are output. The method of identifying the high-density area in the feature map by using the clustering algorithm and outputting the hotspot area coordinates comprises the following steps: extracting the spatial coordinates and feature intensity values of all feature points from the feature map, calculating the distance between each feature point and the adjacent feature points, generating a distance matrix, applying a density clustering algorithm based on the distance matrix, dividing the feature points into multiple clusters, identifying the feature point density of each cluster, selecting the cluster whose feature point density exceeds the preset density threshold as the high-density area, calculating the geometric center coordinates of each high-density area, and outputting the hotspot area coordinates.

[0075] In a specific implementation, the operation of the howling identification network starts from inputting the sound pressure space-time change feature vector into a convolutional neural network layer, the convolutional neural network layer is composed of multiple convolution kernels, the size of each convolution kernel is set to 3x3, the step is set to 1, and the feature map size is maintained by zero padding, the convolutional neural network layer is used to learn the local mode of the sound pressure gradient field, the local mode includes the edge and texture information of the sound pressure change, the output of the convolutional neural network layer is processed by a ReLU activation function, the output feature map is mapped to a recurrent neural network layer, the recurrent neural network layer adopts a long short-term memory unit structure, the number of hidden layer units of the long short-term memory unit structure is set to 128 to track the time evolution of the sound energy density distribution, the hidden state of the recurrent neural network layer is transmitted to a fully connected layer, the fully connected layer contains 256 neurons and is used to calculate the anomaly score of the phase coherence map, the output of the fully connected layer is converted to an anomaly score between 0 and 1 through a Sigmoid activation function, the local mode, the time evolution and the anomaly score are fused through a feature splicing operation, and the spliced feature vector generates the howling occurrence probability through a softmax function. It can be understood that the number of convolution kernels of the convolutional neural network layer can be adjusted according to the dimension of the input feature. Optionally, the recurrent neural network layer can be set to a bidirectional structure to capture the forward and backward time dependence.

[0076] The process of identifying high-density regions in the feature map and outputting the hotspot region coordinates using the clustering algorithm includes extracting the spatial coordinates and feature intensity values of all feature points from the feature map, the spatial coordinates are stored in a two-dimensional matrix form, the feature intensity values are normalized to the interval of 0 to 1, the distance between each feature point and the adjacent feature point is calculated using the Euclidean metric, a distance matrix is generated as the input of the density clustering algorithm, the density clustering algorithm is applied based on the distance matrix, the density clustering algorithm divides the feature points into multiple clusters, the feature point density of each cluster is identified by counting the number of feature points within a unit radius, the cluster whose feature point density exceeds the preset density threshold is selected as the high-density region, the geometric center coordinates of each high-density region are calculated using the arithmetic average method, and the hotspot region coordinates are output to the intervention scheme construction module. In some embodiments, the density clustering algorithm can be implemented using the DBSCAN algorithm, and the cluster division is based on the minimum point number parameter. It can be understood that the calculation of the feature point density needs to consider the normalization of the spatial scale. Optionally, the calculation of the geometric center coordinates can be weighted average with intensity weight for different feature points.

[0077] The generation process of the hotspot region coordinates can be expressed in mathematical form, and the clustering objective function is defined as: for optimizing cluster division, wherein: represents the value of the clustering objective function, represents the total number of feature points, represents the feature point of the neighborhood set, and represent the spatial coordinate vectors of the and the feature points, respectively. The minimization process of the objective function drives the density clustering algorithm to form high-density regions. In some embodiments, the generation of the distance matrix can introduce Mahalanobis distance to consider the correlation of feature distribution. Optionally, the feature intensity value can be used to adjust the weight coefficient in the distance calculation.

[0078] Referring to Figure 4The figure is the core visualization result of the howling identification module in the microphone anti-howling system, which intuitively presents the spatial distribution of sound field feature points and the howling hotspot identification effect: the light gray scattered points in the figure represent normal feature points, corresponding to sound field features without howling risk, and are distributed relatively dispersedly; the orange, blue and red dense scattered points correspond to three hotspot areas respectively, which are high-density feature clusters selected from the feature mapping by the density clustering algorithm, representing sound field areas with high howling risk; the black star marks are the geometric centers of the hotspot areas, that is, the hotspot area coordinates output by the system. The core function of the figure is to clearly distinguish between normal sound field and howling risk area, the distribution of its dense clusters reflects the recognition accuracy of the clustering algorithm, and the hotspot center coordinates provide accurate spatial positioning basis for the subsequent intervention scheme construction module, which not only intuitively displays the technical process of feature point clustering-> hotspot center calculation, but also reflects the spatial resolution and recognition accuracy of the system in the howling positioning link.

[0079] In embodiment 3, the step of correcting the howling occurrence probability in real time according to environmental acoustic parameters is as follows: collecting environmental acoustic parameters including environmental noise spectrum and reverberation time, calculating the deviation amount of the environmental noise spectrum from the reference noise spectrum, generating a noise correction factor, adjusting the sensitivity threshold of the howling occurrence probability according to the reverberation time, the sensitivity threshold being inversely proportional to the reverberation time, applying the noise correction factor and the sensitivity threshold to weighted average the howling occurrence probability, and outputting the corrected howling occurrence probability. The step of adjusting the sensitivity threshold of the howling occurrence probability according to the reverberation time includes measuring the reverberation time value of the current environment, obtaining a mapping relationship table of predefined reverberation time and sensitivity threshold, querying the corresponding reference sensitivity threshold in the mapping relationship table according to the reverberation time value, calculating the deviation proportion of the reverberation time value from the standard reverberation time, and linearly scaling the reference sensitivity threshold based on the deviation proportion, and taking the scaled value as the adjusted sensitivity threshold.

[0080] In a specific implementation, the step of correcting the howling occurrence probability in real time according to the environmental acoustic parameters starts from collecting the environmental acoustic parameters, including the environmental noise spectrum and the reverberation time, the environmental noise spectrum is measured by a spectrum analyzer in a silent state, and the reverberation time is calculated by the impulse response method. The deviation of the environmental noise spectrum from the reference noise spectrum is calculated using the frequency point difference method, and the reference noise spectrum is a pre-stored typical environmental noise sample. The noise correction factor is generated by weighted summation of the deviation. The sensitivity threshold of the howling occurrence probability is adjusted according to the reverberation time, and the sensitivity threshold is inversely proportional to the reverberation time. The noise correction factor and the sensitivity threshold are used to weight and average the howling occurrence probability, the weight coefficient of the weighted average is dynamically adjusted according to the environmental stability, and the corrected howling occurrence probability is output to the intervention scheme construction module. In some embodiments, the collection of the environmental noise spectrum can set multiple sampling points for spatial averaging to improve representativeness. Optionally, the generation of the noise correction factor can introduce a frequency weighting function to highlight the influence of key frequency bands.

[0081] Adjusting the sensitivity threshold of the howling occurrence probability according to the reverberation time includes measuring the reverberation time value of the current environment, the reverberation time value is extracted from the decay curve using the T30 or T60 method, obtaining a mapping relationship table of the predefined reverberation time and the sensitivity threshold, the mapping relationship table is stored in the non-volatile memory in the form of a lookup table, querying the corresponding reference sensitivity threshold in the mapping relationship table according to the reverberation time value, the query operation uses the nearest neighbor interpolation algorithm, calculating the deviation proportion of the reverberation time value and the standard reverberation time, the standard reverberation time takes a typical value between 0.5 seconds and 1.0 seconds, linearly scaling the reference sensitivity threshold based on the deviation proportion, the slope coefficient of linear scaling can be configured, and the scaled value is used as the adjusted sensitivity threshold for probability correction. It can be understood that the construction of the mapping relationship table is based on a large amount of measured data of acoustic environments. Optionally, the calculation of the deviation proportion can use a logarithmic scale to adapt to the large dynamic range of the reverberation time.

[0082] The adjustment relationship of the reverberation time to the sensitivity threshold can be quantitatively expressed by the formula: wherein: represents the adjusted sensitivity threshold, represents the reference sensitivity threshold obtained by querying, represents a preset proportion coefficient for controlling the adjustment amplitude, represents the reverberation time value of the current environment, represents the standard reverberation time value, and the formula realizes the inverse proportional adjustment mechanism of the sensitivity threshold and the reverberation time. In some embodiments, the linear scaling process can introduce upper and lower limit constraints to prevent the threshold from overflowing the effective range. It can be understood that the value of the proportion coefficient affects the adaptation speed of the system to different reverberation environments.

[0083] Referring to Figure 5 The figure is a comparative histogram, taking five typical acoustic environments such as quiet office, noisy conference room as the horizontal axis, and the probability of howling as the vertical axis. Through the difference between pink and green columns, the correction effect of environmental parameters on the probability of howling is intuitively presented. The core of this figure embodies the technical process of environmental acoustic parameter collection → noise correction factor / sensitivity threshold adjustment → probability weighted correction, which not only shows the interference difference of different environments on howling recognition, but also verifies the environmental adaptive ability of the probability correction module. By quantifying the influence of environmental factors, the probability of howling is more in line with the actual scene, providing more accurate decision basis for subsequent intervention scheme construction.

[0084] Embodiment 4: The method for constructing an acoustic intervention scheme includes analyzing the spatial relationship of hotspot area coordinates, determining the boundary of the acoustic interference area, selecting the intervention intensity level based on the corrected probability of howling occurrence, and designing the frequency response adjustment strategy for each intervention intensity level. The frequency response adjustment strategy involves notch filter setting and gain attenuation curve. The beamforming parameters of the acoustic interference area are generated, including the steering vector and the null depth. The combination of intervention intensity level, frequency response adjustment strategy and beamforming parameters forms a complete acoustic intervention scheme. The implementation of the frequency response adjustment strategy includes determining the center frequency and bandwidth of the notch filter according to the intervention intensity level, the center frequency of the notch filter corresponding to the dominant frequency of the howling, setting the starting point and slope of the gain attenuation curve to ensure smooth transition and avoid auditory mutation, verifying the compatibility of the notch filter and the system phase response, adjusting the group delay parameter, applying the adaptive algorithm to dynamically update the notch filter coefficients, and outputting the optimized frequency response configuration file.

[0085] In practical implementation, the method for constructing an acoustic intervention scheme begins with analyzing the spatial relationships of hotspot region coordinates. These coordinates are provided by the howling detection module. Spatial relationship analysis is achieved by calculating the Euclidean distance and angular relationships between coordinate points. The boundary of the acoustic interference region is determined using a convex hull algorithm or boundary detection method. The intervention intensity level is selected based on the corrected howling occurrence probability, categorized into low, medium, and high levels. This selection process involves comparing the corrected howling occurrence probability with a preset threshold. For each intervention intensity level, a frequency response adjustment strategy is designed, involving notch filter settings and gain attenuation curves. Notch filter settings include center frequency and bandwidth parameters, while the gain attenuation curve defines the starting point and slope of amplitude attenuation. Beamforming parameters for the acoustic interference region are generated, including a steering vector and null depth. The steering vector is calculated using an array signal processing algorithm, and the null depth is adjusted according to the interference region intensity. Combining the intervention intensity level, frequency response adjustment strategy, and beamforming parameters forms a complete acoustic intervention scheme, which is stored in a configuration file and transmitted to the execution module. In some embodiments, a mesh generation method can be introduced to improve accuracy in determining the boundary of the acoustic interference region. Optionally, the threshold for selecting the intervention intensity level can be dynamically adjusted based on historical howling data.

[0086] The implementation of the frequency response adjustment strategy includes determining the center frequency and bandwidth of the notch filter based on the intervention intensity level. The center frequency of the notch filter corresponds to the dominant feedback frequency, which is extracted from the spatiotemporal variation characteristics of sound pressure. The starting point and slope of the gain attenuation curve are set, ensuring a smooth transition and avoiding auditory abrupt changes. The starting point is set based on the feedback probability, and the slope is adjusted according to the intervention intensity level. The compatibility of the notch filter with the system phase response is verified by comparing phase offsets. The group delay parameter is adjusted to minimize phase distortion. An adaptive algorithm is applied to dynamically update the notch filter coefficients. The adaptive algorithm uses either the least mean square algorithm or the recursive least squares algorithm. The optimized frequency response configuration file is output. It is understood that the slope setting of the gain attenuation curve needs to consider the characteristics of human hearing. Optionally, the group delay parameter can be adjusted using an all-pass filter. Referring to Table 1, the correspondence between the intervention intensity level and the notch filter parameters is managed through a predefined table.

[0087] Table 1: Correspondence between Intervention Intensity Levels and Notch Filter Parameters

[0088] Intervention intensity level Center frequency (Hz) Bandwidth (Hz) Low 1000 100 Medium 2000 200 High 3000 300

[0089] In some embodiments, the parameter values ​​in the table can be calibrated according to the actual acoustic environment. Optionally, the center frequency and bandwidth values ​​can be extended to a frequency range to enhance adaptability.

[0090] The mathematical expression of the gain attenuation curve is given by the formula:

[0091]

[0092] wherein: G(f) represents the gain attenuation at frequency f, Gmax represents the maximum attenuation determined by the intervention intensity level, S represents the slope parameter of the attenuation curve, f0 represents the center frequency of the notch filter, and the formula ensures that the attenuation changes smoothly with the frequency offset from the center frequency. It can be understood that the value of the slope parameter S affects the steepness of the attenuation. Optionally, the exponential form in the formula can be replaced by a linear model to simplify the calculation. In some embodiments, the corresponding relationship of the notch filter parameters can be mathematically expressed by the formula:

[0093] In some embodiments, the corresponding relationship of the notch filter parameters can be mathematically expressed by the formula:

[0094] In some embodiments, the corresponding relationship of the notch filter parameters can be mathematically expressed by the formula:

[0095] In some embodiments, the corresponding relationship of the notch filter parameters can be mathematically expressed by the formula: wherein: f0 represents the center frequency of the notch filter,​​ a numerical code representing the intervention intensity level (e.g. low = 1, medium = 2, high = 3), a proportional coefficient, an offset, the formula is used to dynamically calculate the center frequency value, the bandwidth value can be similarly calculated by a linear relationship: determined, wherein: a bandwidth of the notch filter, and is a preset coefficient. It can be understood that the coefficient in the formula needs to be calibrated by experiment to ensure accuracy. Optionally, the setting of the center frequency and the bandwidth can introduce a non-linear mapping to handle complex frequency responses.

[0096] The operation of retrieving the corresponding center frequency and bandwidth values involves resolving the identifier of the current intervention intensity level, which is passed from the intervention scheme construction module, the query process uses binary search or hash algorithm to improve efficiency, the retrieved parameter values are temporarily stored in the register for subsequent use, the resonance point of the notch filter is set based on the retrieved center frequency, the resonance point setting is completed by configuring the filter coefficients through the digital signal processor, the filter coefficients are calculated according to the center frequency, the attenuation range of the notch filter is set based on the retrieved bandwidth value, the attenuation range setting is achieved by adjusting the stopband width of the filter, ensuring that the attenuation effect covers the target frequency interval. In some embodiments, the resonance point setting can be combined with an adaptive algorithm to fine-tune in real time to compensate for environmental changes. Optionally, the setting of the attenuation range can verify the compatibility with the overall frequency response of the system.

[0097] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and variations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of the present application being defined by the appended claims and their equivalents.

Claims

1. A microphone feedback prevention system, characterized in that, The system includes: An audio capture module is used to continuously capture the audio input stream of the target microphone in the operating scenario, wherein the audio input stream includes time-domain waveform data and background acoustic parameters; The dynamic range compression module is used to perform dynamic range compression processing on the audio input stream to generate a standardized audio frame sequence. The feature extraction module is used to extract sound pressure spatiotemporal variation features from the standardized audio frame sequence. The sound pressure spatiotemporal variation features include sound pressure gradient field, sound energy density distribution, and phase coherence spectrum. The howling detection module includes a howling detection network, which is used to feed the spatiotemporal variation characteristics of the sound pressure into the howling detection network and output the howling occurrence probability and hot spot area coordinates; The probability correction module is used to correct the howling occurrence probability in real time based on environmental acoustic parameters and generate a corrected howling occurrence probability. An intervention scheme construction module is used to construct an acoustic intervention scheme based on the corrected howling occurrence probability and the coordinates of the hotspot area.

2. The microphone anti-feedback system according to claim 1, characterized in that, The specific implementation of dynamic range compression is as follows: Monitor the peak and average levels of the audio input stream and calculate the instantaneous dynamic range coefficient; The gain control curve is adjusted based on the instantaneous dynamic range coefficient, and the gain control curve is jointly defined by the logarithmic amplifier and the limiter. The audio input stream is amplitude shaped using a gain control curve to eliminate overload distortion components. Band equalization is performed on the amplitude-shaped signal to improve mid-frequency clarity. Output a normalized audio frame sequence with a uniform energy distribution.

3. A microphone anti-feedback system according to claim 2, characterized in that, The process of extracting the spatiotemporal variation characteristics of sound pressure includes: The standardized audio frame sequence is divided into overlapping time windows, each time window corresponding to a sound pressure sampling period; for each time window, the sound pressure gradient field is calculated, which is obtained through three-dimensional spatial difference operation and reflects the rate of change of sound pressure with position. The acoustic energy density distribution is derived, which is generated by convolving the square of the sound pressure with the medium properties. The phase coherence spectrum is measured, and the phase coherence spectrum is obtained by cross-correlation analysis of the phase difference of different microphone units; The sound pressure gradient field, sound energy density distribution, and phase coherence spectrum of continuous time windows are aggregated to form a spatiotemporal fusion feature vector.

4. A microphone anti-feedback system according to claim 3, characterized in that, The operation of a howling identification network is as follows: The sound pressure spatiotemporal variation feature vector is input into a convolutional neural network layer, which learns the local pattern of the sound pressure gradient field. The output features are mapped to a recurrent neural network layer, which tracks the temporal evolution of the acoustic energy density distribution; the hidden state of the recurrent neural network layer is passed to a fully connected layer, which calculates the anomaly score of the phase coherence spectrum. By integrating local patterns, temporal evolution, and anomaly scores, the probability of howling occurrence is generated through the softmax function; at the same time, a clustering algorithm is used to identify high-density regions in the feature map and output the coordinates of hotspot regions.

5. A microphone anti-feedback system according to claim 4, characterized in that, The method of using clustering algorithms to identify high-density regions in feature maps and outputting the coordinates of hotspot regions includes: Extract the spatial coordinates and feature intensity values ​​of all feature points from the feature map; Calculate the distance between each feature point and its neighboring feature points to generate a distance matrix; Based on the distance matrix, a density clustering algorithm is applied to divide the feature points into multiple clusters. Identify the feature point density of each cluster and select clusters whose feature point density exceeds a preset density threshold as high-density regions. Calculate the geometric center coordinates of each high-density region and output the coordinates of the hotspot region.

6. A microphone anti-feedback system according to claim 1, characterized in that, The steps for real-time correction of the whistling probability based on environmental acoustic parameters are as follows: Collect environmental acoustic parameters, including the environmental noise spectrum and reverberation time; Calculate the deviation between the environmental noise spectrum and the reference noise spectrum, and generate a noise correction factor; The sensitivity threshold for adjusting the probability of howling occurrence is based on the reverberation time, wherein the sensitivity threshold is inversely proportional to the reverberation time; The noise correction factor and sensitivity threshold are applied to perform a weighted average of the howling probability, and the corrected howling probability is output.

7. The microphone anti-feedback system as described in claim 6, characterized in that, The sensitivity threshold for adjusting the probability of howling based on reverberation time includes: Measure the reverberation time of the current environment and obtain a predefined mapping table between reverberation time and sensitivity threshold; Based on the reverberation time value, look up the corresponding baseline sensitivity threshold in the mapping table; Calculate the deviation ratio between the reverberation time value and the standard reverberation time, and linearly scale the reference sensitivity threshold based on the deviation ratio; The scaled value is used as the adjusted sensitivity threshold.

8. A microphone anti-feedback system according to claim 1, characterized in that, Methods for constructing acoustic intervention programs include: Analyze the spatial relationships of the hotspot region coordinates to determine the boundary of the acoustic interference region; The intervention intensity level is selected based on the corrected probability of howling occurrence, and the intervention intensity level is divided into three levels: low, medium and high. For each intervention intensity level, a frequency response adjustment strategy is designed, which involves notch filter settings and gain attenuation curves; Beamforming parameters are generated for the acoustic interference region, including the steering vector and null depth. By combining the intervention intensity level, frequency response adjustment strategy, and beamforming parameters, a complete acoustic intervention scheme is formed.

9. A microphone anti-feedback system according to claim 8, characterized in that, The implementation of the frequency response adjustment strategy includes: The center frequency and bandwidth of the notch filter are determined according to the intervention intensity level. The center frequency of the notch filter corresponds to the dominant frequency of the howling. The starting point and slope of the gain attenuation curve are set to ensure a smooth transition and avoid auditory abrupt changes. Verify the compatibility between the notch filter and the system phase response, and adjust the group delay parameters; An adaptive algorithm is applied to dynamically update the notch filter coefficients, and an optimized frequency response configuration file is output.

10. The microphone anti-feedback system as described in claim 9, characterized in that, The determination of the notch filter center frequency and bandwidth based on the intervention intensity level includes: A predefined correspondence between intervention intensity levels and notch filter parameters is defined, where each intervention intensity level is associated with a set of center frequency and bandwidth values; Based on the currently selected intervention intensity level, retrieve the corresponding center frequency and bandwidth value from the correspondence; The resonant point of the notch filter is set based on the retrieved center frequency; The attenuation range of the notch filter is set based on the retrieved bandwidth value.

Images