Voice noise reduction method suitable for different noise environments

Abstract

The invention particularly relates to a voice noise reduction method suitable for different noise environments, and relates to the technical field of voice signal processing. Analyzing a noise environment and characteristics; adaptive selection of a noise reduction algorithm; adaptive noise reduction processing is executed; and noise reduction effect evaluation and iterative optimization are carried out. According to the invention, refined noise classification is matched with the algorithm, so that the limitation of a traditional single algorithm in a complex environment is solved; the steady-state noise is dynamically counteracted by adopting an adaptive filter, the frequency domain gain of the unsteady-state noise is adjusted in real time through Wiener filtering, and the pulse noise is subjected to dynamic threshold processing by combining median filtering and wavelet transform; especially, the design of cooperation factors in wavelet transform can accurately distinguish signal details and noise: intensively suppress pulse noise, loosely retain details such as voices and consonants, and realize dynamic balance of noise reduction intensity and signal distortion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of speech signal processing, and in particular to a speech noise reduction method suitable for different noise environments. BACKGROUND

[0002] The existing speech noise reduction technology has significant limitations in complex noise environments: First, the noise type recognition is rough, and the noise is generally divided into stationary and non-stationary. No special model is constructed for impulse noise (burst, short-time, wide-frequency characteristics), resulting in insufficient suppression of impulse interference such as keyboard tapping and door closing sound. Second, the algorithm selection lacks pertinence. For example, a single adaptive filter cannot cope with the time-varying characteristics of non-stationary noise, Wiener filtering is prone to speech distortion in impulse noise scenarios due to signal-to-noise ratio estimation bias, and the threshold of wavelet transform is usually a fixed value, which cannot be dynamically adjusted in combination with signal continuity and noise sparsity, and cannot balance noise reduction and detail preservation. Third, the dynamic adaptation capability is weak, and it is difficult to track the changes of noise type proportion, energy and spectrum fluctuation in real time. The algorithm parameters (such as step size and decomposition layer number) are adjusted with lag, and the performance drops sharply when mixed noise switches. Fourth, the optimization mechanism is incomplete. The effect evaluation mainly depends on a single objective index (such as signal-to-noise ratio), lacks subjective listening verification and closed-loop iteration, and leads to poor speech intelligibility and listening experience in actual scenarios.

[0003] Therefore, a speech noise reduction method suitable for different noise environments is needed to address the above-mentioned problems. SUMMARY

[0004] The purpose of the present application is to solve the above-mentioned problems by providing a speech noise reduction method suitable for different noise environments.

[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme: A speech noise reduction method suitable for different noise environments, comprising: Noise speech signal acquisition and preprocessing: obtaining a noisy speech signal and performing basic processing to reduce interference in subsequent analysis; Noise environment and characteristic analysis: determining the type, spectral characteristics and time variation of the current noise to provide a basis for subsequent algorithm selection; Adaptive noise reduction algorithm selection: selecting a corresponding noise reduction algorithm according to the noise characteristics and application scenarios; Adaptive noise reduction processing execution: applying the selected algorithm to the noisy speech signal to achieve real-time noise reduction and dynamically adapt to changes in the noise environment; Noise reduction effect evaluation and iterative optimization: verify the noise reduction effect, balance the noise reduction degree and speech distortion, and optimize the process according to the feedback.

[0006] Preferably, the noisy speech signal acquisition and preprocessing specifically includes: According to the noise environment, the corresponding pickup hardware is selected, the fixed noise source signal is acquired through a single pickup device, and the multi-source noise signal is acquired through a multi-pickup device array; The collected noise signal is processed, specifically including removing the direct current component, pre-filtering, and frame division and windowing.

[0007] Preferably, the noise environment and characteristic analysis specifically includes: Classification of noise types, including stationary noise, non-stationary noise and impulse noise; The classified noise is respectively analyzed and processed, including spectrum analysis, statistical characteristic analysis and time-varying characteristic analysis, wherein: The spectrum analysis is performed by short-time Fourier transform, FFT is performed on the frame signal to obtain a time-frequency diagram; The mean, variance and probability distribution of the noise are calculated, the change rate of the noise with time is analyzed by short-time Fourier transform, and a noise model is established based on the analysis result.

[0008] Preferably, the adaptive selection of the noise reduction algorithm specifically includes: The algorithms for stationary noise, non-stationary noise and impulse noise are respectively determined, wherein the algorithm for stationary noise is selected as an adaptive filter: by referring to the correlation between the reference noise and the noisy speech, the filter coefficient is dynamically adjusted, and the clean speech is output; The algorithm for non-stationary noise is selected as Wiener filter, according to the energy of the speech and the noise, by Calculating the gain, weighting the noisy speech to obtain the estimated clean speech ; The algorithm for impulse noise includes median filtering and wavelet change. The median filtering sets a sliding window on the time domain signal, sorts the signal values in the window, and then replaces the signal value at the center position of the window with the median value.

[0009] Preferably, the noise reduction process of the wavelet change algorithm includes: Selecting a suitable wavelet basis function and decomposition level, performing multi-scale wavelet decomposition on the noisy signal; Determining the threshold T for distinguishing signal coefficients and impulse noise coefficients; Assuming that the noise follows a Gaussian distribution in high frequency components, the noise standard deviation is estimated through the highest layer high frequency components, and then the threshold T is calculated; The threshold T calculation formula is ; is the noise standard deviation; is the signal length; For coordination factor; By dynamically adjusting the coordination factor, the threshold is changed accordingly; Based on the determined threshold, the wavelet coefficients are processed to suppress coefficients representing noise and retain coefficients representing useful signals. The high-frequency detail components and low-frequency approximation components, after thresholding, are recombined into a time-domain signal through inverse wavelet transform to obtain the denoised signal.

[0010] Preferably, the coordination factor The acquisition process includes: Based on the correlation of the coefficient neighborhood, for the first For each coefficient in the layer, calculate the correlation coefficient of its neighborhood; Statistics The average correlation of all coefficients in the layer is normalized to obtain the scale correlation coefficient.

[0011] Preferably, the method further includes: Based on the kurtosis of wavelet coefficients, obtain the first... The kurtosis of the high-frequency components of the layer is calculated, and the kurtosis coefficient is obtained by normalizing the kurtosis. The scale correlation coefficient and kurtosis coefficient are normalized and used as the major and minor semi-axes of the ellipse, respectively, to construct an elliptical model. The area of ​​the elliptical model is then calculated to obtain the coordination factor.

[0012] Preferably, the adaptive noise reduction processing includes: Real-time acquisition of noise features of the current frame, including: noise type; noise energy; and dynamic adjustment algorithm parameters for noise time-varying rate. Multi-algorithm collaborative scheduling is performed, machine learning is used to identify the environment and preload parameters, the dominant algorithm is switched according to the noise type, and temporary algorithms are inserted to suppress sudden impulse noise; Real-time performance is ensured through pipelined processing, and resources are dynamically allocated based on hardware load to ensure real-time and high-quality noise reduction under different noise environments.

[0013] Preferably, the noise reduction effect evaluation and iterative optimization specifically includes: The noise reduction effect is measured by calculating the signal-to-noise ratio, segmented signal-to-noise ratio, and noise attenuation, while the degree of speech distortion is assessed by short-time objective intelligibility and logarithmic spectral distance. Next, subjective listening verification is conducted by inviting professionals or ordinary users to listen blindly and score the audio, and collecting feedback on noise residue and unclear speech to locate defects. Iterative optimization is carried out based on the evaluation results. The algorithm parameters are adjusted in reverse according to objective data and subjective feedback. The algorithm is automatically adapted by using reinforcement learning to train the model. Collaborative processes are added or optimized for scenarios with poor performance. The optimized content is redeployed, and verification is performed again, forming a closed loop of evaluation, optimization and verification.

[0014] In summary, due to the adoption of the technical solutions described above, the application has the following beneficial effects: 1. The application solves the limitations of traditional single algorithms in complex environments through fine noise classification and algorithm adaptation. For steady-state noise, an adaptive filter is used for dynamic cancellation, for non-steady-state noise, a Wiener filter is used for real-time adjustment of frequency domain gain, and for impulse noise, a median filter and wavelet transform dynamic threshold processing are combined. In particular, the design of the factor in wavelet transform can accurately distinguish signal details and noise: it strengthens the inhibition of impulse noise and relaxes the retention of voice clear consonants and other details, achieving dynamic balance between noise reduction intensity and signal distortion.

[0015] 2. The application quantifies the noise reduction effect by using signal-to-noise ratio, segmented signal-to-noise ratio, etc. on the objective level, and evaluates the speech distortion by combining short-term objective intelligibility and log spectral distance. On the subjective level, it captures the human ear perception defects through blind listening feedback. Based on the evaluation results, the algorithm parameters are adjusted reversely, and the environment and parameter mapping model are trained using reinforcement learning, realizing the self-adaptation of the algorithm to various scenes. BRIEF DESCRIPTION OF DRAWINGS

[0016] In the following description of exemplary embodiments in conjunction with the accompanying drawings, more details, features and advantages of the application are disclosed, in which: Figure 1 The method structure diagram of the application. DETAILED DESCRIPTION

[0017] Several embodiments of the application will be described below in more detail with reference to the accompanying drawings so as to enable those skilled in the art to implement the application. The application can be embodied in many different forms and purposes and should not be limited to the embodiments described herein. These embodiments are provided to make the application comprehensive and complete, and to fully convey the scope of the application to those skilled in the art. The embodiments do not limit the application.

[0018] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and / or the present specification, and should not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

[0019] Example 1

[0020] The specific embodiments are combined with the accompanying drawings Figure 1 will be described in detail.

[0021] Attach Figure 1 A speech noise reduction method flowchart suitable for different noise environments is provided for the embodiments of the present application, which shows the complete steps from noisy speech signal collection and preprocessing to noise reduction effect evaluation and iterative optimization.

[0022] In the present embodiment, it comprises: Noisy speech signal collection and preprocessing: high-quality noisy speech signals are obtained, and basic processing is performed to reduce interference in subsequent analysis; Specifically, it comprises: According to the noise environment, the corresponding pickup hardware is selected, the fixed noise source signal is obtained through a single pickup device, and the multi-source noise signal is obtained through a multi-pickup device array; The single pickup device includes a microphone: the signal is collected through a single pickup point, the structure is simple, the cost is low, but it cannot distinguish the spatial direction; suitable scene: noise source is fixed (such as office air conditioner, computer fan), and the distance from the speech source is short (such as close-talking); Technical parameters: sensitivity (-38dB-42dB, too high to be saturated, too low to introduce noise), frequency response range (need to cover 300Hz~3400Hz speech core frequency band); The multi-pickup device array includes a microphone array: 2~16 microphones are arranged in a linear / circular manner, and the signals are separated by using the orientation difference between speech and noise (such as target speech coming from the front, noise coming from the side); suitable scene: multi-source noise (such as conference room discussion, restaurant background noise), complex environment (such as station, shopping mall); Core parameters: array aperture (affects spatial resolution, the larger the aperture, the stronger the ability to distinguish close sound sources), microphone spacing (usually 1~10cm, need to avoid spatial aliasing); The sampling rate needs to meet the Nyquist criterion (sampling rate ≥ 2 times the highest frequency of the signal). The speech core frequency band is 300Hz~3400Hz, so 8kHz can cover the basic requirement; if more details need to be preserved (such as music mixed speech), select 16kHz or 48kHz; 16bit can represent 65536 levels, the dynamic range is 96dB, which is enough to cover the speech signal (dynamic range is about 60dB); 24bit is suitable for high-fidelity scenes, but it will increase the data volume; Too high sampling rate will increase the computing burden, which needs to be balanced between signal integrity and processing efficiency (such as 8kHz is commonly used for real-time conversation, and 16kHz is used for post-recording); The collected noise signal is processed, which specifically includes removing the direct current component, pre-filtering, and framing and windowing; Remove DC component: The original signal may introduce a DC component (constant DC voltage) due to the microphone circuit offset, which appears as a shift in the overall baseline of the signal. Processing method: Calculate the long-term average of the signal (e.g. the average of the first 100ms of non-speech segments), and subtract this average from each sample point to ensure that the signal fluctuates around the 0 point; Pre-filtering (anti-aliasing and frequency band limitation): High-pass filtering: filter out low-frequency noise (<300Hz, such as mechanical vibration, footsteps), usually use a second-order Butterworth filter, cutoff frequency 300Hz, attenuation rate -40dB / octave; Low-pass filtering: filter out high-frequency noise (>8kHz, such as electromagnetic interference, equipment noise), to avoid high-frequency components aliasing into the speech frequency band, cutoff frequency 8kHz (matching 8kHz sampling rate); Frame division: divide the continuous signal into short frames (10~30ms), because the speech signal has "short-time stationarity" (the spectral characteristics are approximately constant within 10~30ms), which is convenient for frequency domain analysis. For example: under 8kHz sampling rate, 16ms frame length corresponds to 128 sample points (8000x0.016=128); Windowing: overlap between frames by 50%~75% (e.g. 16ms frame length, overlap 8ms), to avoid signal discontinuity at frame boundaries; windowing function (Hanning window, Hamming window) reduces the amplitude of the frame edges, reducing the spectral leakage after FFT (Hanning window has better suppression of leakage, Hamming window retains more DC components); Noise environment and characteristic analysis: determine the type, spectral characteristics and time variation of the current noise, to provide a basis for subsequent algorithm selection; Specifically includes: Classification of noise types, including stationary noise, non-stationary noise and impulse noise; Stationary noise: amplitude and spectral characteristics change <5% / second (slowly), statistical characteristics (mean, variance) are approximately constant; Typical examples: air conditioner / fan noise (low frequency 50~500Hz concentrated), computer host noise (wide frequency but energy stable), stable background human voice (e.g. constant volume of conversation in the distance).

[0023] Core features: smooth power spectral density (PSD) curve, no sudden peaks; Non-stationary noise: amplitude or spectral variation rate >10% / second, including random fluctuations or periodic changes (but the period is not fixed); it can be further divided into slow-changing non-stationary and fast-changing non-stationary; Slow-changing non-stationary: such as traffic flow noise (vehicle flow increases or decreases over time, frequency spectrum fluctuates in the range of 500~2000Hz); Fast non-stationary: such as multiple people talking (different people's voices alternate, spectrum quickly switches between 300~3400Hz), shopping mall background sound (mix of broadcast, footsteps, conversation, spectrum irregular fluctuation); Core features: PSD curve fluctuates significantly over time, with short energy peaks; Impulse noise: isolated signal with sudden, short (<100ms), high intensity (peak value exceeds background noise by more than 10dB); Typical examples: closing a door, typing on a keyboard, electric spark, car horn (single sound wave); Core features: sharp pulse in time domain, wide frequency distribution (energy from low to high frequency) in frequency domain; After classification, the noise is analyzed and processed, including spectral analysis, statistical characteristic analysis and time-varying characteristic analysis, where: Spectral analysis is done by short-time Fourier transform, which does FFT (frame length 1024 points, overlap 50%) on the framed signal to get a time-frequency diagram (time on the horizontal axis, frequency on the vertical axis, color representing energy); Calculate the mean, variance, and probability distribution of the noise (such as Gaussian distribution for stationary noise and heavy-tailed distribution for impulse noise); analyze the rate of change of the noise over time through short-time Fourier transform (such as the frequency of spectral fluctuations for non-stationary noise); Based on the analysis results, a noise model is established; For example: stationary noise can be modeled with a fixed power spectrum; non-stationary noise can be modeled with a time-varying power spectrum; and impulse noise can be modeled with a sparse impulse model; Adaptive selection of noise reduction algorithm: according to the characteristics of the noise and the application scenario (such as real-time communication, speech recognition, and recording noise reduction), select the corresponding noise reduction algorithm; Specifically includes: Determine the algorithm for stationary noise, non-stationary noise and impulse noise respectively, where the algorithm for stationary noise is an adaptive filter: by referring to the correlation between the noise and the noisy speech, dynamically adjust the filter coefficients to make the output approach the clean speech, the formula is: ; Where is the filter coefficient at the current time (a set of numbers that determine how to filter noise, such as the parameters in the noise reduction algorithm of a mobile phone); is the filter coefficient at the next time (adjusted according to the current error, new parameters to be used at the next moment); is the step size (adjustment strength, like a volume knob: too big is easy to overcorrect, too small is too slow to adjust); The error signal is the difference between noisy speech and the desired clean speech. For example, if the actual recorded speech is human voice and the sound of an air conditioner, but the desired speech is only human voice, the difference is the error. The reference signal is a specially collected piece of pure noise, such as ambient sound recorded separately by the noise-canceling mic of a headphone, used to tell the algorithm what the noise to be eliminated looks like. The energy of the reference signal (simply put, how loud the reference noise is, used to normalize and adjust the intensity to avoid loss of intensity control); The algorithm first compares the noisy speech with the signal processed by the current filter, and then calculates the difference. ; For example, if the actual recording includes a weak human voice and a strong air conditioner sound, the filter tries to preserve the human voice and remove the air conditioner sound, but the parameters are not adjusted correctly. As a result, the human voice is weakened, but the air conditioner sound remains. In this case, there is an error. It's very big; Using error Adjustment coefficient , set the current coefficient Add the "adjustment amount" to obtain the coefficient for the next time step. ; For non-steady-state noise, Wiener filtering was chosen as the algorithm, based on the energy of speech and noise, by... Calculate gain for noisy speech Weighted estimation of clean speech ; Achieving a balance between noise reduction and speech preservation hinges on decision-oriented estimation, utilizing... ,in It is the posterior signal-to-noise ratio; =0.9 ensures a smooth transition; It can estimate the prior signal-to-noise ratio in real time and dynamically adapt to noise changes. It is suitable for scenarios with slow noise changes, such as street traffic and coffee shops. It can balance noise reduction and speech preservation in speech recognition preprocessing. With a mathematically optimal compromise strategy, the filtering effect is both clean and natural. The algorithm for impulse noise includes median filtering and wavelet transform; Median filtering is a simple and efficient nonlinear filtering method. It involves setting a sliding window (e.g., window length of 3-7 points) for the time-domain signal, sorting the signal values ​​within the window, and then replacing the signal value at the center of the window with the median value. Since impulse noise manifests as short-duration, high-intensity outliers in the time domain, it is often located at both ends of the sequence after sorting. Through median replacement, isolated impulse interference can be removed quickly and effectively. Low computational complexity, easy to implement, and can quickly reduce the intensity of the pulse noise for subsequent processing. For example, in the process of collecting voice signals, the occasional keyboard click sound can be quickly weakened or eliminated by median filtering; The original noisy voice signal is preliminarily processed to remove most isolated and high-amplitude pulse noise, so that the pulse interference in the signal is preliminarily suppressed, and the signal faced by the subsequent wavelet transform is cleaner, reducing the processing difficulty; The noise reduction process of the wavelet transformation algorithm includes: Selecting a suitable wavelet base function and decomposition layer number to perform multi-scale wavelet decomposition on the noisy signal; During the decomposition process, the signal is decomposed into a low-frequency approximation component (retaining the main trend of the signal) and a high-frequency detail component (containing signal details and noise). The more layers, the higher the frequency resolution of the high-frequency component (can distinguish more subtle time-frequency features); Through multi-scale wavelet decomposition, the signal is split into low-frequency approximation components (retaining core trends) and high-frequency detail components (containing noise and signal details), and the decomposition layer number can be flexibly adjusted. The frequency resolution of the high-frequency component improves with the number of layers, allowing more detailed differentiation of signal characteristics and noise in different frequency bands. This hierarchical processing avoids one-size-fits-all noise reduction, allowing subsequent threshold processing to focus on frequency bands with high noise concentration and reduce interference with low-frequency core signals; Determine the threshold T for distinguishing between signal coefficients and pulse noise coefficients; Coefficients greater than T may be noise and need to be suppressed; coefficients less than T may be signal details and need to be retained; Assuming that the noise in the high-frequency component follows a Gaussian distribution (the sparse large coefficients of pulse noise can be considered as outliers), estimate the noise standard deviation from the highest layer high-frequency component (which is believed to contain almost only noise), and then calculate the threshold T; The threshold T calculation formula is ; is the noise standard deviation; is the signal length; is the cooperation factor; By dynamically adjusting the cooperation factor, the threshold value changes accordingly; The threshold T is calculated based on the noise standard deviation (accurately estimated through the highest-level high-frequency components, closely matching the actual noise level), signal length, and a dynamically adjusted coordination factor, rather than a fixed value. The design of the coordination factor is particularly crucial: the scale correlation coefficient reflects the continuity of signal details (avoiding misjudging continuous signals as noise), and the kurtosis coefficient identifies impulse noise (sparse, large coefficients in impulse noise will exhibit high kurtosis). The two are fused into a coordination factor through an elliptic model, allowing the threshold to dynamically change with the noise type (steady-state / impulse) and signal detail characteristics. The threshold is more stringent for impulse noise to suppress large coefficients, and more lenient for high-frequency components containing signal details to preserve details, effectively balancing noise reduction intensity and signal distortion. Based on a determined threshold, the wavelet coefficients are processed to suppress or remove coefficients representing noise and retain coefficients representing useful signals, thereby achieving the purpose of noise reduction. The high-frequency detail components and low-frequency approximation components after thresholding are recombined into a time-domain signal through inverse wavelet transform to obtain the denoised signal. Following the reverse process of wavelet decomposition, the processed components are reconstructed. In this process, since the noise coefficients have already been suppressed or removed in the thresholding step, the noise components in the reconstructed signal are greatly reduced, while the main features of the useful signal are preserved, thus achieving signal denoising. coordination factor The acquisition process includes: Based on the correlation of the coefficient neighborhood, for the first For each coefficient in a layer, calculate the correlation coefficient of its neighborhood (e.g., 3 coefficients to the left and 3 coefficients to the right); ; in: The scale index represents the level of signal decomposition (e.g., the level of wavelet decomposition). (high-frequency components of the layer), different j correspond to different frequencies; For position Signal value at the location / wavelet coefficients; The ability to distinguish between signals and noise is enhanced by leveraging neighborhood correlation: useful details in a signal (such as formants in speech or edges in an image) typically exhibit strong local continuity (high correlation of neighborhood coefficients), while noise (especially impulse noise) often shows a random distribution (low correlation of neighborhood coefficients). Calculating the neighborhood correlation coefficient for each coefficient directly quantifies this difference, providing a reliable basis for subsequently distinguishing between signal details to be preserved and noise to be suppressed. Statistics The average correlation of all coefficients in the layer is normalized to obtain the scale correlation coefficient. ; ; wherein: , are the minimum and maximum of all scale correlation coefficients, respectively; is the average correlation of all coefficients in the th layer; is the detail preservation strength, which is set to 0.3 by default; Different wavelet decomposition layers correspond to different frequency scales (high-frequency layers correspond to rapidly changing details, and low-frequency layers correspond to the main body of the signal). The signal and noise characteristics of each layer are significantly different. By statistically analyzing the average correlation of each layer and normalizing it (eliminating the magnitude difference of different levels and different signals), the scale correlation coefficient accurately reflects the signal and noise distribution characteristics of the frequency layer, allowing subsequent processing (such as threshold calculation) to adapt to the characteristics of each layer, avoiding misjudgment caused by using a uniform standard to process all frequencies (such as excessive noise reduction in high-frequency layers, which loses details, or insufficient noise reduction in low-frequency layers, which leaves residual noise); The setting of the detail preservation strength (such as the default 0.3) plays a weight adjustment role in the normalization process, preventing the sacrifice of signal details due to excessive emphasis on noise suppression. This ensures that the algorithm can reduce noise while not misjudging relevant but signal-related subtle features as noise, thereby reducing signal distortion (such as the clarity of speech and the loss of image texture); Based on the kurtosis of the wavelet coefficients, the kurtosis of the high-frequency components in the th layer is obtained, and the kurtosis is normalized to obtain the kurtosis coefficient; The kurtosis of the high-frequency components in the th layer is: ; is the i th wavelet coefficient in the th layer; The scale correlation coefficient and the kurtosis coefficient are normalized, and the scale correlation coefficient and the kurtosis coefficient are used as the major and minor semi-axes of an ellipse, respectively, to construct an elliptical model, calculate the area of the elliptical model, and obtain the cooperation factor; The kurtosis coefficient is used to capture the sparsity of impulse noise, the scale correlation coefficient is used to preserve the continuity of signal details, and the fusion logic of the elliptical model is used to make the algorithm capable of handling both Gaussian distributed steady noise and effectively suppressing sudden impulse noise (such as popping sounds and impact sounds), solving the problem of single threshold or single feature being unable to adapt to mixed noise scenarios. In complex and variable noise environments (such as background noise and sudden interference in voice communication), the algorithm can still maintain stable performance; The whole process focuses on preserving the details of useful signals: the low-frequency approximation component directly preserves the main trend of the signal, and the high-frequency component only suppresses the noise coefficient (not all high frequencies) through a dynamic threshold, and the dynamic adjustment of the factor further avoids the mis-suppression of signal details (such as the clear auxiliary in speech and the overtone in music). When the signal is reconstructed by the wavelet inverse transform, the noise can be greatly reduced while the integrity and detail characteristics of the original signal are maximally preserved, making the processed signal closer to the original signal in terms of hearing, recognition, and other aspects. Adaptive noise reduction processing is performed: the selected algorithm is applied to the noisy speech signal to achieve real-time or offline noise reduction and dynamically adapt to changes in the noise environment. Specifically, it includes: Real-time acquisition of noise characteristics of the current frame, including: noise type (proportion of change of steady-state / non-steady-state / pulse); noise energy (such as average amplitude of high-frequency component, energy proportion); noise time-varying rate (spectrum fluctuation frequency, amplitude change speed); dynamic adjustment of algorithm parameters; Adaptive filter real-time update step, adjustment of reference signal acquisition strategy, Wiener filter update of prior SNR weight and frequency domain gain smoothing coefficient, wavelet transform noise reduction adjustment of decomposition level and threshold function; Multi-algorithm collaborative scheduling is performed, machine learning is used to identify the environment and pre-load parameters, the dominant algorithm is switched according to the noise type, and temporary algorithms are inserted to suppress burst pulse noise; Real-time processing is ensured through pipeline processing, and resources are dynamically allocated according to hardware load to ensure that the algorithm can adapt flexibly and run efficiently in different noise environments, achieving real-time and high-quality noise reduction; By capturing the type proportion, energy distribution, and time-varying characteristics of the noise in real time, the algorithm can dynamically adjust the filter step, Wiener filter weight, and wavelet decomposition parameters, and match the current noise characteristics; For example, temporary suppression algorithms are quickly called for burst pulse noise, and smoothing coefficients are optimized for steady-state noise to avoid the problem of signal distortion or insufficient noise reduction caused by over-reduction of noise under fixed parameters, and to maintain high-quality noise reduction effect in complex and variable environments; Multi-algorithm collaborative scheduling combined with machine learning pre-loaded parameters greatly reduces the computational redundancy during environmental switching; pipeline processing and dynamic allocation of hardware load ensure that the algorithm can adapt flexibly to different noise environments in scenarios with strict real-time requirements (such as voice calls and real-time monitoring), without causing delays due to increased computational load, while achieving efficient operation and flexible adaptation; Whether it is steady-state noise (such as air conditioning sound), non-steady-state noise (such as traffic sound), or burst pulse noise (such as collision sound), this mechanism can be stably dealt with through dynamic parameter adjustment and algorithm switching, avoiding the limitations of a single algorithm in complex environments, allowing the noise reduction system to maintain stable and reliable performance in multiple fields such as speech processing, audio recording, and industrial monitoring, significantly improving the practical value of the technology; Noise reduction effect evaluation and iterative optimization: verify the noise reduction effect, balance the noise reduction degree and speech distortion, and optimize the process according to the feedback; Specifically includes: Calculate the signal-to-noise ratio, segmented signal-to-noise ratio, and noise attenuation to measure the noise reduction effect, and evaluate the speech distortion degree through short-term objective intelligibility and log spectral distance; Then perform subjective listening verification, invite professionals or ordinary users to listen blindly and score, collect feedback on noise residue, speech blurring, etc. to locate defects; Based on the evaluation results, iterative optimization is performed, algorithm parameters are adjusted based on objective data and subjective feedback, and the model is trained using reinforcement learning to achieve automatic adaptation of the algorithm, and new or optimized collaborative processes are added for scenes with poor performance; The optimized content is redeployed and verified again to form a closed loop of evaluation, optimization, and verification, continuously improving the processing effect of the noise reduction method in different noise environments; The signal-to-noise ratio, segmented signal-to-noise ratio, and other objective indicators can quantify the cleanliness of noise reduction, and the short-term objective intelligibility and log spectral distance can accurately measure the subtle differences in speech distortion, avoiding the ambiguity of relying solely on subjective experience. Subjective listening verification complements the limitations of objective indicators; For example, in some scenarios, the objective indicators meet the standards, but the human ear still feels unnatural. Through the blind listening feedback of professionals or users, we can locate the details of noise residue, speech blurring, etc. to make the evaluation both data-driven and practical experience-based; Based on the reverse adjustment of objective data and subjective feedback, targeted optimization can be achieved: for example, if the segmented signal-to-noise ratio shows that low-frequency noise reduction is insufficient, the low-frequency filter parameters can be specifically enhanced; if the user feedback is that the speech is muffled, the high-frequency attenuation coefficient can be reduced to restore the details. Reinforcement learning training model further improves the automation of optimization, allowing the algorithm to adapt to various scenarios independently, avoiding the experience dependence of manual parameter tuning; After each optimization, redeployment and verification can consolidate the effect in known scenarios and find potential problems in new scenarios (such as adaptation defects in extreme noise), ultimately allowing the system to maintain stable performance in steady-state, non-steady-state, and pulse noise environments, significantly improving the ability to go from the laboratory to practical application.

[0024] The above formulas are all de-dimensioned to calculate the numerical values, the formulas are obtained by collecting a large amount of data to simulate a formula of the most recent real situation, and preset parameters in the formulas are set by a person skilled in the art according to actual conditions.

[0025] The above merely describes certain exemplary embodiments of the present application by way of illustration, and it is needless to say that those skilled in the art can modify the described embodiments in various manners without departing from the spirit and scope of the present application. Therefore, the above drawings and descriptions are merely illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present application.

[0026] It should be noted that, in this document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0027] It should be understood that, in various embodiments of the present application, the size of the sequence number of each process described above does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0028] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0029] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0030] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiment scheme according to actual needs.

[0031] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit.

[0032] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited to this. Any skilled person in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0033] The above only describes some exemplary embodiments of the present application by way of illustration, and it is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present application. Therefore, the above figures and descriptions are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the present application.

Claims

1. A speech noise reduction method for different noise environments, characterized by, Comprise: Noise speech signal acquisition and preprocessing: obtain noisy speech signal and perform basic processing to reduce interference in subsequent analysis; Noise environment and characteristic analysis: determine the type, spectral characteristics and time variation law of the current noise, and provide basis for subsequent algorithm selection; Adaptive noise reduction algorithm selection: select the corresponding noise reduction algorithm according to the noise characteristics and application scene; Adaptive noise reduction processing execution: apply the selected algorithm to the noisy speech signal to realize real-time noise reduction and dynamically adapt to the change of noise environment; Noise reduction effect evaluation and iterative optimization: verify the noise reduction effect, balance the noise reduction degree and speech distortion, and optimize the process according to the feedback.

2. The speech noise reduction method for different noise environments according to claim 1, characterized in that, Noise speech signal acquisition and preprocessing, specifically including: According to the noise environment, select the corresponding pickup hardware, obtain the fixed noise source signal through a single pickup device, and obtain the multi-source noise signal through a multi-pickup device array; The collected noise signal is processed, specifically including removing the direct current component, prefiltering, and framing and windowing.

3. The speech noise reduction method for different noise environments according to claim 1, characterized in that, Noise environment and characteristic analysis, specifically including: Classification of noise types, including stationary noise, non-stationary noise and impulse noise; The classified noise is respectively analyzed and processed, including spectral analysis, statistical characteristic analysis and time-varying characteristic analysis, wherein: Spectral analysis is performed by short-time Fourier transform to obtain time-frequency diagram by FFT on framed signal; Calculate the mean, variance and probability distribution of noise; analyze the change rate of noise over time by short-time Fourier transform; establish a noise model based on the analysis results.

4. The speech noise reduction method for different noise environments according to claim 1, characterized in that, Adaptive noise reduction algorithm selection, specifically including: Determine the algorithm for stationary noise, non-stationary noise and impulse noise respectively, wherein the algorithm selection for stationary noise is adaptive filter: dynamically adjust the filter coefficients by referring to the correlation between the reference noise and the noisy speech, and output clean speech; The algorithm selection for the non-steady-state noise is the Wiener filter. According to the energy of the speech and the noise, the gain is calculated, and the noisy speech is weighted to obtain the estimated clean speech. ;​​ The algorithm for impulse noise includes median filtering and wavelet change. Median filtering replaces the signal value at the center of the window with the median value by setting a sliding window on the time domain signal and sorting the signal values within the window.

5. The speech noise reduction method for different noise environments according to claim 4, characterized in that, The noise reduction process of wavelet change algorithm includes: Selecting appropriate wavelet basis function and decomposition level, multi-scale wavelet decomposition is performed on the noisy signal; Determine the threshold T for distinguishing signal coefficients and impulse noise coefficients; Assuming that the noise follows Gaussian distribution in high frequency components, estimate the noise standard deviation from the highest layer high frequency components, and then calculate the threshold T; The threshold T is calculated by the formula ; is the noise standard deviation; is the signal length; is the fit factor; Adjust the threshold value by adjusting the cooperation factor; According to the determined threshold, process the wavelet coefficients to suppress the coefficients representing noise and retain the coefficients representing useful signals; The high frequency detail components and low frequency approximation components processed by the threshold are recombined into time domain signals by inverse wavelet transform to obtain the noise-reduced signal.

6. The speech noise reduction method for different noise environments according to claim 5, characterized in that, Cooperating factors The acquisition process comprises: Based on the correlation of the coefficient neighborhood, for each coefficient of the layer the correlation coefficients of its neighborhood are calculated; Statistics of the average correlation of all coefficients in a layer, normalized to give scale-dependent correlation coefficients. Statistics of the average correlation of all coefficients in a layer, normalized to give scale-dependent correlation coefficients.

7. The speech noise reduction method for different noise environments according to claim 6, characterized in that, Also includes: Based on the kurtosis of wavelet coefficients, the first The kurtosis of the high-frequency component of the layer is obtained, and the kurtosis is normalized to obtain a kurtosis coefficient; Normalize the scale correlation coefficient and kurtosis coefficient, and construct an elliptical model by taking the scale correlation coefficient and kurtosis coefficient as the major and minor axes of the ellipse respectively, calculate the area of the elliptical model, and obtain the cooperation factor.

8. The speech noise reduction method for different noise environments according to claim 1, wherein, Adaptive noise reduction processing execution, specifically including: Real-time acquisition of current frame noise characteristics, including: noise type; noise energy; noise time-varying rate dynamic adjustment algorithm parameters; Multi-algorithm collaborative scheduling is performed, machine learning is used to identify the environment and pre-load parameters, the dominant algorithm is switched according to the noise type, and a temporary algorithm is inserted to suppress burst pulse noise; Real-time performance is ensured through pipeline processing, resources are dynamically allocated according to hardware load, and real-time and high-quality noise reduction is achieved in different noise environments.

9. The speech noise reduction method for different noise environments according to claim 1, wherein, Noise reduction effect evaluation and iterative optimization, including: Calculate the signal-to-noise ratio, segmented signal-to-noise ratio, and noise attenuation to measure the noise reduction effect, and evaluate the speech distortion degree through short-term objective intelligibility and log spectral distance; Then perform subjective listening verification, invite professionals or ordinary users to listen blindly and score, collect feedback on noise residue and speech ambiguity to locate defects; Based on the evaluation results, iterative optimization is performed, algorithm parameters are adjusted based on objective data and subjective feedback, the model is trained using reinforcement learning to automatically adapt the algorithm, and new or optimized collaborative processes are added for scenes with poor performance; The optimized content is redeployed and verified again to form a closed loop of evaluation, optimization, and verification.

Images