A noise suppression method based on differential microphone array

Abstract

The present invention relates to the technical field of noise suppression, and discloses a noise suppression method based on a differential microphone array, the method comprising: calculating the difference signal between different sound source signals, performing signal feature extraction on multi-channel difference sound source signals, and obtaining the difference signal feature vectors of different sound source signals; constructing a low-frequency noise determination model to perform low-frequency noise determination on the difference signal feature vectors; and performing low-frequency noise suppression processing on the sound source signal with low-frequency noise. The present invention calculates the difference signal between different sound source signals, characterizes the signal change characteristics between different channels, performs dimensionality reduction processing on the difference signal feature vectors in combination with weight coefficients of different dimensions, performs feature mapping on the difference signal dimensionality reduction feature vectors using training parameters, realizes the determination of low-frequency noise, performs channel position filtering and signal filtering on the sound source signal with low-frequency noise, removes the low-frequency noise information therein, and realizes low-frequency noise suppression in the microphone array.

Description

Technical Field

[0001] The present invention relates to the technical field of noise suppression, and in particular to a noise suppression method based on a differential microphone array. Background Art

[0002] With the continuous development of voice technology, especially its wide application in human-computer interaction, intelligent assistants and communication devices, the performance of noise suppression directly affects the experience quality of the device. Traditional noise suppression methods, such as single-microphone noise suppression technology, usually rely on the analysis and filtering of the statistical characteristics of noise and speech signals, but their performance is limited by the complexity of the noise environment and the non-stationarity of the signal. In order to overcome the limitations of the single-microphone system, multi-microphone array technology has been widely used. Microphone arrays can enhance the signal processing capabilities through spatial information, which can not only improve the signal-to-noise ratio (SNR) of the signal, but also realize functions such as source localization and spatial filtering. Among them, differential microphone arrays have become an important tool in noise suppression research due to their simple structure, low cost and good spatial resolution. The differential microphone array uses the small spatial distance between multiple microphones to spatially differentiate the input signal, thereby selectively responding to the sound source signals in different directions. This technology can effectively suppress noise from the interference direction without relying on specific sound field modeling, while enhancing the speech signal from the target direction. Because the differential array has strong directional selectivity for noise, it has obvious advantages in noise suppression in noisy environments.

[0003] However, the differential microphone array has a good effect on suppressing high-frequency noise, but a relatively poor effect on suppressing low-frequency noise (such as air conditioning noise, traffic noise, etc.). This is because the wavelength of low-frequency noise is long, and the structure size of the differential array is limited, which cannot effectively perform spatial filtering on low-frequency signals. In existing research, the suppression effect of low-frequency noise is usually improved by increasing the number of microphones or adjusting the array structure, but these methods often increase the complexity and cost of the system, and the effect is limited. Summary of the invention

[0004] In view of this, the present invention provides a noise suppression method based on a differential microphone array, which can effectively filter low-frequency noise and improve the noise suppression effect.

[0005] To achieve the above object, the present invention provides a noise suppression method based on a differential microphone array, comprising the following steps:

[0006] S1: Collecting multi-channel original sound source signals and performing differential processing to obtain sound source signals and calculate difference signals between different sound source signals, and forming multi-channel difference sound source signals from the difference signals;

[0007] S2: extracting signal features from multi-channel difference sound source signals to obtain difference signal feature vectors of different sound source signals;

[0008] S3: construct a low-frequency noise determination model to perform low-frequency noise determination on the feature vector of the difference signal, and obtain a determination result of whether the sound source signal has low-frequency noise, wherein the low-frequency noise determination adopts an improved support vector machine algorithm;

[0009] S4: performing low-frequency noise suppression processing on the sound source signal with low-frequency noise, outputting the sound source signal after the low-frequency noise suppression and the sound source signal without the low-frequency noise, and converting the sound source signal into the original sound source signal.

[0010] As a further improvement method of the present invention:

[0011] Optionally, the step S1 of collecting multi-channel sound source signals and calculating difference signals between different sound source signals includes:

[0012] Collect multi-channel sound source signals from microphone array:

[0013] ;

[0014] in:

[0015] represents the sound source signal of the mth channel, M represents the number of channels in the microphone array; wherein the sound source signal is represented in the form of a differential signal;

[0016] Indicates the sound source signal The differential signal value at N signal moments, Indicates the sound source signal The differential signal value at the nth signal moment, ,Right now is the difference between the signal value of the original sound source signal at the n+1th signal time and the signal value at the nth signal time;

[0017] The sound source signal of the m-th channel is the sound signal captured by the m-th microphone in the microphone array, and the number of channels in the microphone array is the number of microphones in the microphone array;

[0018] Calculate the difference signal between different sound source signals, where the sound source signal and The difference signal between them is:

[0019] ;

[0020] in:

[0021] Indicates the sound source signal and The difference signal between ;

[0022] Indicates the difference signal The nth differential signal value in .

[0023] Optionally, the step of forming a multi-channel difference sound source signal from the difference signal includes:

[0024] The difference signal is used to form a multi-channel difference sound source signal of different sound source signals, where the sound source signal The multi-channel difference sound source signal is:

[0025] ;

[0026] in:

[0027] Indicates the sound source signal The multi-channel difference sound source signal The sound source signal A signal matrix composed of difference signals between the sound source signals of M channels;

[0028] T stands for transpose.

[0029] Optionally, the step S2 of extracting signal features from the multi-channel difference sound source signal includes:

[0030] The signal feature extraction is performed on the multi-channel difference sound source signal, and the feature extraction result is used as the difference signal feature vector of the sound source signal, wherein the multi-channel difference sound source signal The signal feature extraction process is as follows:

[0031] S21: Calculate the multi-channel difference sound source signal Center frequency :

[0032] ;

[0033] in:

[0034] Indicates the number of signal moments per second;

[0035] S22: Calculate the multi-channel difference sound source signal Kurtosis :

[0036] ;

[0037] ;

[0038] ;

[0039] in:

[0040] Indicates the difference signal The kurtosis component of

[0041] Indicates the difference signal The standard deviation of the N differential signal values ​​in ;

[0042] Indicates the difference signal The mean of the N differential signal values ​​in ;

[0043] S23: Calculate the multi-channel difference sound source signal The power spectral density :

[0044] ;

[0045] ;

[0046] in:

[0047] Indicates the difference signal The power spectral density component of

[0048] represents an exponential function with a natural constant as base;

[0049] j represents the imaginary unit, ;

[0050] S24: Calculate the multi-channel difference sound source signal Sound pressure ;

[0051] S25: Multi-channel difference sound source signal Center frequency , Kurtosis , power spectral density And sound pressure As a result of feature extraction, the sound source signal is formed The difference signal eigenvector : .

[0052] Optionally, the multi-channel difference sound source signal is calculated in step S24 Sound pressure ,include:

[0053] S241: Calculate and obtain multi-channel difference sound source signal The acoustic power of different difference signals in The sound power is :

[0054] ;

[0055] S242: Convert the sound power of the difference signal into a sound pressure component of the difference signal, wherein the difference signal The sound pressure component is :

[0056] ;

[0057] in:

[0058] represents the air density, c represents the speed of sound;

[0059] represents the distance between the sound source position and the mth microphone position; in the embodiment of the present invention, the sound source position is the main podium position;

[0060] S243: Constructing a multi-channel difference sound source signal from the sound pressure components Sound pressure : .

[0061] Optionally, constructing a low-frequency noise determination model in step S3 includes:

[0062] Constructing a low-frequency noise determination model, wherein the low-frequency noise determination model includes an input layer, a feature reconstruction layer, a feature mapping layer, and a determination layer, wherein the input layer is used to receive a difference signal feature vector of a sound source signal, the feature reconstruction layer is used to perform dimensionality reduction and reconstruction processing on the difference signal feature vector to obtain a difference signal dimensionality reduction feature vector, the feature mapping layer is used to perform feature mapping on the difference signal dimensionality reduction feature vector using training parameters to obtain a mapping feature vector, and the determination layer is used to perform low-frequency noise determination on the mapping feature vector and output a determination result;

[0063] The difference signal dimension reduction feature vector of the sound source signal of group D and the actual judgment result of whether there is low-frequency noise are collected to construct a training data set :

[0064] ;

[0065] in:

[0066] Represents the dimension reduction feature vector of the difference signal of the dth group of sound source signals, Indicates the true judgment result of whether the dth group of sound source signals has low-frequency noise. It means that there is no low-frequency noise in the sound source signal of group d. It means that there is low-frequency noise in the sound source signal of group d;

[0067] Based on the training data set The model parameters are optimized and solved, and the optimization process is as follows:

[0068] S31: Initialize and generate U groups of Lagrange multipliers with a length of D, where the u-th group of Lagrange multipliers generated by initialization is :

[0069] ;

[0070] Set the current iteration number of the Lagrange multiplier to t, the initial value of t is 0, and the maximum value is Max. Then the t-th iteration result of the u-th group of Lagrange multipliers is ;

[0071] S32: Based on the training data set Construct a fitness function with Lagrange multipliers as input , where the Lagrange multiplier The corresponding fitness function value is :

[0072] ;

[0073] in:

[0074] represents the L1 norm;

[0075] S33: The U groups of Lagrange multipliers obtained in the t-th iteration are used as the input values ​​of the fitness function in sequence, and the Lagrange multiplier with the largest fitness function value is used as the optimal value obtained in the t-th iteration ;

[0076] S34: Iterate the Lagrange multiplier, where the iteration formula is:

[0077] ;

[0078] ;

[0079] in:

[0080] Represents a random number between 0 and 1;

[0081] express The fitness probability of

[0082] represents any Lagrange multiplier in the U group of Lagrange multipliers obtained at the t-th iteration;

[0083] S35: Let t = t + 1, return to step S33, until the maximum number of iterations is reached, and extract the Lagrange multiplier with the maximum fitness function value at this time ;

[0084] S36: Using the extracted Lagrange multipliers , calculate the training parameters:

[0085] ;

[0086] ;

[0087] in:

[0088] are the training parameters in the feature map layer.

[0089] Optionally, the using of a low-frequency noise determination model to perform low-frequency noise determination on the feature vector of the difference signal to obtain a determination result of whether the sound source signal has low-frequency noise includes:

[0090] The input layer receives the difference signal feature vector of the sound source signal;

[0091] The feature reconstruction layer performs dimensionality reduction and reconstruction on the feature vector of the difference signal, where the feature vector of the difference signal The dimension reduction and reconstruction processing formula is:

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] in:

[0098] Represents the difference signal feature vector The corresponding difference signal dimension reduction feature vector;

[0099] T stands for transpose;

[0100] represents the L2 norm;

[0101] Indicates the difference signal Dimensionality reduction features;

[0102] Indicates normalization processing;

[0103] represents the standard deviation of all kurtosis components after normalization, represents the mean of all kurtosis components after normalization;

[0104] represents the standard deviation of all power spectral density components after normalization, Represents the mean of all power spectral density components after normalization;

[0105] They are the weight coefficients of the kurtosis component, the power spectrum density component, and the sound pressure component respectively;

[0106] The feature mapping layer uses the training parameters to perform feature mapping on the difference signal dimension reduction feature vector to obtain the mapping feature vector, where the difference signal dimension reduction feature vector The feature mapping formula is:

[0107] ;

[0108] in:

[0109] The dimension reduction feature vector of the difference signal The mapped feature vector obtained by mapping;

[0110] The decision layer performs low-frequency noise decision on the mapped feature vector. If the mapped feature vector is lower than a preset threshold, the corresponding sound source signal has low-frequency noise. Otherwise, the corresponding sound source signal does not have low-frequency noise.

[0111] Optionally, in step S4, performing low-frequency noise suppression processing on the sound source signal with low-frequency noise includes:

[0112] The low-frequency noise suppression process is performed on the sound source signal with low-frequency noise, where the sound source signal The low-frequency noise suppression processing flow is:

[0113] S41: Get the sound source signal x that is closest to the mth channel and does not have low-frequency noise:

[0114] ;

[0115] in:

[0116] Represents the differential signal value of the sound source signal x at the nth signal moment;

[0117] S42: Based on the sound source signal x, the sound source signal with low-frequency noise Double noise reduction filtering is performed, in which the sound source signal Medium differential signal value The double noise reduction filter processing formula is:

[0118] ;

[0119] in:

[0120] Indicates the differential signal value The double noise reduction filter processing result;

[0121] Indicates the sound source signal The distance between the channel corresponding to x;

[0122] represents the distance filter scale, represents the signal filtering scale; in the embodiment of the present invention, Set to 1, Set to 2;

[0123] Composing sound source signal The low-frequency noise suppression processing result:

[0124] ;

[0125] in:

[0126] Indicates the sound source signal The low-frequency noise suppression processing result.

[0127] In order to solve the above problem, the present invention provides an electronic device, the electronic device comprising:

[0128] A memory storing at least one instruction;

[0129] Communication interface, enabling electronic equipment to communicate; and

[0130] The processor executes the instructions stored in the memory to implement the above-mentioned noise suppression method based on differential microphone array.

[0131] In order to solve the above problem, the present invention also provides a computer-readable storage medium, in which at least one instruction is stored. The at least one instruction is executed by a processor in an electronic device to implement the above-mentioned noise suppression method based on a differential microphone array.

[0132] Compared with the prior art, the present invention proposes a noise suppression method based on a differential microphone array, which has the following advantages:

[0133] First, this scheme proposes a signal feature extraction method, which uses a differential array to extract the trend information and dynamic characteristics of the original sound source signal in different channels. The signal after differential processing is used as the sound source signal, and the difference signal between different sound source signals is calculated to characterize the signal change characteristics between different channels. The difference signal is processed by extracting signal characteristics such as power spectral density and sound pressure to obtain the sound source signal between different channels and the signal change characteristics in different dimensions, thereby improving the detection accuracy of low-frequency noise.

[0134] At the same time, this scheme eliminates a low-frequency noise detection and suppression method, adopts an improved support vector machine algorithm to build a low-frequency noise determination model, combines weight coefficients of different dimensions to reduce the dimension of the difference signal feature vector, uses a dual method to construct a fitness function and trains and optimizes to obtain training parameters, uses the training parameters to perform feature mapping on the reduced-dimensional feature vector of the difference signal to achieve low-frequency noise determination, performs channel position filtering and signal filtering on the sound source signal with low-frequency noise to remove the low-frequency noise information, and uses an inverse difference method to restore the sound source signal after removing the low-frequency noise information to the original sound source signal, thereby achieving low-frequency noise suppression in the microphone array. BRIEF DESCRIPTION OF THE DRAWINGS

[0135] Figure 1 A schematic diagram of a flow chart of a noise suppression method based on a differential microphone array provided by an embodiment of the present invention.

[0136] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0137] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0138] The embodiment of the present application provides a noise suppression method based on a differential microphone array. The execution subject of the noise suppression method based on a differential microphone array includes but is not limited to at least one of the electronic devices such as a server and a terminal that can be configured to execute the method provided in the embodiment of the present application. In other words, the noise suppression method based on a differential microphone array can be executed by software or hardware installed in a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc.

[0139] A noise suppression method based on a differential microphone array comprises the following steps:

[0140] S1: Collect original sound source signals of multiple channels and perform differential processing to obtain sound source signals and calculate difference signals between different sound source signals, and form multi-channel difference sound source signals from the difference signals.

[0141] The step S1 includes collecting multi-channel sound source signals and calculating difference signals between different sound source signals, including:

[0142] Collect multi-channel sound source signals from microphone array:

[0143] ;

[0144] in:

[0145] represents the sound source signal of the mth channel, M represents the number of channels in the microphone array; wherein the sound source signal is represented in the form of a differential signal;

[0146] Indicates the sound source signal The differential signal value at N signal moments, Indicates the sound source signal The differential signal value at the nth signal moment, ,Right now is the difference between the signal value of the original sound source signal at the n+1th signal time and the signal value at the nth signal time;

[0147] The sound source signal of the m-th channel is the sound signal captured by the m-th microphone in the microphone array, and the number of channels in the microphone array is the number of microphones in the microphone array;

[0148] Calculate the difference signal between different sound source signals, where the sound source signal and The difference signal between them is:

[0149] ;

[0150] in:

[0151] Indicates the sound source signal and The difference signal between ;

[0152] Indicates the difference signal The nth differential signal value in .

[0153] The step of forming a multi-channel difference sound source signal from the difference signal comprises:

[0154] The difference signal is used to form a multi-channel difference sound source signal of different sound source signals, where the sound source signal The multi-channel difference sound source signal is:

[0155] ;

[0156] in:

[0157] Indicates the sound source signal The multi-channel difference sound source signal The sound source signal A signal matrix composed of difference signals between the sound source signals of M channels;

[0158] T stands for transpose.

[0159] S2: Extract signal features from multi-channel difference sound source signals to obtain difference signal feature vectors of different sound source signals.

[0160] The step S2 extracts signal features from the multi-channel difference sound source signal, including:

[0161] The signal feature extraction is performed on the multi-channel difference sound source signal, and the feature extraction result is used as the difference signal feature vector of the sound source signal, wherein the multi-channel difference sound source signal The signal feature extraction process is as follows:

[0162] S21: Calculate the multi-channel difference sound source signal Center frequency :

[0163] ;

[0164] in:

[0165] Indicates the number of signal moments per second;

[0166] S22: Calculate the multi-channel difference sound source signal Kurtosis :

[0167] ;

[0168] ;

[0169] ;

[0170] in:

[0171] Indicates the difference signal The kurtosis component of

[0172] Indicates the difference signal The standard deviation of the N differential signal values ​​in ;

[0173] Indicates the difference signal The mean of the N differential signal values ​​in ;

[0174] S23: Calculate and obtain multi-channel difference sound source signal The power spectral density :

[0175] ;

[0176] ;

[0177] in:

[0178] Indicates the difference signal The power spectral density component of

[0179] represents an exponential function with a natural constant as base;

[0180] j represents the imaginary unit, ;

[0181] S24: Calculate the multi-channel difference sound source signal Sound pressure ;

[0182] S25: Multi-channel difference sound source signal Center frequency , Kurtosis , power spectral density And sound pressure As a result of feature extraction, the sound source signal is formed The difference signal eigenvector : .

[0183] The multi-channel difference sound source signal is calculated in step S24 Sound pressure ,include:

[0184] S241: Calculate and obtain multi-channel difference sound source signal The acoustic power of different difference signals in The sound power is :

[0185] ;

[0186] S242: Convert the sound power of the difference signal into a sound pressure component of the difference signal, wherein the difference signal The sound pressure component is :

[0187] ;

[0188] in:

[0189] represents the air density, c represents the speed of sound;

[0190] represents the distance between the sound source position and the mth microphone position; in the embodiment of the present invention, the sound source position is the main podium position;

[0191] S243: Constructing a multi-channel difference sound source signal from the sound pressure components Sound pressure : .

[0192] S3: Construct a low-frequency noise determination model to perform low-frequency noise determination on the feature vector of the difference signal, and obtain a determination result of whether the sound source signal has low-frequency noise.

[0193] The low-frequency noise determination model is constructed in step S3, including:

[0194] Constructing a low-frequency noise determination model, wherein the low-frequency noise determination model includes an input layer, a feature reconstruction layer, a feature mapping layer, and a determination layer, wherein the input layer is used to receive a difference signal feature vector of a sound source signal, the feature reconstruction layer is used to perform dimensionality reduction and reconstruction processing on the difference signal feature vector to obtain a difference signal dimensionality reduction feature vector, the feature mapping layer is used to perform feature mapping on the difference signal dimensionality reduction feature vector using training parameters to obtain a mapping feature vector, and the determination layer is used to perform low-frequency noise determination on the mapping feature vector and output a determination result;

[0195] The difference signal dimension reduction feature vector of the sound source signal of group D and the actual judgment result of whether there is low-frequency noise are collected to construct a training data set :

[0196] ;

[0197] in:

[0198] Represents the dimension reduction feature vector of the difference signal of the dth group of sound source signals, Indicates the true judgment result of whether the dth group of sound source signals has low-frequency noise. It means that there is no low-frequency noise in the sound source signal of group d. It means that there is low-frequency noise in the sound source signal of group d;

[0199] Based on the training data set The model parameters are optimized and solved, and the optimization process is as follows:

[0200] S31: Initialize and generate U groups of Lagrange multipliers with a length of D, where the u-th group of Lagrange multipliers generated by initialization is :

[0201] ;

[0202] Set the current iteration number of the Lagrange multiplier to t, the initial value of t is 0, and the maximum value is Max. Then the t-th iteration result of the u-th group of Lagrange multipliers is ;

[0203] S32: Based on the training data set Construct a fitness function with Lagrange multipliers as input , where the Lagrange multiplier The corresponding fitness function value is :

[0204] ;

[0205] in:

[0206] represents the L1 norm;

[0207] S33: The U groups of Lagrange multipliers obtained in the t-th iteration are used as the input values ​​of the fitness function in sequence, and the Lagrange multiplier with the largest fitness function value is used as the optimal value obtained in the t-th iteration ;

[0208] S34: Iterate the Lagrange multiplier, where the iteration formula is:

[0209] ;

[0210] ;

[0211] in:

[0212] Represents a random number between 0 and 1;

[0213] express The fitness probability of

[0214] represents any Lagrange multiplier in the U group of Lagrange multipliers obtained at the t-th iteration;

[0215] S35: Let t = t + 1, return to step S33, until the maximum number of iterations is reached, and extract the Lagrange multiplier with the maximum fitness function value at this time ;

[0216] S36: Using the extracted Lagrange multipliers , calculate the training parameters:

[0217] ;

[0218] ;

[0219] in:

[0220] are the training parameters in the feature map layer.

[0221] The low-frequency noise determination model is used to perform low-frequency noise determination on the feature vector of the difference signal to obtain a determination result of whether the sound source signal has low-frequency noise, including:

[0222] The input layer receives the difference signal feature vector of the sound source signal;

[0223] The feature reconstruction layer performs dimensionality reduction and reconstruction on the feature vector of the difference signal, where the feature vector of the difference signal The dimension reduction and reconstruction processing formula is:

[0224] ;

[0225] ;

[0226] ;

[0227] ;

[0228] ;

[0229] in:

[0230] Represents the difference signal feature vector The corresponding difference signal dimension reduction feature vector;

[0231] T stands for transpose;

[0232] represents the L2 norm;

[0233] Indicates the difference signal Dimensionality reduction features;

[0234] Indicates normalization processing;

[0235] represents the standard deviation of all kurtosis components after normalization, represents the mean of all kurtosis components after normalization;

[0236] represents the standard deviation of all power spectral density components after normalization, Represents the mean of all power spectral density components after normalization;

[0237] They are the weight coefficients of the kurtosis component, the power spectrum density component, and the sound pressure component respectively;

[0238] The feature mapping layer uses the training parameters to perform feature mapping on the difference signal dimension reduction feature vector to obtain the mapping feature vector, where the difference signal dimension reduction feature vector The feature mapping formula is:

[0239] ;

[0240] in:

[0241] The dimension reduction feature vector of the difference signal The mapped feature vector obtained by mapping;

[0242] The decision layer performs low-frequency noise decision on the mapped feature vector. If the mapped feature vector is lower than a preset threshold, the corresponding sound source signal has low-frequency noise. Otherwise, the corresponding sound source signal does not have low-frequency noise.

[0243] S4: performing low-frequency noise suppression processing on the sound source signal with low-frequency noise, outputting the sound source signal after the low-frequency noise suppression and the sound source signal without the low-frequency noise, and converting the sound source signal into the original sound source signal.

[0244] In the step S4, the low-frequency noise suppression process is performed on the sound source signal having low-frequency noise, including:

[0245] The low-frequency noise suppression process is performed on the sound source signal with low-frequency noise, where the sound source signal The low-frequency noise suppression processing flow is:

[0246] S41: Get the sound source signal x that is closest to the mth channel and does not have low-frequency noise:

[0247] ;

[0248] in:

[0249] Represents the differential signal value of the sound source signal x at the nth signal moment;

[0250] S42: Based on the sound source signal x, the sound source signal with low-frequency noise Double noise reduction filtering is performed, in which the sound source signal Medium differential signal value The double noise reduction filter processing formula is:

[0251] ;

[0252] in:

[0253] Indicates the differential signal value The double noise reduction filter processing result;

[0254] Indicates the sound source signal The distance between the channel corresponding to x;

[0255] represents the distance filter scale, Indicates the signal filtering scale;

[0256] Composing sound source signal The low-frequency noise suppression processing result:

[0257] ;

[0258] in:

[0259] Indicates the sound source signal The low-frequency noise suppression processing result.

[0260] It should be understood that the embodiment is for illustration only and the scope of the patent application is not limited to this structure.

[0261] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0262] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0263] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A noise suppression method based on a differential microphone array, characterized in that: The method comprises: S1: Collecting multi-channel original sound source signals and performing differential processing to obtain sound source signals and calculate difference signals between different sound source signals, and forming multi-channel difference sound source signals from the difference signals; S2: extracting signal features from multi-channel difference sound source signals to obtain difference signal feature vectors of different sound source signals; S3: construct a low-frequency noise determination model to perform low-frequency noise determination on the feature vector of the difference signal, and obtain a determination result of whether the sound source signal has low-frequency noise; S4: performing low-frequency noise suppression processing on the sound source signal with low-frequency noise, outputting the sound source signal after the low-frequency noise suppression and the sound source signal without the low-frequency noise, and converting the sound source signal into the original sound source signal.

2. A noise suppression method based on a differential microphone array as claimed in claim 1, characterized in that: The S1 step includes: Collect the original sound source signals of multiple channels in the microphone array and perform differential processing to obtain the sound source signals of multiple channels: ； in: represents the sound source signal of the mth channel, M represents the number of channels in the microphone array; wherein the sound source signal is represented in the form of a differential signal; Indicates the sound source signal The differential signal value at N signal moments, Indicates the sound source signal The differential signal value at the nth signal moment, ; The sound source signal of the m-th channel is the sound signal captured by the m-th microphone in the microphone array, and the number of channels in the microphone array is the number of microphones in the microphone array; Calculate the difference signal between different sound source signals, where the sound source signal and The difference signal between them is: ； in: Indicates the sound source signal and The difference signal between ; Indicates the difference signal The nth differential signal value in .

3. A noise suppression method based on a differential microphone array as claimed in claim 2, characterized in that: The step of forming a multi-channel difference sound source signal from the difference signal comprises: The difference signal is used to form a multi-channel difference sound source signal of different sound source signals, where the sound source signal The multi-channel difference sound source signal is: ； in: Indicates the sound source signal The multi-channel difference sound source signal The sound source signal A signal matrix composed of difference signals between the sound source signals of M channels; T stands for transpose.

4. A noise suppression method based on a differential microphone array as claimed in claim 3, characterized in that: The S2 step includes: Signal feature extraction is performed on the multi-channel difference sound source signal, and the feature extraction result is used as the difference signal feature vector of the sound source signal, wherein the multi-channel difference sound source signal The signal feature extraction process is as follows: S21: Calculate the multi-channel difference sound source signal Center frequency : ； in: Indicates the number of signal moments per second; S22: Calculate the multi-channel difference sound source signal Kurtosis : ； ； ； in: Indicates the difference signal The kurtosis component of Indicates the difference signal The standard deviation of the N differential signal values ​​in ; Indicates the difference signal The mean of the N differential signal values ​​in ; S23: Calculate the multi-channel difference sound source signal The power spectral density : ； ； in: Indicates the difference signal The power spectral density component of represents an exponential function with a natural constant as base; j represents the imaginary unit, ; S24: Calculate the multi-channel difference sound source signal Sound pressure ; S25: Multi-channel difference sound source signal Center frequency , Kurtosis , power spectral density And sound pressure As a result of feature extraction, the sound source signal is formed The difference signal eigenvector : .

5. A noise suppression method based on a differential microphone array as claimed in claim 4, characterized in that: The multi-channel difference sound source signal is calculated in step S24 Sound pressure ,include: S241: Calculate and obtain multi-channel difference sound source signal The acoustic power of different difference signals in The sound power is : ； S242: Convert the sound power of the difference signal into a sound pressure component of the difference signal, wherein the difference signal The sound pressure component is : ； in: represents the air density, c represents the speed of sound; represents the distance between the sound source position and the mth microphone position; S243: Constructing a multi-channel difference sound source signal from the sound pressure components Sound pressure : .

6. A noise suppression method based on a differential microphone array as claimed in claim 1, characterized in that: The S3 step includes: Constructing a low-frequency noise determination model, wherein the low-frequency noise determination model includes an input layer, a feature reconstruction layer, a feature mapping layer, and a determination layer, wherein the input layer is used to receive a difference signal feature vector of a sound source signal, the feature reconstruction layer is used to perform dimensionality reduction and reconstruction processing on the difference signal feature vector to obtain a difference signal dimensionality reduction feature vector, the feature mapping layer is used to perform feature mapping on the difference signal dimensionality reduction feature vector using training parameters to obtain a mapping feature vector, and the determination layer is used to perform low-frequency noise determination on the mapping feature vector and output a determination result; The difference signal dimension reduction feature vector of the sound source signal of group D and the actual judgment result of whether there is low-frequency noise are collected to construct a training data set : ； in: Represents the dimension reduction feature vector of the difference signal of the dth group of sound source signals, Indicates the true judgment result of whether the dth group of sound source signals has low-frequency noise. It means that there is no low-frequency noise in the sound source signal of group d. It means that there is low-frequency noise in the sound source signal of group d; Based on the training data set The model parameters are optimized and solved, and the optimization process is as follows: S31: Initialize and generate U groups of Lagrange multipliers with a length of D, where the u-th group of Lagrange multipliers generated by initialization is : ； Set the current iteration number of the Lagrange multiplier to t, the initial value of t is 0, and the maximum value is Max. Then the t-th iteration result of the u-th group of Lagrange multipliers is ; S32: Based on the training data set Construct a fitness function with Lagrange multipliers as input , where the Lagrange multiplier The corresponding fitness function value is : ； in: represents the L1 norm; S33: The U groups of Lagrange multipliers obtained in the t-th iteration are used as the input values ​​of the fitness function in sequence, and the Lagrange multiplier with the largest fitness function value is used as the optimal value obtained in the t-th iteration ; S34: Iterate the Lagrange multiplier, where the iteration formula is: ； ； in: Represents a random number between 0 and 1; express The fitness probability of represents any Lagrange multiplier in the U group of Lagrange multipliers obtained at the t-th iteration; S35: Let t = t + 1, return to step S33, until the maximum number of iterations is reached, and extract the Lagrange multiplier with the maximum fitness function value at this time ; S36: Using the extracted Lagrange multipliers , calculate the training parameters: ； ； in: are the training parameters in the feature map layer.

7. A noise suppression method based on a differential microphone array as claimed in claim 6, characterized in that: The low-frequency noise determination model is used to perform low-frequency noise determination on the feature vector of the difference signal to obtain a determination result of whether the sound source signal has low-frequency noise, including: The input layer receives the difference signal feature vector of the sound source signal; The feature reconstruction layer performs dimensionality reduction and reconstruction on the feature vector of the difference signal, where the feature vector of the difference signal The dimension reduction and reconstruction processing formula is: ； ； ； ； ； in: Represents the difference signal feature vector The corresponding difference signal dimension reduction feature vector; T stands for transpose; represents the L2 norm; Indicates the difference signal Dimensionality reduction features; Indicates normalization processing; represents the standard deviation of all kurtosis components after normalization, represents the mean of all kurtosis components after normalization; represents the standard deviation of all power spectral density components after normalization, Represents the mean of all power spectral density components after normalization; They are the weight coefficients of the kurtosis component, the power spectrum density component, and the sound pressure component respectively; The feature mapping layer uses the training parameters to perform feature mapping on the difference signal dimension reduction feature vector to obtain the mapping feature vector, where the difference signal dimension reduction feature vector The feature mapping formula is: ； in: The dimension reduction feature vector of the difference signal The mapped feature vector obtained by mapping; The decision layer performs low-frequency noise decision on the mapped feature vector. If the mapped feature vector is lower than a preset threshold, the corresponding sound source signal has low-frequency noise. Otherwise, the corresponding sound source signal does not have low-frequency noise.

8. A noise suppression method based on a differential microphone array as claimed in claim 7, characterized in that: The S4 step includes: The low-frequency noise suppression process is performed on the sound source signal with low-frequency noise, where the sound source signal The low-frequency noise suppression processing flow is: S41: Get the sound source signal x that is closest to the mth channel and does not have low-frequency noise: ； in: Represents the differential signal value of the sound source signal x at the nth signal moment; S42: Based on the sound source signal x, the sound source signal with low-frequency noise Double noise reduction filtering is performed, in which the sound source signal Medium differential signal value The double noise reduction filter processing formula is: ； in: Indicates the differential signal value The double noise reduction filter processing result; Indicates the sound source signal The distance between the channel corresponding to x; represents the distance filter scale, Indicates the signal filtering scale; Composing sound source signal The low-frequency noise suppression processing result: ； in: Indicates the sound source signal The low-frequency noise suppression processing result.

Images