A method and system for suppressing noise of audio signal

The noise characteristics are analyzed through Gaussian hybrid model and wavelet transform, combined with the integrated noise index selection processing strategy, efficient noise suppression in complex noise environments is achieved, and the quality of the audio signal is improved.

CN118782067BActive Publication Date: 2024-11-15SHENZHEN AXCEL TECH CO LTD
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411266179.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-11
Publication Date
2024-11-15
Estimated Expiration
2044-09-11

AI Technical Summary

Technical Problem

The prior art has poor noise suppression effect in complex noise environments, especially when the spectrum changes violently or the noise characteristics change greatly, it is difficult to adapt to the dynamically changing noise environment.

Method used

Audio signals are collected through audio sensors, noise statistical modeling is performed using Gaussian hybrid model, multi-scale analysis is performed by combining wavelet transformation, comprehensive noise index is obtained, and spectral noise processing or adaptive filtering and blind source separation algorithm are selected according to the index for noise suppression.

Benefits of technology

It improves the accuracy and adaptability of noise suppression, improves the clarity and quality of audio signals, and overcomes the shortcomings of traditional technologies in complex noise environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118782067B_ABST
    Figure CN118782067B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and system for suppressing noise of an audio signal, and relates to the technical field of audio processing. The method comprises: collecting audio signals of a target environment, and using a Gaussian mixture model to extract features and perform statistical modeling on noise, constructing a noise distribution statistical model, performing multi-scale analysis on the audio signal, capturing different frequency components, and dynamically adjusting the frequency components to effectively separate noise and target signals. By analyzing the stability, dynamics, and mixing degree in the noise distribution statistical model, a comprehensive noise index is obtained, and matched with a threshold interval. When the comprehensive noise index is in a first threshold interval, a spectrum noise processing technique is applied to filter out interfering audio signal noise; when the comprehensive noise index is in a second threshold interval, an adaptive filtering and blind source separation algorithm are used to further filter out interfering audio signal noise. The clarity and quality of audio signals in complex noise environments are improved to meet high-standard audio processing requirements.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of audio processing, and in particular to a method and system for suppressing noise of an audio signal. Background Art

[0002] With the acceleration of urbanization, environmental noise issues have become more prominent in modern society. Noise pollution caused by factors such as transportation, industry, construction and crowd activities has a negative impact on personal health and quality of life. Especially in city centers and busy public places, these noise sources often affect the clarity and quality of audio signals.

[0003] For example, a method and terminal device for reducing audio signal noise with announcement number CN111640447B includes: collecting a first audio signal; determining whether there is an interfering audio signal in the first audio signal that does not conform to a pre-stored user voice feature, wherein the pre-stored user voice feature includes: at least one of the timbre parameters of the user voice, the pitch parameters of the user voice, and the loudness parameters of the user voice; if there is an interfering audio signal that does not conform to the pre-stored user voice feature, filtering the interfering audio signal from the first audio signal to obtain a second audio signal, wherein the second audio signal is an audio signal of the user's voice. The method is applied to scenes with noisy sound environments.

[0004] However, this method has limitations in dealing with dynamic changes and diverse noise characteristics in complex noise environments. In particular, when the spectrum changes dramatically or the noise characteristics are variable, its noise suppression effect may be affected. Therefore, for more complex noise environments, more efficient and adaptable noise suppression technology is needed. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a method and system for suppressing noise in an audio signal, which solves the problems of the above-mentioned background technology.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for suppressing audio signal noise, comprising the following steps: S1. collecting audio signals of a target environment through an audio sensor, and extracting noise features, statistically modeling the noise through a Gaussian mixture model, and obtaining a noise distribution statistical model; S2. performing multi-scale analysis on the signal through wavelet transform, capturing different frequency components in the audio signal, and dynamically adjusting the frequency components of the audio signal; S3. analyzing the noise stability, noise dynamics, and noise mixing through a noise distribution statistical model, and obtaining a comprehensive noise index; S4. based on the comprehensive noise index, matching it with a set comprehensive noise index threshold interval, when the comprehensive noise index belongs to the first threshold interval, performing spectral noise processing on the audio signal to filter out interference audio signal noise, and when the comprehensive noise index belongs to the second threshold interval, filtering out interference audio signal noise through adaptive filtering and blind source separation algorithm.

[0007] Furthermore, the noise feature extraction of the audio signal of the target environment includes: converting the audio signal into a frequency domain signal by short-time Fourier transform to extract spectrum features; analyzing the time variation and spectrum distribution of the spectrum features to identify the noise components of the audio signal; and preliminarily dividing the frequency components in the audio signal into noise components and target audio signal components through a set frequency range based on the spectrum features and time variation characteristics.

[0008] Furthermore, the specific process of using the Gaussian mixture model to perform statistical modeling of noise is as follows: cluster analysis is performed on the extracted noise features to determine the preliminary distribution form of the noise; the spectral characteristics of the noise are modeled using the Gaussian mixture model, and the spectral distribution of the noise is a combination of multiple Gaussian distributions; the expectation-maximization algorithm is used for parameter estimation to optimize the mean, variance and weight coefficient of each Gaussian distribution to generate a noise distribution statistical model.

[0009] Furthermore, the specific process of using wavelet transform to perform multi-scale analysis on the signal, capture different frequency components in the audio signal, and dynamically adjust the frequency components of the audio signal is as follows: determine the scale number of wavelet decomposition through wavelet basis function analysis of the characteristics of the audio signal, identify the frequency characteristics and variation range of the audio signal; decompose the audio signal into detail coefficients and approximation coefficients of multiple scales through wavelet transform, perform threshold processing on the decomposed wavelet coefficients, remove the low-amplitude noise components of the audio signal, and retain the main audio signal information. The processed detail coefficients and approximation coefficients are synthesized through inverse wavelet transform to obtain the audio signal after noise suppression, retaining the main characteristics of the target audio signal.

[0010] Furthermore, the specific process of analyzing noise stability, noise dynamics and noise mixing through the noise distribution statistical model is as follows: through the mean and variance parameters in the noise distribution statistical model, the time mean and variance of the noise signal are calculated to obtain the noise stability index; by analyzing the changes of each Gaussian distribution in the noise distribution statistical model, the change trend of the mean and variance is identified to obtain the noise dynamics index; by calculating the Pearson correlation coefficient between Gaussian distributions, the correlation and superposition degree between each Gaussian distribution in the noise model are evaluated to obtain the noise mixing index.

[0011] Furthermore, the specific process of obtaining the comprehensive noise index is as follows: weights are assigned and weighted operations are performed on the noise stability index, the noise dynamics index and the noise mixing index respectively to obtain the comprehensive noise index.

[0012] Furthermore, the specific process of performing spectral noise processing on the audio signal and filtering out the noise interfering with the audio signal is as follows: when the comprehensive noise index belongs to the first threshold interval, performing short-time Fourier transform on the audio signal, converting the time domain signal into a frequency domain signal, and obtaining a spectrum representation; analyzing the noise components and the target signal components in the spectrum, identifying the frequency interval of the noise interfering with the audio signal according to the noise distribution statistical model, subtracting the noise spectrum components, and retaining the spectrum of the target signal; performing inverse Fourier transform on the processed spectrum, converting the frequency domain signal back to the time domain, and obtaining a denoised audio signal.

[0013] Furthermore, the interference audio signal noise is screened out through adaptive filtering and blind source separation as follows: when the comprehensive noise index belongs to the second threshold interval, the audio signal is estimated and suppressed in real time by adjusting the adaptive filter coefficient to generate a preliminary denoised audio signal; the preliminary denoised audio signal is input into a blind source separation algorithm based on deep learning, and the residual noise components are accurately identified and separated through feature extraction and pattern recognition. Finally, the denoised audio signal is reconstructed and enhanced using multi-level signal reconstruction technology to output a high-quality audio signal.

[0014] An audio signal noise suppression system comprises the following modules: a model building module, an audio signal adjustment module, a noise analysis module and a noise processing module; the model building module is used to collect audio signals of a target environment through an audio sensor, extract noise features, perform statistical modeling on the noise through a Gaussian mixture model, and obtain a noise distribution statistical model; the audio signal adjustment module is used to perform multi-scale analysis on the signal through wavelet transform, capture different frequency components in the audio signal, and dynamically adjust the frequency components of the audio signal; the noise analysis module is used to analyze noise stability, noise dynamics and noise mixing through a noise distribution statistical model, and obtain a comprehensive noise index; the noise processing module is used to match the comprehensive noise index with a set comprehensive noise index threshold range based on the comprehensive noise index, perform spectrum noise processing on the audio signal when the comprehensive noise index belongs to the first threshold range, and filter out the interference audio signal noise; when the comprehensive noise index belongs to the second threshold range, filter out the interference audio signal noise through adaptive filtering and blind source separation algorithm.

[0015] The present invention has the following beneficial effects:

[0016] (1) This method for suppressing noise in an audio signal can accurately describe the distribution characteristics of noise by using a Gaussian mixture model to statistically model the noise, thereby improving the accuracy of noise feature modeling. It uses wavelet transform to perform multi-scale analysis on the signal, which can capture different frequency components in the audio signal and dynamically adjust the frequency components, thereby solving the limitation that it is difficult to comprehensively analyze the frequency characteristics of the audio signal at a single scale, thereby improving the accuracy of noise suppression.

[0017] (2) This audio signal noise suppression system generates a comprehensive noise index by analyzing the stability, dynamics and mixing degree of noise. Based on the threshold interval matching of the comprehensive noise index, different noise processing strategies can be selected. This solves the problem that traditional noise processing technology cannot adapt to different noise complexities and provides a flexible and efficient processing solution. By combining spectral noise processing, adaptive filtering and blind source separation technologies, the clarity and quality of audio signals can be effectively improved, overcoming the problem that the existing technology is not effective in complex noise environments.

[0018] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The present invention is a flow chart of a method for suppressing noise in an audio signal.

[0020] Figure 2 The present invention is a flow chart of an audio signal noise suppression system. DETAILED DESCRIPTION

[0021] The embodiment of the present application provides an audio signal noise suppression method and system, which accurately models the noise through a Gaussian mixture model, performs multi-scale analysis of the audio signal through wavelet transform, comprehensively evaluates the stability, dynamics and mixing degree of the noise, and dynamically selects an appropriate noise processing strategy based on the comprehensive noise index, thereby solving the problems of inaccurate modeling, insufficient frequency component analysis, incomplete noise evaluation, inflexible processing strategy and unsatisfactory suppression effect of traditional noise suppression technology in complex noise environments.

[0022] The overall idea of ​​the problem in the embodiment of this application is as follows:

[0023] The audio signal of the target environment is collected through the audio sensor, and the noise features are extracted. The noise is statistically modeled through the Gaussian mixture model to obtain the noise distribution statistical model.

[0024] The signal is analyzed at multiple scales through wavelet transform to capture different frequency components in the audio signal and dynamically adjust the frequency components of the audio signal.

[0025] The noise stability, noise dynamics and noise mixing degree are analyzed through the noise distribution statistical model to obtain the comprehensive noise index.

[0026] Based on the comprehensive noise index, it is matched with the set comprehensive noise index threshold range. When the comprehensive noise index belongs to the first threshold range, the audio signal is spectrally processed to filter out the interference audio signal noise. When the comprehensive noise index belongs to the second threshold range, the interference audio signal noise is filtered out through adaptive filtering and blind source separation algorithm.

[0027] See also Figure 1 The embodiment of the present invention provides a technical solution: a method for suppressing audio signal noise, comprising the following steps: S1. collecting audio signals of a target environment through an audio sensor, extracting noise features, statistically modeling the noise through a Gaussian mixture model, and obtaining a noise distribution statistical model; S2. performing multi-scale analysis on the signal through wavelet transform, capturing different frequency components in the audio signal, and dynamically adjusting the frequency components of the audio signal; S3. analyzing the noise stability, noise dynamics, and noise mixing through the noise distribution statistical model, and obtaining a comprehensive noise index; S4. matching the comprehensive noise index with a set comprehensive noise index threshold interval based on the comprehensive noise index, performing spectral noise processing on the audio signal when the comprehensive noise index belongs to the first threshold interval, filtering out interference audio signal noise, and filtering out interference audio signal noise through adaptive filtering and blind source separation algorithm when the comprehensive noise index belongs to the second threshold interval.

[0028] In this embodiment, the Gaussian mixture model is a probabilistic model that assumes that the data is composed of the weighted sum of multiple Gaussian distributions (normal distributions). Each Gaussian distribution represents a potential component in the data. By modeling the Gaussian mixture of noise characteristics, complex noise distributions can be better described and analyzed so as to make accurate estimates and processing in the noise suppression process. The noise distribution statistical model is a mathematical model based on the statistical modeling of noise data, which is used to describe the distribution characteristics of noise. The noise distribution statistical model established by the Gaussian mixture model can provide the distribution of noise in different frequency ranges, so as to effectively process it in the subsequent noise suppression. Wavelet transform is a signal processing method for decomposing a signal into different scales and frequency components. Unlike Fourier transform, wavelet transform can provide time and frequency information at the same time, so that the local characteristics of the signal can be analyzed in detail. In noise suppression, wavelet transform is used for multi-scale analysis to capture different frequency components in the audio signal, so as to dynamically adjust the frequency components of the signal. The comprehensive noise index is a quantitative index calculated based on the noise distribution statistical model, which is used to comprehensively evaluate the stability, dynamics and mixing degree of noise. It reflects the complexity of noise in time and frequency. The calculation of the comprehensive noise index helps to select the appropriate noise processing strategy and take corresponding processing measures according to different noise characteristics. Spectral noise processing is the process of identifying and suppressing noise by analyzing the spectrum of the audio signal. It includes using spectral subtraction technology to subtract noise components from the audio signal and retain useful signal components. This method improves the noise suppression effect by processing the signal in the frequency domain. Adaptive filtering is a technology that dynamically adjusts filter parameters to adapt to changes in signals and noise. The filter updates its parameters in real time according to the characteristics of the input signal to optimally remove noise and retain the target signal. This method is particularly effective in processing noise that changes over time and in the environment. Blind source separation is a signal processing technique used to extract the original source signal from multiple mixed signals without knowing the specific information of the mixing process in advance. It is used to separate different sound sources in multi-channel audio signals, thereby more effectively removing interfering audio signal noise.

[0029] Specifically, the noise feature extraction of the audio signal of the target environment includes: converting the audio signal into a frequency domain signal through short-time Fourier transform to extract the spectrum feature; analyzing the time variation and spectrum distribution of the spectrum feature to identify the noise component of the audio signal; and preliminarily dividing the frequency components in the audio signal into noise components and target audio signal components through a set frequency range based on the spectrum feature and time variation characteristics.

[0030] In this embodiment, the audio signal is converted from the time domain to the frequency domain signal by short-time Fourier transform. This means that the time signal is decomposed into its frequency components so that the frequency information of the signal can be analyzed and processed. STFT is a commonly used spectrum analysis tool that can describe the signal in both the time and frequency domains. In the frequency domain, the spectrum characteristics of the audio signal are extracted. The spectrum characteristics reflect the intensity distribution of each frequency component in the signal. The purpose of this step is to obtain the frequency distribution of the audio signal in order to identify the noise and target signal therein. By analyzing the time-varying nature of the spectrum characteristics (i.e., the characteristics of the signal spectrum changing over time) and the spectrum distribution, the noise component in the audio signal is identified. Time-varying analysis helps understand the frequency changes of the signal at different time points and identify which frequency components are noise and which are target audio signals. Based on the previous analysis results, by setting a specific frequency range, the frequency components in the audio signal are preliminarily divided into noise components and target audio signal components. This step is the basis of noise suppression, and by clearly distinguishing between noise and target signals, it lays the foundation for further noise suppression. The short-time Fourier transform formula is as follows, which is used to convert the audio signal from the time domain to the frequency domain:

[0031] in, Indicates at time The spectrum below, represents the original audio signal, represents the window function, is the angular frequency, Represents the complex exponential function in Fourier transform, which is used to convert time domain signals into frequency domain signals. k represents the center moment of the current analysis window, and m represents the time variable, which represents the sampling point of the audio signal in the time domain.

[0032] Specifically, the specific process of using the Gaussian mixture model to perform statistical modeling of noise is as follows: cluster analysis is performed on the extracted noise features to determine the preliminary distribution form of the noise; the spectral characteristics of the noise are modeled using the Gaussian mixture model, and the spectral distribution of the noise is a combination of multiple Gaussian distributions; the expectation maximization algorithm is used for parameter estimation to optimize the mean, variance and weight coefficient of each Gaussian distribution to generate a noise distribution statistical model.

[0033] In this embodiment, cluster analysis is performed on the extracted noise features. Cluster analysis is an unsupervised machine learning method for dividing data points into different groups according to their similarities. Here, through cluster analysis, the noise features can be divided into several categories to determine the preliminary distribution form of the noise. This process helps to understand the different components of the noise and provides a basis for subsequent modeling. The spectral characteristics of the noise are modeled using a Gaussian mixture model. Multiple expectation maximization algorithms are used for parameter estimation: In order to optimize the various Gaussian distributions in the Gaussian mixture model, the parameters of each Gaussian distribution (including mean, variance, and weight coefficient) need to be estimated. The expectation maximization algorithm is a commonly used iterative algorithm for estimating these parameters. The algorithm continuously optimizes the parameters of each Gaussian distribution by repeatedly calculating expectations and maximization, so that the model better fits the spectral characteristics of the noise. Through the above steps, a noise distribution statistical model is finally generated. This model contains the spectral characteristics of the noise and its combination between different Gaussian distributions, and can accurately describe the distribution characteristics of the noise. This statistical model can be used for subsequent noise identification and suppression processes. This process uses cluster analysis and Gaussian mixture model to perform detailed statistical modeling on the spectral characteristics of the noise, generating a statistical model that can reflect the distribution characteristics of the noise. This model plays a key role in the noise suppression algorithm and can improve the accuracy of noise identification and processing. The Gaussian mixture model formula is as follows: ,in, represents the probability density function of noise, which indicates the possibility of noise occurring at a certain frequency. z represents the noise spectrum feature vector, which indicates the specific performance of noise in the frequency domain. M represents the number of Gaussian distributions, that is, the number of mixed Gaussian distributions in the Gaussian mixture model. represents the weight coefficient of the i-th Gaussian distribution, which represents the proportion of the Gaussian distribution in the overall model and satisfies , Represents the i-th Gaussian distribution, with the mean as parameter and the covariance matrix , describing the shape of the distribution. The expectation maximization algorithm, ;in, Represents the jth data point The responsibility value belonging to the i-th Gaussian component represents the data point The probability from the ith Gaussian distribution, represents the noise feature vector of the jth data point, represents the weight coefficient of the i-th Gaussian distribution, Represents the probability density value of the i-th Gaussian distribution, applied to the data point . in, represents the updated mean of the ith Gaussian distribution, recalculated based on the current responsibility value, represents the noise feature vector of the jth data point, Represents the jth data point is the responsibility value belonging to the i-th Gaussian component, and N represents the total number of data points. ,in, Represents the covariance matrix of the i-th Gaussian distribution, indicating the degree and direction of expansion of the Gaussian distribution. Represents data points Relative to the mean deviation. Represents the transposed matrix of the above deviations.

[0034] Specifically, the specific process of using wavelet transform to perform multi-scale analysis on the signal, capturing different frequency components in the audio signal, and dynamically adjusting the frequency components of the audio signal is as follows: analyzing the characteristics of the audio signal through wavelet basis functions to determine the scale number of wavelet decomposition, and identifying the frequency characteristics and variation range of the audio signal; decomposing the audio signal into detail coefficients and approximation coefficients of multiple scales through wavelet transform, performing threshold processing on the decomposed wavelet coefficients, removing low-amplitude noise components of the audio signal, and retaining the main audio signal information; synthesizing the processed detail coefficients and approximation coefficients through inverse wavelet transform to obtain an audio signal after noise suppression, and retaining the main characteristics of the target audio signal.

[0035] In this embodiment, the characteristics of the audio signal are analyzed using wavelet basis functions. Wavelet basis functions are mathematical functions that can decompose a signal into components of different frequencies and times. By analyzing the characteristics of the audio signal, the appropriate number of wavelet decomposition scales is determined, that is, how many different frequency scales the signal is decomposed into. This step helps to identify the frequency characteristics and variation range of the audio signal, so that different frequency components can be captured more accurately in subsequent processing. The audio signal is decomposed by wavelet transform. Wavelet transform is a signal processing technology that can decompose a signal into detail coefficients and approximation coefficients of different scales. The detail coefficient contains high-frequency information (such as edges and mutations) in the signal, while the approximation coefficient contains low-frequency information (such as smooth parts) in the signal. This decomposition method allows the multi-scale characteristics of the signal to be analyzed, which is convenient for signal processing in different frequency ranges. After the wavelet transform decomposition, the wavelet coefficients obtained by the decomposition are threshold processed. Threshold processing is a noise suppression technology that can remove low-amplitude noise components in the signal while retaining the main audio signal information. By setting a threshold, coefficients (usually noise) less than the threshold are suppressed or removed, thereby reducing the impact of noise on the signal. The processed detail coefficients and approximation coefficients are resynthesized into a complete signal through inverse wavelet transform. Inverse wavelet transform is the inverse process of wavelet transform, which is used to reassemble the decomposed signal into an approximate form of the original signal. After the processed signal is inversely transformed, a noise-suppressed audio signal is obtained, which retains the main features of the target audio signal. The wavelet transform formula is as follows: ,in, Represents the wavelet transform coefficient, indicating the signal at scale and location The transformation result is: Represents the scale parameter, which controls the width of the wavelet; larger The value corresponds to the low-frequency component, and the smaller The value corresponds to the high frequency component, represents the translation parameter, which controls the position of the wavelet on the time axis. represents the original input audio signal, Represents the mother wavelet function, which is used to extract local features in the signal. Represents the complex conjugate function of the mother wavelet. The threshold processing formula is as follows, ,in, represents the wavelet coefficient after threshold processing, : The original wavelet coefficients, representing the signal at a specific scale and location frequency components. Represents the wavelet coefficient threshold, which is used to decide whether to retain a wavelet coefficient and is set to an estimated value of the noise level.

[0036] Specifically, the specific process of analyzing noise stability, noise dynamics and noise mixing through the noise distribution statistical model is as follows: through the mean and variance parameters in the noise distribution statistical model, the time mean and variance of the noise signal are calculated to obtain the noise stability index; by analyzing the changes of each Gaussian distribution in the noise distribution statistical model, the change trend of the mean and variance is identified to obtain the noise dynamics index; by calculating the Pearson correlation coefficient between Gaussian distributions, the correlation and superposition degree between each Gaussian distribution in the noise model are evaluated to obtain the noise mixing index.

[0037] In this embodiment, the noise stability index is calculated as follows: the parameters of the Gaussian mixture model (GMM) are obtained, including the mean of each Gaussian distribution ,variance( ), time mean Indicates that the mean value of the noise is calculated within a certain time window, reflecting the center position of the noise signal. It indicates the variance of the noise calculated in the same time window, reflecting the degree of noise fluctuation. The noise stability index expresses the stability of the noise through the inverse of the time variance, and the formula is as follows: , where SI represents the noise stability index. The higher the stability index, the more stable the noise. Indicates the variance of the noise calculated in the same time window. The calculation process of the noise dynamic index is as follows: obtain the Gaussian distribution change data of the noise model in different time periods, including the change of mean and variance. ) indicates the calculation of the change amplitude of the mean value in different time windows, and the change point detection algorithm is used to identify the mutation points and fluctuations in the noise signal. The calculation formula of the noise dynamics index is as follows, which represents the dynamics of the noise by the change amplitude and mutation of the noise. The formula is as follows: ,in, represents the noise dynamics index, and W represents the number of mutation points detected in the noise signal. The higher the dynamics index, the more significant the dynamic changes of the noise. Indicates the change range of the mean value calculated in different time windows. The formula for calculating the mixing index is as follows. The mixing index (MI) represents the mixing degree of noise through the weighted average of correlation and superposition degree. The formula is as follows: , where MI represents the noise mixing index, is the correlation coefficient between the i-th and j-th Gaussian distributions, and N is the total number of Gaussian distributions. A higher mixing index indicates a higher degree of mixing between noise components.

[0038] Specifically, the specific process of obtaining the comprehensive noise index is as follows: weights are assigned and weighted operations are performed on the noise stability index, the noise dynamics index and the noise mixing index respectively to obtain the comprehensive noise index.

[0039] In this embodiment, the calculation process of obtaining the comprehensive noise index is as follows: each noise index is multiplied by its corresponding weight, and a weighted operation is performed to calculate the comprehensive noise index (CNI).

[0040] The formula is as follows: , where CNI stands for the comprehensive noise index, which is an index for comprehensively evaluating the overall characteristics of noise and reflects the overall impact of noise. SI stands for the noise stability index, which reflects the volatility of the noise signal. DI stands for the noise dynamics index, which indicates the degree of dynamic change of the noise signal. MI stands for the noise mixing index, which describes the mixing and correlation between noise components. represents the noise stability index weight coefficient, represents the noise dynamics index weight coefficient, Represents the noise mixing index weight coefficient, which indicates the relative importance of each noise index in the comprehensive noise index.

[0041] Specifically, the specific process of performing spectral noise processing on the audio signal and filtering out the noise that interferes with the audio signal is as follows: perform short-time Fourier transform on the audio signal, convert the time domain signal into a frequency domain signal, and obtain a spectral representation; analyze the noise components and target signal components in the spectrum, identify the frequency interval of the noise that interferes with the audio signal according to the noise distribution statistical model, subtract the noise spectrum components, and retain the spectrum of the target signal; perform inverse Fourier transform on the processed spectrum, convert the frequency domain signal back to the time domain, and obtain the denoised audio signal. The audio signal is converted from the time domain to the frequency domain through short-time Fourier transform. This step decomposes the audio signal into different frequency components and calculates its corresponding spectral representation in each time window.

[0042] In this embodiment, by converting the time domain signal into the frequency domain signal, the various frequency components in the audio signal can be observed and analyzed more clearly, so as to identify which frequencies contain noise and which frequencies contain the target signal. The spectrum of the audio signal is analyzed in the frequency domain to distinguish the noise component and the target signal component. According to the data in the noise distribution statistical model, it can be determined which frequency intervals mainly contain noise components. By identifying the noise frequency interval, targeted noise suppression operations can be performed to ensure that the target signal is retained to the greatest extent during the processing. The spectrum components of the identified noise frequency interval are reduced or removed. The specific method is to subtract these noise frequency components from the spectrum of the audio signal, so that only the spectrum components of the target signal are retained. By removing the components of the noise frequency interval, the interference noise in the audio signal can be significantly reduced, making the final signal clearer. Inverse Fourier transform is to perform inverse Fourier transform on the spectrum after noise processing, convert the frequency domain signal back to the time domain, and obtain the final denoised audio signal. Inverse Fourier transform restores the frequency domain information to the time domain signal, thereby outputting the processed audio signal with suppressed noise.

[0043] Specifically, the interference audio signal noise is screened out through adaptive filtering and blind source separation as follows: when the comprehensive noise index belongs to the second threshold interval, the audio signal is estimated and suppressed in real time by adjusting the adaptive filter coefficients to generate a preliminary denoised audio signal; the preliminary denoised audio signal is input into a blind source separation algorithm based on deep learning, and the residual noise components are accurately identified and separated through feature extraction and pattern recognition. Finally, the denoised audio signal is reconstructed and enhanced using multi-level signal reconstruction technology to output a high-quality audio signal.

[0044] In this embodiment, the noise level in the current audio signal is determined according to the calculated comprehensive noise index. If the comprehensive noise index falls into the preset second threshold interval, it indicates that there is a more complex noise interference in the audio signal, and a more advanced noise suppression technology needs to be adopted. By judging the comprehensive noise index, it is decided whether to enable complex noise suppression means, such as adaptive filtering and blind source separation. The coefficients of the adaptive filter are adjusted so that it can track the noise characteristics in the audio signal in real time and perform dynamic noise estimation and suppression. The adaptive filter adapts to the changes in noise by continuously updating its internal parameters, thereby generating a preliminary denoised audio signal. The use of the adaptive filter can suppress the noise components in the audio signal in real time, improve the initial clarity of the signal, and provide a purer input signal for subsequent blind source separation. The preliminary denoised signal generated by the adaptive filter is input into the blind source separation algorithm based on deep learning. The blind source separation algorithm performs feature extraction and pattern recognition through a deep learning model, and can accurately identify the residual noise components in the audio signal. By separating these noise components, they are removed from the target audio signal. The blind source separation algorithm further removes the residual noise components after adaptive filtering through the powerful feature extraction capability of the deep learning model, thereby improving the purity of the audio signal. After the noise components are separated and removed, the denoised audio signal is reconstructed through multi-level signal reconstruction technology. During the reconstruction process, the signal is enhanced to make up for the details that may be lost during the denoising process. The final output audio signal not only removes the noise, but also undergoes reconstruction and enhancement to ensure the high fidelity and clarity of the audio signal, achieving the goal of high-quality audio output.

[0045] See also Figure 2 , an audio signal noise suppression system, comprising the following modules: a model building module, an audio signal adjustment module, a noise analysis module, and a noise processing module; the model building module is used to collect audio signals of a target environment through an audio sensor, and perform noise feature extraction, perform statistical modeling on the noise through a Gaussian mixture model, and obtain a noise distribution statistical model; the audio signal adjustment module is used to perform multi-scale analysis on the signal through wavelet transform, capture different frequency components in the audio signal, and dynamically adjust the frequency component of the audio signal; the noise analysis module is used to analyze noise stability, noise dynamics, and noise mixing through a noise distribution statistical model, and obtain a comprehensive noise index; the noise processing module is used to match the comprehensive noise index with a set comprehensive noise index threshold range based on the comprehensive noise index, and when the comprehensive noise index belongs to the first threshold range, perform spectrum noise processing on the audio signal to filter out the interference audio signal noise, and when the comprehensive noise index belongs to the second threshold range, filter out the interference audio signal noise through adaptive filtering and blind source separation algorithm.

[0046] In this implementation scheme, the model building module is responsible for collecting audio signals of the target environment through audio sensors and extracting noise features from these signals. Then, the Gaussian mixture model (GMM) is used to statistically model the noise and generate a noise distribution statistical model. These statistical models can be used for subsequent noise analysis and processing. The model building module is the basis of the system, which aims to accurately model the noise characteristics in the target environment and provide data support for subsequent noise analysis and processing. The audio signal adjustment module uses wavelet transform to perform multi-scale analysis on the audio signal to capture components of different frequencies in the signal. By dynamically adjusting the frequency components of the audio signal, the signal quality is optimized, making subsequent noise analysis and processing more effective. By analyzing the frequency characteristics of the audio signal, this module can effectively separate components of different frequencies, providing a basis for the system to accurately capture and adjust the frequency components of the audio signal. The noise analysis module uses the previously constructed noise distribution statistical model to conduct a detailed analysis of the stability, dynamics and mixing degree of the noise, and finally obtains a comprehensive noise index. This index reflects the complexity of the current noise environment and provides a basis for the selection of noise suppression strategies. This module is responsible for evaluating the current noise environment and providing a quantitative indicator (comprehensive noise index) so that the system can select an appropriate noise processing method. The noise processing module matches the comprehensive noise index provided by the noise analysis module with the preset threshold interval. When the comprehensive noise index falls into the first threshold interval, the system performs spectral noise processing on the audio signal to filter out the noise. When the comprehensive noise index falls into the second threshold interval, the system enables adaptive filtering and blind source separation algorithms for more complex noise suppression processing. This module is the execution unit of the system, which determines the specific noise suppression method and selects different processing methods according to the complexity of the noise environment to ensure the best audio signal quality.

[0047] In summary, this application has at least the following effects:

[0048] A method and system for suppressing noise in an audio signal collects audio signals in a target environment, uses a Gaussian mixture model to extract features and perform statistical modeling on the noise, constructs a noise distribution statistical model, performs multi-scale analysis on the audio signal, captures different frequency components, and dynamically adjusts the frequency components to effectively separate noise and target signals. By analyzing the stability, dynamics, and mixing degree in the noise distribution statistical model, a comprehensive noise index is obtained and matched with a threshold interval. When the comprehensive noise index is in the first threshold interval, a spectrum noise processing technique is used to filter out interfering audio signal noise; when the comprehensive noise index is in the second threshold interval, an adaptive filtering and blind source separation algorithm are used to further filter out interfering audio signal noise. The clarity and quality of audio signals in complex noise environments are improved to meet high-standard audio processing requirements.

[0049] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0050] The present invention is described with reference to flowcharts and / or block diagrams of systems, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0051] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0052] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0053] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0054] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for suppressing noise in an audio signal, characterized in that: The following steps are involved: S1. Collect the audio signal of the target environment through the audio sensor, extract the noise feature, perform statistical modeling on the noise through the Gaussian mixture model, and obtain the noise distribution statistical model; S2. Perform multi-scale analysis on the signal through wavelet transform, capture different frequency components in the audio signal, and dynamically adjust the frequency components of the audio signal; S3. Analyze the noise stability, noise dynamics and noise mixing through the noise distribution statistical model to obtain the comprehensive noise index; S4. Based on the comprehensive noise index, the comprehensive noise index threshold interval is matched. When the comprehensive noise index belongs to the first threshold interval, the audio signal is subjected to spectral noise processing to filter out the interference audio signal noise. When the comprehensive noise index belongs to the second threshold interval, the interference audio signal noise is filtered out by adaptive filtering and blind source separation algorithm; The specific process of analyzing noise stability, noise dynamics and noise mixing by using the noise distribution statistical model is as follows: The time mean and variance of the noise signal are calculated through the mean and variance parameters in the noise distribution statistical model to obtain the noise stability index; By analyzing the changes of each Gaussian distribution in the noise distribution statistical model, the change trend of the mean and variance is identified, and the noise dynamic index is obtained; By calculating the Pearson correlation coefficient between Gaussian distributions, the correlation and superposition degree between Gaussian distributions in the noise model are evaluated to obtain the noise mixing index.

2. The method for suppressing audio signal noise according to claim 1, characterized in that: Noise feature extraction of the audio signal of the target environment includes: The audio signal is converted into a frequency domain signal through short-time Fourier transform to extract the spectrum features; Analyze the time-varying and spectral distribution of spectral features and identify the noise components of audio signals; According to the spectral characteristics and time-varying characteristics, the frequency components in the audio signal are preliminarily divided into noise components and target audio signal components through the set frequency range.

3. The method for suppressing audio signal noise according to claim 2, characterized in that: The specific process of using Gaussian mixture model to statistically model noise is as follows: Perform cluster analysis on the extracted noise features to determine the preliminary distribution form of the noise; The spectral characteristics of noise are modeled using a Gaussian mixture model, where the spectral distribution of noise is a combination of multiple Gaussian distributions; The expectation maximization algorithm is used for parameter estimation to optimize the mean, variance and weight coefficient of each Gaussian distribution and generate a noise distribution statistical model.

4. The method for suppressing audio signal noise according to claim 3, characterized in that: The specific process of using wavelet transform to perform multi-scale analysis on the signal, capture different frequency components in the audio signal, and dynamically adjust the frequency components of the audio signal is as follows: Analyze the characteristics of audio signals through wavelet basis functions to determine the scale number of wavelet decomposition and identify the frequency characteristics and variation range of audio signals; The audio signal is decomposed into detail coefficients and approximation coefficients of multiple scales through wavelet transform, and the decomposed wavelet coefficients are threshold processed to remove the low-amplitude noise components of the audio signal and retain the main audio signal information; The processed detail coefficients and approximation coefficients are synthesized through inverse wavelet transform to obtain the audio signal after noise suppression, retaining the main features of the target audio signal.

5. The method for suppressing audio signal noise according to claim 4, characterized in that: The specific process of obtaining the comprehensive noise index is as follows: The noise stability index, noise dynamics index and noise mixing index are respectively assigned weights and weighted operations are performed to obtain a comprehensive noise index.

6. The method for suppressing audio signal noise according to claim 5, characterized in that: The specific process of performing spectrum noise processing on the audio signal and filtering out the noise that interferes with the audio signal is as follows: When the comprehensive noise index belongs to the first threshold interval, short-time Fourier transform is performed on the audio signal to convert the time domain signal into a frequency domain signal to obtain a spectrum representation; Analyze the noise components and target signal components in the spectrum, identify the frequency interval of the noise interfering with the audio signal according to the noise distribution statistical model, subtract the noise spectrum components, and retain the spectrum of the target signal; Perform inverse Fourier transform on the processed spectrum, convert the frequency domain signal back to the time domain, and obtain the denoised audio signal.

7. The method for suppressing audio signal noise according to claim 6, characterized in that: The specific method of filtering out the interference audio signal noise through adaptive filtering and blind source separation is as follows: When the comprehensive noise index belongs to the second threshold interval, the audio signal is estimated and suppressed in real time by adjusting the adaptive filter coefficient to generate a preliminary denoised audio signal; The preliminary denoised audio signal is input into a blind source separation algorithm based on deep learning. Through feature extraction and pattern recognition, the residual noise components are accurately identified and separated. Finally, the denoised audio signal is reconstructed and enhanced using multi-level signal reconstruction technology to output a high-quality audio signal.

8. An audio signal noise suppression system, applied to an audio signal noise suppression method according to any one of claims 1 to 7, characterized in that: It includes the following modules: model building module, audio signal adjustment module, noise analysis module, and noise processing module; The model building module is used to collect audio signals of the target environment through an audio sensor, extract noise features, perform statistical modeling on the noise through a Gaussian mixture model, and obtain a noise distribution statistical model; The audio signal adjustment module is used to perform multi-scale analysis on the signal through wavelet transform, capture different frequency components in the audio signal, and dynamically adjust the frequency components of the audio signal; The noise analysis module is used to analyze the noise stability, noise dynamics and noise mixing through the noise distribution statistical model to obtain the comprehensive noise index; The noise processing module is used to match the comprehensive noise index with the set comprehensive noise index threshold range based on the comprehensive noise index. When the comprehensive noise index belongs to the first threshold range, the audio signal is subjected to spectrum noise processing to filter out the interference audio signal noise. When the comprehensive noise index belongs to the second threshold range, the interference audio signal noise is filtered out by adaptive filtering and blind source separation algorithm. The specific process of analyzing noise stability, noise dynamics and noise mixing by using the noise distribution statistical model is as follows: The time mean and variance of the noise signal are calculated through the mean and variance parameters in the noise distribution statistical model to obtain the noise stability index; By analyzing the changes of each Gaussian distribution in the noise distribution statistical model, the change trend of the mean and variance is identified, and the noise dynamic index is obtained; By calculating the Pearson correlation coefficient between Gaussian distributions, the correlation and superposition degree between Gaussian distributions in the noise model are evaluated to obtain the noise mixing index.

Citation Information

Patent Citations

  • A method and terminal device for reducing audio signal noise

    CN111640447B

  • Layered adaptive denoising method and system for home noise environment

    CN111477241A

  • Adaptive noise reduction method based on scene perception, terminal and server

    CN118230751A

  • Noise suppressor, noise suppression method, and program

    JP2015036769A

  • Method and apparatus for improved estimation of non-stationary noise for speech enhancement

    US20070055508A1