Active howling elimination method and system based on acoustic signal pre-judgment

Abstract

The invention discloses an active howling elimination method and system based on acoustic signal pre-judgment, and relates to the technical field of sound system howling elimination, and the elimination method comprises the following steps: collecting an acoustic signal from at least one microphone in real time, and segmenting the acoustic signal according to frames; performing front-end preprocessing on the acoustic signal of each frame; extracting a plurality of acoustic features from the preprocessed frame signal; the acoustic features are input into a real-time predictor, and the predictor outputs a squeal risk score in a future short time window, one or more possible dominant squeal frequency bands and corresponding predicted phase information; and comparing the risk score with at least one preset threshold value. Through risk prediction of the short-time window, prediction of the howling risk, the dominant frequency band and the phase trend in the short-time window is realized, the continuous risk score is output for the controller to make a decision, the traditional time delay limitation of only afterward detection or slow self-adaption is overcome, and passive response of howling formation and amplification is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sound system howling elimination, in particular to an active howling elimination method and system based on acoustic signal prediction. BACKGROUND

[0002] Howling in a sound system is caused by the output sound of a loudspeaker propagating through space or structure and being picked up again by a microphone and amplified to form a positive feedback loop. Traditional howling elimination mainly relies on passive gain reduction, fixed notch or post-inhibition based on adaptive filtering. These methods often have difficulty in balancing stability and sound quality: passive gain reduction severely loses loudness, fixed notch requires pre-knowledge of frequency and affects the speech band, and traditional adaptive algorithms are slow to adapt to nonlinearity, multipath and rapid environmental changes. More importantly, most existing solutions are "post-inhibition", lacking the ability to make short-term prediction and actively intervene before howling is formed, resulting in large response delay, frequent misjudgment or excessive inhibition. In the face of real application scenarios such as multi-mic array, complex reflection field and nonlinear loudspeaker distortion, there is an urgent need for a howling elimination method and system that is low-latency, robust and minimally damages sound quality.

[0003] Patent CN105228056B discloses a method and system for eliminating microphone howling, which detects and eliminates the frequency of the sound that produces howling.

[0004] The above patent eliminates the echo signal of a certain frequency point that causes the microphone to howl, but lacks the ability to make short-term prediction and actively intervene before howling is formed.

[0005] Therefore, the present application proposes an active howling elimination method and system based on acoustic signal prediction, which makes short-term prediction and actively intervenes before howling is formed. SUMMARY

[0006] The present application aims to provide an active howling elimination method and system based on acoustic signal prediction to solve the technical problem of traditional howling elimination relying mainly on passive gain reduction, fixed notch or post-inhibition based on adaptive filtering, which often has difficulty in balancing stability and sound quality.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical solution: an active howling elimination method based on acoustic signal prediction, the elimination method comprising the following steps: Real-time acquisition of acoustic signals from at least one microphone and frame segmentation of the acoustic signals; Performing front-end preprocessing on each frame of acoustic signals, including DC removal, window function, band-pass filtering and energy normalization; extracting several acoustic features from the pre-processed frame signal, the acoustic features including short-time spectral amplitude spectrum, short-time energy of each frequency band, envelope rise rate, inter-band cross-correlation peak, instantaneous phase stability index, and short-time energy change rate in time domain; inputting the acoustic features into a real-time predictor, the predictor outputting: a risk score of howling in a future short-time window and one or more possible dominant howling frequency bands and corresponding predicted phase information; comparing the risk score with at least one preset threshold value, if the risk score exceeds the threshold value, triggering a set of active intervention strategies in priority order; evaluating system effect in real time during execution of the active intervention strategy, deciding to upgrade to the next high-priority intervention or fallback to the previous safe parameter according to real-time energy and stability index; using the triggering event and recorded acoustic features, control action and effect as samples to update the predictor and controller parameters online or offline.

[0008] Preferably, the set of active intervention strategies includes: output gain fine-tuning to reduce loop gain without degrading sound quality; variable Q narrowband adaptive notch based on predicted dominant frequency band, center frequency provided by the predictor, Q value dynamically adjusted within the prediction range; phase cancellation, generating a reverse or phase-shifted compensation signal to inject into the output path based on predicted phase information; spatial domain control, spatially nulling or suppressing feedback direction through beam adjustment of microphone array or loudspeaker unit; injecting a low-amplitude probe sequence during speech gaps to carry out probe-based feedback path identification.

[0009] Preferably, the predictor is a time series model or a hybrid of time series model and rule engine, the time series model selected from the following candidates: small long short-term memory network LSTM, causal convolutional network TCN, one-dimensional Transformer or quantized distillation lightweight model of one-dimensional Transformer, the predictor having a reasoning delay of less than 10 ms.

[0010] Preferably, the parameters of the variable Q narrowband adaptive notch satisfy: the center frequency ±1 / 2 frequency point is determined by the predictor output, the Q value is dynamically adjusted within the range of 15-60, the minimum bandwidth of the notch band does not exceed 3% of the center frequency, and the maximum duration of the notch amplitude is limited to avoid speech perceptible distortion; The probe sequence is a pseudo-random low-amplitude signal, the amplitude of the probe sequence is less than or equal to -30 dBFS of the maximum perceptible amplitude of the system, and is injected only in the "gap" or low-energy section of the voice activity VAD detection decision; the response after the probe injection is used to construct or update the loudspeaker-to-microphone transfer function, including approximation of the nonlinear term.

[0011] Preferably, the real-time effect evaluation includes at least two indicators: Short-term frequency band energy rate of change and phase energy stability index; When both indicators show improvement, the intervention is considered effective, otherwise the intervention intensity is increased according to the preset escalation strategy and the safety rollback strategy is triggered if necessary, which restores the system parameters to the last known stable parameter set.

[0012] Preferably, the cancellation method further includes direction of arrival DOA estimation in a multi-microphone scene to assist spatial domain control, the DOA estimation is based on short-term phase difference TDOA or array correlation method, and the spatial domain support is based on MVDR or minimum variance criterion beam nulling.

[0013] Preferably, the cancellation system includes: A microphone array or at least one microphone for collecting live acoustic signals; A front-end analog-to-digital conversion unit ADC and a front-end preprocessing module for removing direct current, window function processing, and baseband bandpass filtering; An acoustic feature extraction module for calculating short-term spectrum, envelope rise rate, frequency band energy, cross-correlation peak value, and phase stability features from the frame signals from the front-end preprocessing module; A predictor module for real-time output of a howling risk score and predicted dominant frequency band and predicted phase information with acoustic features as input; An active intervention controller connected to the predictor module and scheduling the execution of intervention strategies in priority order when the risk score exceeds a preset threshold, the intervention strategies including an output gain fine-tuning unit, a variable Q narrowband notch filter unit, a phase cancellation generator, a beamforming nulling unit, and a probe injection unit; A feedback path identification module for constructing or updating the loudspeaker-microphone transfer function and the nonlinear term approximation model under the conditions of probe injection or blind identification; An effect evaluation and rollback module for real-time monitoring of short-term energy and phase stability and performing parameter rollback when the stability criterion is not met; A storage and update module for recording trigger events, training samples, and supporting online or offline parameter update of the predictor; A hardware platform, which is a combination of audio DSP, ARM processor, FPGA and NN accelerator, guarantees that the end-to-end inference and filtering delay is lower than a preset upper limit.

[0014] Preferably, the predictor module is implemented as a quantized or distilled lightweight neural network to meet real-time inference requirements, and eliminates the system in local deployment lightweight predictor and controller, while working with edge, cloud model training and parameter optimization platform, the data interaction between local and cloud only uploads de-identified statistical features and event summaries to protect privacy.

[0015] Preferably, the feedback path identification module includes a blind identification submodule and a probe identification submodule: the blind identification submodule is started when there is continuous speech to continuously estimate the feedback path, and the probe identification submodule is started when a speech gap is detected to inject a probe for a short time and refine the transfer function based on the probe response; at the same time, the feedback path identification module supports modeling and online updating of nonlinear terms.

[0016] Preferably, the active intervention controller implements a hierarchical priority strategy and a safety fallback mechanism: When the predictor indicates that the howling risk reaches a first threshold, only a low-intrusive action is performed; when the risk continues to rise or the low-intrusive action is invalid, a medium-level action, i.e., variable Q notch and short-time phase cancellation, is triggered in sequence, and a high-level action, i.e., beam nulling, probe identification, and updating of the feedback model based thereon, is triggered; in any case, if the effect evaluation module does not detect stability improvement within a preset time window, the system automatically falls back to the last known stable parameters and generates a log alarm information.

[0017] Compared with the prior art, the present application has the following advantages: 1. The present application realizes prediction of howling risk, dominant frequency band and phase trend within a short time window through short-time window risk prediction, outputs continuous risk scores for controller decision-making, overcomes the time delay limitation of traditional post-detection or slow adaptation, avoids passive response to howling formation and amplification, reduces the situation of false judgment of frequent activation of strong suppression measures, significantly reduces response time and required suppression strength, enables the system to block howling development with small amplitude actions without sacrificing sound quality, thereby improving available system gain and listening experience; 2. The present application realizes step-by-step scheduling of control actions from low intrusion to high intrusion according to predetermined priority through a set of hierarchical active intervention strategies, avoids "one-size-fits-all" strong notch or large gain reduction leading to significant distortion and loudness loss of voice / music, reduces the negative impact of false triggering on user experience, solves the problem through low-intrusive actions in most scenarios, and only uses stronger means when necessary, thereby achieving a better balance between stability and sound quality and reducing the frequency of manual debugging; 3. The present application solves the problem of low accuracy of traditional identification under continuous speech or high nonlinear distortion conditions, leading to slow convergence or failure of adaptive compensation, and the problem of single blind identification being easily disturbed by noise and multipath interference, to obtain a more accurate and faster feedback path model without significantly affecting the listening experience, and to improve the directivity and effectiveness of notch, phase cancellation and spatial control, especially in complex multipath or nonlinear conditions. Robustness is significantly enhanced; 4. The present application solves the problem of single domain control (only frequency domain or only spatial domain) in multi-source, multi-reflection and nonlinear environment, and the problem of single measure easily causing new instability or excessive impact on sound quality. Through joint control, the loop gain is more accurately and locally suppressed while maintaining high loudness and sound quality. Low delay closed loop and fallback mechanism ensures that any control action will not introduce new instability in a short time, improving the safety and usability of the system in actual engineering deployment. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The system overall process schematic diagram of the present application is shown in the figure; Figure 2 The predictor internal processing schematic diagram of the present application is shown in the figure; Figure 3 The probe injection and feedback path identification and model update schematic diagram of the present application is shown in the figure. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described clearly and completely 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, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0020] Please refer to Figure 1 , Figure 2 and Figure 3 An embodiment provided by the present application: an active howling elimination system and method based on acoustic signal prediction, which is suitable for single microphone-single speaker conference terminal, and performs real-time prediction and hierarchical intervention frame by frame: Hardware and module configuration: 1 microphone (single channel), 1 speaker; ADC 48 kHz, 24 bit; local processor: audio DSP or ARM Cortex-A series; modules: front-end preprocessing, feature extraction, predictor, active intervention controller, effect evaluator, log recorder.

[0021] Frame and timing parameters: Frame length N = 256 samples; window function: Hann window, hop = 128; feature history buffer M = 8 frames for short-term trend analysis and prediction input window; shortest control response target: complete control command issuance within < 1 frame from risk decision to first action execution.

[0022] Feature computation: STFT magnitude spectrum: X[k] = |STFT(frame)| (256-point FFT); band energy B_i = Σ_{k∈band_i}|X[k]|^2 (partitioned per 1 / 3 octave or fixed narrow band, e.g. per 1 / 12 octave or 50 Hz bandwidth); short-term energy E_t = Σ_{n=0}^{N-1}x[n]^2; envelope rise rate S_t = (E_t - E_{t-1}) / frame_duration (frame_duration = 5.333333 ms); band cross-correlation peak C_{i,j} = max_τcorr(B_i[t-M+1:t], B_j[t-M+1:t]); phase stability indicator P = 1 - circular_variance(phase(k) over last M frames), computed for main frequency bands; normalize each feature to zero-mean unit-variance for model input.

[0023] Risk score and trigger rules: risk score R = σ(w1 · norm(EnergySlope) + w2 · norm(peakBandEnergy) + w3 · norm(Correlation) + w4 · norm(1-P)), σ is Sigmoid. Example weights: w1 = 0.35, w2 = 0.35, w3 = 0.15, w4 = 0.15; threshold settings: low risk T1 = 0.35; medium risk T2 = 0.60; high risk T3 = 0.85; Decision flow: if R < T1: no action, only monitoring; if T1 ≤ R < T2: perform low-invasive action; if T2 ≤ R < T3: perform low-invasive action while enabling intermediate actions; if R ≥ T3: enable all intermediate actions and prepare to launch high-level actions, and start probe injection opportunity.

[0024] Active intervention actions: Low-invasive action: output gain fine-tuning: immediately reduce the overall gain of the speaker output by ΔG = 0.5 dB; if there is still no improvement in 3 consecutive frames, reduce by another 0.5 dB, cumulative reduction not exceeding -3.0 dB. Each step of reduction has a duration observation window T_obs = 50 ms.

[0025] Intermediate action: adaptive narrow-band notch: center frequency f0 output by predictor; set Q initial value 30, if speech perceptible distortion needs to be shortened, maximum duration of notch T_notch_max = 200 ms; phase cancellation: generate a narrow-band compensation signal opposite to the predicted phase (with a + pi offset) for the dominant frequency band, set the amplitude according to 80% of the estimated closed-loop gain G_loop(f0), and the duration is the same as or shorter than the notch.

[0026] Advanced action: if the device supports spatial control, switch to a short-time "mute" local frequency band or use phase modulation; at the same time, trigger probe injection to identify the feedback path.

[0027] Effect evaluation and rollback: Calculate ERR = (E_before - E_after) / E_before every frame after the action is executed; if ERR >= 0.20 and the phase stability improvement >= 15%, it is considered effective; otherwise, perform stronger actions or rollback to the last stable parameter set according to the escalation strategy. The rollback action will gradually restore all changes to the original value to avoid sudden loudness jumps.

[0028] Calibration and debugging steps (first deployment or site change): initial calibration: play a low-amplitude chirp (1kHz-8kHz) or white noise without speech, collect 2s, calculate the initial loudspeaker -> microphone transfer function H0(f), and save it as the baseline on the device; if the site noise is high, repeat 3 times to take the average. Parameter fine-tuning: record 50 events during the meeting and manually review them, adjust the weights w_i and thresholds T1-T3 until the false positive rate < 5% and the false negative rate < 10% (empirical threshold).

[0029] Acceptance criteria: test in 3 rooms and 3 microphone positions: the system can allow the average system gain to rise by at least +6dB without howling relative to no active control; and the speech perceptible distortion < 0.5 MOS degradation.

[0030] See Figure 1 , Figure 2 and Figure 3 , an embodiment provided by the present application: an active howling elimination system and method based on acoustic signal prediction, using multi-microphone array + space-frequency domain joint, illustrating DOA-driven beam nulling + narrow-band notch: Microphone array M = 8 channels (linear or circular array), each channel synchronously sampled at 48kHz; multi-channel front-end ADC, channel clock synchronization, array time delay calibration; modules: multi-channel preprocessing, array cooperative feature extraction, DOA algorithm, predictor, MVDR / nulling beam unit, narrow-band notch filter group, effect evaluation module.

[0031] STFT (same parameters as above) for each channel, compute the cross-channel cross-correlation matrix Φ(ω) = E[X(ω)X(ω)^H]; DOA estimation: use GCC-PHAT to estimate TDOA for two channels and do direction voting by multiple pairs; or use MUSIC algorithm to get a set of possible DOA candidates on each frame. Output DOA θ* and its confidence conf_DOA (0-1); combine DOA with band energy migration: if a band has significant energy rise at DOA θ* and R rises, then this band and spatial direction are potential loop sources.

[0032] Spatial action detailed steps: estimate the covariance matrix R = (1 / K)∑_tx_tx_t^H (K = window frame number, example K = 8); construct the expected direction vector v(θ*), calculate the weight vector w_null = R^{-1}v(θ*) / (v(θ*)^HR^{-1}v(θ*)); construct constraints for "nulling": set the minimum response in the θ* direction while keeping the target direction gain unchanged; implementation: adjust the beam weight vector to introduce a zero point with a depth ≥ 15 dB in the θ* direction; implementation details: the weight vector is calculated every 8 frames, and the update time is not more than 42.666664 ms. When applied, do limited amplitude smoothing on the output signal to avoid sound field mutations.

[0033] Frequency domain cooperative action (when spatial nulling is insufficient): when the predicted dominant frequency band still has residual energy after beam control, enable the same frequency band narrow band notch (Q = 30), and the notch center f0 is combined with DOA to confirm whether the target source is fixed (if the DOA fluctuation is large, reduce the notch duration to avoid killing voice); notch application sequence: spatial nulling → if ERR < 0.15, then add a notch in the same frequency band.

[0034] Data flow and timing (frame level): calculate multi-channel features every frame and accumulate to M windows (8 frames) for DOA and predictor use; when DOA confidence conf_DOA < 0.5, reduce dependence on spatial action and switch to frequency domain dominant suppression.

[0035] Fault tolerance and stability control: if continuous 3 rounds of spatial weight vector adjustment cause voice core energy to drop more than 2 dB and subjective report is "voice darkening", automatically back off 50% of the last weight vector adjustment; for any failure of array channels (a certain road is disconnected or noise bursts), the system switches to the remaining channels and recalculates the covariance matrix within 100 ms.

[0036] Calibration and acceptance: Calibration procedure: Play known direction white noise and chirp from -90° to +90° with 5° sampling, build steering vector table; Acceptance criterion: When simulating two or more feedback sources (two sources separated by ≥ 30°), the system can reduce the average energy of the dominant loop direction by ≥ 20% and the speech intelligibility (repeat accuracy) by ≤ 5%.

[0037] See Figure 1 , Figure 2 and Figure 3 , an embodiment of the present invention provides an active feedback-based feedback cancellation system and method, which is explained by the following: Multiple power amplifiers drive large line array loudspeakers; each loudspeaker unit can be individually phase shifted (hardware supports delay / phase control, resolution ≤ 10 µs); the central controller needs to communicate with the power amplifier / loudspeaker control interface (e.g. Dante / AES67 / Control Bus) to configure the gain and phase delay.

[0038] Risk assessment and hierarchical strategy (time table): longer risk assessment window: use M = 16 frames to reduce false triggering; hierarchical response time nodes (example): t0 (assessment): trigger low-invasive action (ΔG = -0.5 dB); observation period T1 = 100 ms; if t0 + T1 is ineffective: enable variable-Q notch, Q initial value 40, T2 = 200 ms; if still ineffective: enable phase distribution adjustment (set ± φ shift for specific loudspeaker units, φ initial value 5°, max ± 30°), T3 = 300 ms; if still ineffective or the system detects a sharp rise in energy (peak exceeds the safety SPL threshold), immediately trigger global protection.

[0039] Phase adjustment details: only adjust the phase of the loudspeaker units related to the predicted dominant frequency band f0: use FIR or short delay to achieve frequency band-dependent phase adjustment. Phase fine-tuning amplitude is calculated according to the closed-loop phase margin: if the closed-loop phase margin φ_margin < 20°, start adjustment, target is to shift the closed-loop loop phase to 180° ± 20°. Adjustment step: ± 5° each time and wait for 50-100 ms to observe the effect, max 6 steps.

[0040] Probe identification and engineering operation: during tuning / rehearsal, allow injection of large amplitude probes (e.g. -10 dBFS, length 500 ms broadband chirp) to obtain high-quality transfer functions; during formal performance, only inject low-amplitude probes when high risk is detected and there is a short gap. Probe results are used to fit high-order nonlinear models to estimate effective compensation strategies under high-power distortion conditions.

[0041] Safety and responsibility mechanism: When the system triggers fastmute, the following must be recorded: trigger time, trigger reason, executed action sequence, and the event must be alarmed through the control panel until manual reset. After fastmute execution, the amplifier output is automatically attenuated to 0 and after 2s, it is allowed to recover with -6dB degradation, and further recovery needs manual confirmation.

[0042] Debugging and acceptance: Before the TDH (total harmonic distortion) rises to the system safety threshold, the system should be able to reduce the dominant frequency band energy by at least 30% within 200ms. Listening test: Under the stage normal sound pressure, the subjective score of voice and music decreases ≤0.7 (MOS difference).

[0043] Please refer to Figure 1 , Figure 2 and Figure 3 , an embodiment provided by the present application: an active howling elimination system and method based on acoustic signal prediction, in which the probe-assisted feedback path identification is implemented as follows: Probe type: short pseudo-random binary sequence or short chirp, preferably m-sequence due to good autocorrelation. Length: L=512 samples or 1024 samples according to scene selection; longer sequence can be used for stage, and short sequence for conference equipment. Amplitude: A_probe≤-30dBFS; if allowed during tuning phase, -20dBFS can be used. Window processing: use Hann window to smooth the rise / fall of both ends of the probe by more than 2ms to avoid impact noise.

[0044] Only inject when "gap" (≥30ms) or low-energy window (short-time energy <E_bg+3dB) is detected. The injection interval is at least 2s.

[0045] Record the microphone signal y(t) during injection. Calculate the cross-correlation r(τ) of probe→response or estimate the system impulse response h(t) by Wiener deconvolution: H(f)=Y(f) / P(f) (avoid division by zero in frequency domain, use small regularization λ=1e-6). Do short-time energy and phase analysis on the obtained linear impulse response, extract the main peak delay and frequency response amplitude-phase characteristics.

[0046] Select a memory polynomial model: y(t)≈Σ_{p=1}^{P}Σ_{m=0}^{M}a_{p,m}x(t-m)|x(t-m)|^{p-1}, commonly used P=3, M=10 as the starting point. Fit a_{p,m} with probe response (least squares or regularized least squares, L2 regularization term λ=1e-3). The fitting result is used to introduce an inverse nonlinear compensation term in subsequent phase cancellation or amplitude compensation.

[0047] The linear transfer function is updated immediately after the probe trigger and the updated result is written to the local model library and the version number is recorded within 1s if the multiple probe data are consistent. The nonlinear parameters are only merged into the model when the difference between two or more probe response sequences is less than 10%, otherwise, the parameters are kept as candidates and wait for more samples to avoid overfitting noise.

[0048] The effective identification is considered when the correlation peak SNR increases by more than 6dB. The linear model fitting mean square error (MSE) is less than 1e-6 or the relative error is less than 10%.

[0049] Please refer to Figure 1 , Figure 2 and Figure 3 , an embodiment provided by the present application: an active howling elimination system and method based on acoustic signal prediction, applicable scenarios: vehicle audio, implementation details of low delay and vehicle condition coupling: Typical vehicle sampling 48kHz; strict segment delay budget for eyes and ears: end-to-end target delay ≤8ms. Preferred hardware: vehicle-grade MCU + DSP + NPU, memory ≥128MB, main processing frame processing delay ≤2ms.

[0050] In addition to acoustic features, vehicle condition features are added: vehicle speed V (m / s), window open / close (binary), wind noise estimation, air conditioning state. These features and acoustic features are input to the predictor at a uniform frequency, which can significantly improve the prediction performance in high-speed driving or window opening state.

[0051] Model: use TC-Net or 1D causal convolution network, model parameter quantity limit ≤50k parameters, quantized to int8. Inference delay <1.5ms. Trap and phase compensation preferentially use short FIR (length ≤32 taps) or IIR (second order) to ensure processing delay ≤1ms. Probe: only in parking or silence; otherwise, rely on blind identification.

[0052] Sudden changes: automatically trigger a short probe or switch to a more conservative control within 200ms after detecting a window state change to avoid frequent false positives. In driving state, prefer to use phase fine-tuning and small gain adjustment (≤0.5dB per step); only allow more aggressive trap or probe operation when the vehicle is parked or the speed is less than 5km / h.

[0053] Initial calibration: play chirp1kHz-8kHz (-20dBFS) under the condition that the vehicle is stationary and in an indoor environment, record and generate the initial H0(f). Online adaptation: use a fine-tuning window every 5 minutes to make offline small-step updates based on collected events; limit the maximum number of model updates per minute to ≤1 to avoid parameter jitter.

[0054] Test in three states of vehicle speed 0 / 30 / 90km / h: under the same microphone-loudspeaker gain conditions, the system can extend the maximum gain threshold before howling by an average of ≥4dB; the system intervention has an impact on voice intelligibility ≤5% (word recognition rate decreases by no more than 5%).

[0055] Please refer to Figure 1 、 Figure 2 and Figure 3 , an embodiment provided by the application: an active howling elimination system and method based on acoustic signal prediction, for earphone, hearing aid personalized howling elimination: Device: in-ear or open earphone / hearing aid, sampling rate 48kHz, limited end-side processing capability. Delay target: end-to-end <6ms.

[0056] Personalized calibration steps (first wear): Step A: short probe sequence (L=256samples, amplitude-20dBFS) is injected three times, and ear canal response is collected. The personalized transfer function H_ear(f) is estimated by deconvolution.

[0057] Step B: based on H_ear(f), the subjective sensitive frequency band is calculated and higher weights are allocated to these frequency bands in the predictor.

[0058] Step C: store personalized parameters in device NVM.

[0059] Extremely low-invasive control strategy: preferred operation: extremely short-time gain adjustment (ΔG=-0.25dB, extremely small step size) and short-time phase compensation. If must be notched: limit Q≤25 and maximum duration ≤100ms, and perform gradual smoothing within 30ms after the start of notching.

[0060] Energy and perceptual evaluation: built-in simple perceptual model (based on Bark band masking) in the device to evaluate the human perceptible distortion caused by intervention; only when the perceptual loss index P_loss≤0.05, intermediate action is allowed. P_loss can be estimated by the change of dominant frequency band energy ratio and amplitude-frequency curve.

[0061] Long-term adaptation and user feedback: If the user reports "sound quality problem" through the APP, an offline refined calibration will be triggered. The device retains the latest 50 times of trigger records and corresponding parameters for remote diagnosis and cloud improvement.

[0062] Acceptance criteria: in the subjective test of 20 subjects, after personalized calibration, the subjective satisfaction of the system in suppressing howling caused by wearing / leakage is ≥4.0 / 5; and the sound quality degradation score is ≤0.5.

[0063] Working principle: The system continuously collects acoustic signals from the microphone or microphone array, and processes them in short frames in time domain and frequency domain. A set of features representing the short-time howling trend is calculated for each frame. The feature sequence of several historical frames is input into a real-time predictor. The predictor outputs the howling risk score, the possible dominant frequency band and its predicted phase information within a very short prediction window, so as to predict whether howling will occur in advance. When the predictor gives a risk signal higher than the threshold, the system performs active suppression measures in a pre-defined set of hierarchical priorities: first, make a slight gain adjustment, if still ineffective, enable variable-Q narrowband notch, while generating in parallel narrowband phase compensation components in phase and anti-phase for short-time phase cancellation; when equipped with array capability, use beam nulling or loudspeaker unit phase adjustment to suppress the feedback direction in space; if necessary, inject low-amplitude probes during speech gaps to accurately identify the feedback path and update the controller parameters. All control actions are based on frame-level closed-loop effect evaluation. If the action is ineffective, it is automatically upgraded or downgraded to avoid adding instability. The system uses the trigger events, feature sequences and effects as samples to fine-tune the predictor and control strategy online or offline, while built-in stability and safety criteria: there is a maximum duration and amplitude limit for each control action. If no improvement is seen within a certain number of frames, it triggers a rollback to the last known stable parameter set and records an alarm. The system also saves operation logs and versioned model parameters, supports periodic optimization on the edge and cloud, but ensures low-latency real-time response locally.

[0064] It is apparent for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, but can be implemented in other concrete forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered in all aspects as illustrative and not restrictive, and the scope of the present application is defined by the appended claims rather than the above description, and all changes falling within the meaning and range of equivalent elements of the claims are intended to be embraced in the present application. Any reference signs in the claims should not be considered as limiting the claims involved.

Claims

1. An active howling cancellation method based on acoustic signal prediction, characterized in that: The elimination method includes the following steps: Acoustic signals from at least one microphone are acquired in real time and segmented into frames; Perform front-end preprocessing on the acoustic signal for each frame, including DC removal, windowing, bandpass filtering, and energy normalization; Several acoustic features are extracted from the preprocessed frame signal. These acoustic features include short-time spectral amplitude spectrum, short-time energy of each frequency band, envelope rise rate, peak value of cross-correlation between frequency bands, instantaneous phase stability index, and time-domain short-time energy change rate. The acoustic features are input into the real-time predictor, and the predictor outputs: the howling risk score within a short future time window and one or more possible dominant howling frequency bands and their corresponding predicted phase information; The risk score is compared with at least one preset threshold. If the risk score exceeds the threshold, a set of proactive intervention strategies in priority order is triggered. The system effectiveness is evaluated in real time during the implementation of proactive intervention strategies, and decisions are made based on real-time energy and stability indicators to upgrade to the next higher priority intervention or revert to the previous safe parameters. Triggering events and recorded acoustic features, control actions, and effects are used as samples to update predictor and controller parameters online or offline.

2. The active howling cancellation method based on acoustic signal prediction according to claim 1, characterized in that: The set of proactive intervention strategies includes: Output gain fine-tuning reduces loop gain without degrading sound quality; Based on the predictive dominant frequency band, the center frequency is provided by the predictor and the Q value is dynamically adjusted within the prediction range; Phase cancellation involves generating a reversed or phase-shifted compensation signal based on predicted phase information and injecting it into the output path. Spatial domain control, through beamforming of microphone arrays or speaker units to spatially empty or suppress feedback direction; Low-amplitude probe sequences are injected into speech gaps to perform probe-based feedback path identification.

3. The active howling cancellation method based on acoustic signal prediction according to claim 1, characterized in that: The predictor is a temporal model or a hybrid of a temporal model and a rule engine. The temporal model is selected from the following: Small Long Short-Term Memory Network (LSTM), Causal Convolutional Network (TCN), One-Dimensional Transformer, or a quantized and lightweight model of One-Dimensional Transformer. The predictor has an inference latency of less than 10ms.

4. The active howling cancellation method based on acoustic signal prediction according to claim 2, characterized in that: The parameters of the variable Q narrowband adaptive notch filter satisfy: The center frequency ±1 / 2 frequency point is determined by the predictor output, the Q value is dynamically adjusted in the range of 15-60, the minimum bandwidth of the notch band does not exceed 3% of the center frequency, and the maximum duration of the notch amplitude is limited to avoid perceptible speech distortion. The probe sequence is a pseudo-random low-amplitude signal with an amplitude less than or equal to -30 dBFS of the system's maximum perceptible amplitude. It is injected only in the "gap" or low-energy segment where the voice activity VAD detection decision is made. The response after probe injection is used to construct or update the speaker-to-microphone transfer function, including approximations of nonlinear terms.

5. The active howling cancellation method based on acoustic signal prediction according to claim 1, characterized in that: Real-time performance evaluation includes at least two metrics: Short-time frequency band energy change rate and phase energy stability index; If both indicators show improvement, the intervention is considered effective; otherwise, the intervention intensity is increased according to the preset upgrade strategy, and a safety rollback strategy is triggered when necessary. The safety rollback strategy restores the system parameters to the last known set of stable parameters.

6. The active howling cancellation method based on acoustic signal prediction according to claim 1, characterized in that: The elimination method further includes estimating the direction of arrival (DOA) in a multi-microphone scenario to assist spatial domain control. The DOA estimation is based on short-time phase difference (TDOA) or array correlation methods, and the spatial domain supports beam vacancy reduction based on MVDR or minimum variance criteria.

7. A method and system for active howling cancellation based on acoustic signal prediction, applicable to the active howling cancellation method based on acoustic signal prediction as described in any one of claims 1-6, characterized in that: The elimination system includes: A microphone array or at least one microphone is used to acquire acoustic signals from the field. The front-end analog-to-digital converter (ADC) and the front-end preprocessing module are used for DC removal, window function processing, and baseband bandpass filtering. The acoustic feature extraction module is used to calculate the short-time spectrum, envelope rise rate, band energy, cross-correlation peak and phase stability features of the frame signal from the front-end preprocessing module. The predictor module is used to output a howling risk score, the predicted dominant frequency band, and the predicted phase information in real time, taking acoustic features as input. An active intervention controller connects to the predictor module and schedules the execution of intervention strategies in priority order when the risk score exceeds a preset threshold. The intervention strategies include an output gain fine-tuning unit, a variable Q narrowband notch filter unit, a phase cancellation generator, a beamforming vacancy reduction unit, and a probe injection unit. The feedback path identification module is used to construct or update the approximate model of the speaker-microphone transfer function and nonlinear terms under probe injection or blind identification conditions. The effect evaluation and rollback module is used to monitor short-time energy and phase stability in real time and perform parameter rollback when the stability criterion is not met. The storage and update module is used to record trigger events and training samples, and to support online or offline parameter updates for the predictor. The hardware platform is a combination of an audio DSP, an ARM processor, an FPGA, and an NN accelerator, ensuring that end-to-end inference and filtering delays are below a preset upper limit.

8. The active howling cancellation system based on acoustic signal prediction according to claim 7, characterized in that: The predictor module is implemented as a quantized or distilled lightweight neural network to meet real-time inference requirements. The system eliminates the need to deploy lightweight predictors and controllers locally, while working in collaboration with edge and cloud-based model training and parameter optimization platforms. Data interaction between local and cloud platforms only uploads de-identified statistical features and event summaries to protect privacy.

9. The active howling cancellation system based on acoustic signal prediction according to claim 7, characterized in that: The feedback path identification module includes a blind identification submodule and a probe-type identification submodule: the blind identification submodule is activated when there is continuous speech to continuously estimate the feedback path, and the probe-type identification submodule injects a probe briefly when a speech gap is detected and refines the transfer function based on the probe response. Meanwhile, the feedback path identification module supports the modeling and online updating of nonlinear terms.

10. The active howling cancellation system based on acoustic signal prediction according to claim 7, characterized in that: The active intervention controller implements a hierarchical priority strategy and a safety fallback mechanism. When the predictor indicates that the whistling risk has reached the first threshold, only low-intrusion actions are performed; when the risk continues to rise or the low-intrusion actions are ineffective, intermediate actions such as Q-switching and short-time phase cancellation are triggered in sequence, along with advanced actions such as beam vacancy reduction and probe identification, and the feedback model is updated accordingly; in any case, if the effect evaluation module does not detect stability improvement within the preset time window, it will automatically fall back to the last known stable parameters and generate log alarm information.

Images