Acoustic noise reduction method and device based on improved INMF

Abstract

The present invention discloses an acoustic noise reduction method based on improved INMF, including: performing short-time Fourier transform on pure leakage signal and interference noise signal to obtain respective amplitude spectra; using improved INMF algorithm to decompose the amplitude spectrum of pure leakage signal into dictionary matrix, decomposing the amplitude spectrum of interference noise signal into dictionary matrix, and merging to form joint dictionary matrix as prior information of training process; real-time acquisition of leakage signal containing noise, and performing short-time Fourier transform to obtain the amplitude spectrum of leakage signal containing noise; using improved INMF algorithm and improved adaptive MMSE-LSA algorithm to estimate the amplitude spectrum of leakage signal containing noise after noise reduction in real time; using phase invariance of noisy signal, obtaining the noise-reduced time domain leakage signal through inverse short-time Fourier transform. The defect that traditional method cannot effectively estimate real-time changing signal is overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of acoustic noise reduction, and in particular to an acoustic noise reduction method and device based on an improved INMF. Background Art

[0002] Acoustic signal denoising is a fundamental problem in the field of acoustics, and its purpose is to separate the clean target signal from the noise-contaminated signal. According to the principle and development history of denoising, the acoustic signal denoising methods can be divided into traditional denoising methods and machine learning-based denoising methods.

[0003] Traditional acoustic noise reduction methods mainly use methods based on digital signal processing technology, such as spectral subtraction, Wiener filtering, statistical model-based methods, and subspace methods. However, these methods have certain assumptions or rarely use prior information about pure sound signals and noise signals, resulting in reduced effectiveness when facing non-steady-state noise.

[0004] Audio noise reduction can be considered as a supervised learning problem. More and more experts and scholars are using machine learning methods to improve the effect of audio noise reduction. Machine learning methods can be divided into methods based on traditional machine learning models and methods based on deep learning models. Since deep learning-based model methods require large computing resources and sufficient sound signals to train an excellent noise reduction model, it is difficult to collect enough leakage data and noise data for training for low-probability events such as gas leaks. Therefore, traditional machine learning model noise reduction methods that require relatively small computing resources and data volume are more suitable for such scenarios. Summary of the invention

[0005] In order to solve the above problems, the present invention provides an acoustic noise reduction method and device based on improved INMF, wherein the method comprises the following steps:

[0006] S1. Get pure leakage signal V L (t), interference noise signal V N (t) is used as training information for the pure leakage signal V L (t), interference noise signal V N (t) performing short-time Fourier transform respectively to obtain their respective amplitude spectra;

[0007] S2. Use the improved INMF algorithm to decompose the amplitude spectrum of the pure leakage signal into a dictionary matrix W L , decompose the amplitude spectrum of the interference noise signal into a dictionary matrix W N , W L With W N Combine to form a joint dictionary matrix W as prior information for the training process;

[0008] S3, collecting the noisy air leakage signal in real time, and performing short-time Fourier transform to obtain the amplitude spectrum of the noisy air leakage signal;

[0009] S4, based on W and the amplitude spectrum of the noisy leakage signal, using the improved INMF algorithm and the improved adaptive MMSE-LSA algorithm to estimate in real time the amplitude spectrum of the noisy leakage signal after denoising;

[0010] S5. Based on the amplitude spectrum of the noisy air leakage signal estimated in real time after denoising, using the phase invariance of the noisy signal, a time-domain air leakage signal after denoising is obtained through inverse short-time Fourier transform.

[0011] The present invention also proposes an acoustic noise reduction device based on the improved INMF, comprising:

[0012] processor;

[0013] a memory having stored thereon a computer program executable on the processor;

[0014] Wherein, when the computer program is executed by the processor, an acoustic noise reduction method based on improved INMF is implemented.

[0015] The beneficial effects brought by the technical solution provided by the present invention are:

[0016] The present invention provides an acoustic denoising method based on improved INMF, using the improved INMF algorithm to decompose the amplitude spectrum of a pure leakage signal into a dictionary matrix, decompose the amplitude spectrum of an interference noise signal into a dictionary matrix, merge the two dictionary matrices to form a joint dictionary matrix, and use the joint dictionary as prior information for the training process; based on the joint dictionary matrix and the amplitude spectrum of a leakage signal containing noise, the improved INMF algorithm is used to estimate the amplitude spectrum of a pure leakage signal and an interference noise signal; using the estimated amplitude spectrum of the pure leakage signal and the interference noise signal, the improved adaptive MMSE-LSA algorithm is used to estimate the amplitude spectrum of the pure leakage signal in real time; based on the real-time estimated amplitude spectrum of the pure leakage signal, the denoised time domain leakage signal is obtained by inverse short-time Fourier transform. The defect that the traditional method cannot effectively estimate the real-time changing signal is overcome. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flowchart of the improved INMF noise reduction algorithm implemented in the present invention;

[0018] Figure 2 It is the dictionary matrix and dictionary atoms of the leakage signal implemented by the present invention. DETAILED DESCRIPTION

[0019] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0020] The flowchart of the acoustic noise reduction method based on the improved INMF of this embodiment is as follows: Figure 1 , specifically including the following steps:

[0021] S1. Get pure leakage signal V L (t), interference noise signal V N (t) is used as training information for the pure leakage signal V L (t), interference noise signal V N (t) Perform short-time Fourier transform on each of them to obtain their respective amplitude spectra.

[0022] S2. Use the improved INMF algorithm to decompose the amplitude spectrum of the pure leakage signal into a dictionary matrix W L , the dictionary matrix of the pure leakage signal is as follows Figure 2 As shown in (a), the columns of the dictionary matrix are dictionary atoms, and the dictionary atoms of the pure leakage signal are as follows: Figure 2 As shown in (b), the dictionary atom is a non-negative function of frequency and can be linearly combined with the corresponding coefficients in real time. The amplitude spectrum of the interference noise signal is decomposed into the dictionary matrix W N , W L With W N The combined dictionary matrix W is used as the prior information for the training process.

[0023] S3. Collect the noisy air leakage signal in real time, and perform short-time Fourier transform to obtain the amplitude spectrum of the noisy air leakage signal.

[0024] S4. Based on W and the amplitude spectrum of the noisy leakage signal, the improved INMF algorithm and the improved adaptive MMSE-LSA algorithm are used to estimate in real time the amplitude spectrum of the noisy leakage signal after denoising.

[0025] S5. Based on the amplitude spectrum of the noisy air leakage signal estimated in real time after denoising, using the phase invariance of the noisy signal, a time-domain air leakage signal after denoising is obtained through inverse short-time Fourier transform.

[0026] The improved INMF algorithm is mainly as follows:

[0027] (1) Introducing the noise term E, X = WH + E. The introduced noise term can effectively reduce the interference between the coefficient matrix H and the basis matrix W.

[0028] The objective function under KL divergence is expressed as:

[0029]

[0030] in, is the objective function under KL divergence, X is the amplitude spectrum of the noisy signal, is the amplitude spectrum of the noisy air leakage signal after denoising, E is the introduced noise term, K is the frequency of the noisy air leakage signal, L is the time frame of the noisy air leakage signal, x∈X, e∈E, (xe) k,l is the objective function of KL divergence after adding noise term, lg represents the logarithmic operation with base 10, It is the amplitude spectrum of the noisy leakage signal after denoising estimated under KL divergence.

[0031] (2) In order to make the above equation more sparse, we consider adding a norm constraint term to the noise term. 0 The norm is difficult to solve, so we choose to add L to the noise term 1 Norm.

[0032]

[0033] Where a is the trade-off coefficient that controls the sparsity of E and the reconstruction error.

[0034] (3) Since the acoustic signal has sparse properties in the time domain and frequency domain, the sparse factor is considered to determine the sparsity of the coefficient matrix H, so that the basis matrix W becomes a complete basis. The sparse constraint term of the coefficient matrix H is added to the formula in (2), so that the sparsity of H and the distortion of the audio signal can be better controlled, so the objective function becomes:

[0035]

[0036] Where H is the coefficient matrix, ||H|| 1 L for H 1 norm matrix, and b is the sparsity factor of the determination coefficient matrix.

[0037] (4) When the threshold operator is introduced, the added noise term can be continuously updated and optimized by fixing W and H, which can solve the convex optimization problem during the update and make it free from the constraints of the noise outlier point form, thus improving the robustness of the algorithm. λ (·) can be expressed as

[0038]

[0039] Among them, λ is the threshold.

[0040] (5) Taking advantage of the fact that the objective function does not have a unique optimal solution, W and H are normalized while keeping the value of the objective function unchanged. Then, the gradient descent method is used to optimize the formula in (3). The values ​​of W, H, and E are determined as follows:

[0041]

[0042]

[0043] E←soft λ (X-WH)

[0044] Among them, ·* and · / represent the multiplication and division in the decomposition matrix, X=WH+E, WH=W L H L +W N H N , X is the amplitude spectrum of the noisy air leakage signal, E is the introduced noise term, H is the coefficient matrix, and T represents the transposition process. is the amplitude spectrum of the noisy leakage signal after denoising, 1 K×L ∈R K×L It is represented as a K×L matrix with all element values ​​1, b is the sparse factor of the determination coefficient matrix, and soft λ () is the soft threshold function.

[0045] (6) The MMSE-LSA algorithm obtains the estimated value of the denoised audio by minimizing the following formula.

[0046]

[0047] |X(n,k)|=G(n,k)|Y(n,k)

[0048]

[0049]

[0050] Where ln is the natural logarithm operation, |X(n,k)|=G(n,k) / Y(n,k), ξ(n,k) is the prior signal-to-noise ratio of the nth frame at the kth frequency point in the denoised air leakage signal, X(n,k) is the spectral amplitude of the denoised air leakage signal at the kth component of the nth frame, It is represented by the spectral amplitude of the noise-reduced leakage signal estimated at the kth component of the nth frame, and Y(n,k) is represented by the amplitude spectrum of the noisy leakage signal at the kth component of the nth frame.

[0051] (7) To estimate the pure audio signal, the prior signal-to-noise ratio must be calculated first. Therefore, the prior signal-to-noise ratio has a direct impact on the audio noise reduction effect. The estimation of the prior signal-to-noise ratio can be expressed as:

[0052]

[0053] in, is the spectral amplitude of the noise-reduced leakage signal estimated at the kth component of the n-1 frame, is the noise signal amplitude spectrum estimated at the kth component of the n-1 frame, α is the weighting factor, and γ(n,k) is the posterior signal-to-noise ratio.

[0054]

[0055] P[] represents a half-wave rectifier function. In order to The difference between ξ(n,k) and ξ(n,k) is as small as possible and estimated using the minimum mean square error (MMSE).

[0056]

[0057] General Substitution Available:

[0058]

[0059] Deriving G and setting its partial derivative to 0, we can get the optimal solution of α:

[0060]

[0061] use to replace the unknown ξ(n,k), because Therefore, when the posterior signal-to-noise ratio changes, the value of α(n,k) approaches 1, and when the value of α(n,k) is small, It will change accordingly, thus achieving an adaptive effect.

[0062] In a further embodiment of the present invention:

[0063] S1. Get pure leakage signal V L (t), interference noise signal V N (t) is used as training information for the pure leakage signal V L (t), interference noise signal V N (t) Perform short-time Fourier transform on each of them to obtain their respective amplitude spectra.

[0064] S2. Use the improved INMF algorithm to decompose the amplitude spectrum of the pure leakage signal into a dictionary matrix W L , decompose the amplitude spectrum of the interference noise signal into a dictionary matrix W N , W L With W N The combined dictionary matrix W is used as the prior information for the training process.

[0065] Set the number of iterations to 100 and iterate the following formulas to decompose the amplitude spectrum of the pure leakage signal into the dictionary matrix WL , decompose the amplitude spectrum of the interference noise signal into a dictionary matrix W N .

[0066]

[0067]

[0068] E←soft λ (X-WH) (3)

[0069] Among them, ·* and · / represent the multiplication and division in the decomposition matrix, X=WH+E, WH=W L H L +W N H N , X is the amplitude spectrum of the noisy air leakage signal, E is the introduced noise term, H is the coefficient matrix, and T represents the transposition process. is the amplitude spectrum of the noisy leakage signal after denoising, 1 K×L ∈R K×L represents a K×L matrix whose elements are all 1, b is the sparse factor of the determination coefficient matrix, and soft λ () is the soft threshold function.

[0070]

[0071] Among them, x is soft λ (x) is the variable, and λ is the threshold.

[0072] S3. Collect the noisy air leakage signal in real time, and perform short-time Fourier transform to obtain the amplitude spectrum of the noisy air leakage signal.

[0073] S4. Based on W and the amplitude spectrum of the noisy leakage signal, the improved INMF algorithm and the improved adaptive MMSE-LSA algorithm are used to estimate in real time the amplitude spectrum of the noisy leakage signal after denoising.

[0074] Use the amplitude spectrum of the leak signal containing noise and the joint dictionary matrix W as the input parameters of formula (2), keep W unchanged, and use formula (2) and formula (3) to iterate and update continuously until the objective function reaches convergence and terminates;

[0075]

[0076] E←soft λ (X-WH)

[0077] The objective function is:

[0078]

[0079] in, is the objective function under KL divergence, X is the amplitude spectrum of the noisy signal, is the amplitude spectrum of the noisy air leakage signal after denoising, E is the introduced noise term, H is the coefficient matrix, K is the frequency of the noisy air leakage signal, L is the time frame of the noisy air leakage signal, x∈X, e∈E, (xe) k,l is the objective function of KL divergence after adding noise term, lg represents the logarithmic operation with base 10, is the amplitude spectrum of the noisy leakage signal after denoising estimated under KL divergence, ||E|| 1 L for E 1 norm matrix, ||H|| 1 L for H 1 norm matrix, a and b are different trade-off coefficients, a is the trade-off coefficient that controls the sparsity of E and the reconstruction error, and b is the sparsity factor that determines the coefficient matrix.

[0080] After the iteration, the pure leakage audio signal is estimated and interference noise signal The amplitude spectrum can be derived as

[0081] The improved adaptive MMSE-LSA algorithm is:

[0082] By minimizing the value of formula (6), the estimated value of the leakage signal after noise reduction is obtained:

[0083]

[0084] Where ln is the natural logarithm, |X(n,k)|=G(n,k) / Y(m,k), ξ(n,k) is the prior signal-to-noise ratio of the nth frame at the kth frequency point in the denoised air leakage signal: X(n,k) represents the spectral amplitude of the noise-reduced leakage signal at the kth component of frame n, It is represented by the spectral amplitude of the noise-reduced leakage signal estimated at the kth component of the nth frame, and Y(n,k) is represented by the amplitude spectrum of the noisy leakage signal at the kth component of the nth frame;

[0085] The estimate of the prior signal-to-noise ratio is expressed as:

[0086]

[0087] in, is the spectral amplitude of the noise-reduced leakage signal estimated at the kth component of the n-1 frame, is the noise signal amplitude spectrum estimated at the kth component of the n-1 frame, α is the weighting factor, P is the half-wave rectification function, and γ(n,k) is the posterior signal-to-noise ratio.

[0088]

[0089] make The difference between ξ(n,k) and ξ(n,k) is as small as possible, and the minimum mean square error method is used to estimate:

[0090]

[0091] Among them, MMSE means using the minimum mean square error method for the following formula, and G is the result of using the minimum mean square error estimation;

[0092] Substituting formula (7) into formula (9), we get formula (10):

[0093]

[0094] The optimal solution of α is:

[0095]

[0096] In formula (11), P{ξ(n,k)-1} is used to replace the unknown number ξ(n,k) to achieve an adaptive effect.

[0097] After the iterative update, the G in the improved adaptive MMSE-LSA algorithm is used as the gain function

[0098]

[0099] Real-time estimation of pure air leakage audio amplitude spectrum is the gain function. This overcomes the defect that traditional methods such as Wiener filtering cannot effectively estimate real-time changing signals.

[0100] S5. Based on the amplitude spectrum of the noisy air leakage signal estimated in real time after denoising, using the phase invariance of the noisy signal, a time-domain air leakage signal after denoising is obtained through inverse short-time Fourier transform.

[0101] The embodiment also includes an acoustic noise reduction device based on the improved INMF, including:

[0102] processor;

[0103] a memory having stored thereon a computer program executable on the processor;

[0104] When the computer program is executed by the processor, an acoustic noise reduction method based on improved INMF is implemented.

[0105] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. Acoustic noise reduction method based on improved INMF, It is characterized in that The following steps are involved: S1. Get pure leakage signal V L (t), interference noise signal V N (t) is used as training information for the pure leakage signal V L (t), interference noise signal V N (t) performing short-time Fourier transform respectively to obtain their respective amplitude spectra; S2. Use the improved INMF algorithm to decompose the amplitude spectrum of the pure leakage signal into a dictionary matrix W L , decompose the amplitude spectrum of the interference noise signal into a dictionary matrix W N , W L With W N Combine to form a joint dictionary matrix W as prior information for the training process; S3, collecting the noisy air leakage signal in real time, and performing short-time Fourier transform to obtain the amplitude spectrum of the noisy air leakage signal; S4, based on W and the amplitude spectrum of the noisy leakage signal, using the improved INMF algorithm and the improved adaptive MMSE-LSA algorithm to estimate in real time the amplitude spectrum of the noisy leakage signal after denoising; S5, based on the amplitude spectrum of the noisy air leakage signal after denoising estimated in real time, using the phase invariance of the noisy signal, and obtaining the time domain air leakage signal after denoising through inverse short-time Fourier transform; Step S2 is specifically as follows: Set the number of iterations c and iterate the following formula c times: E←soft λ (X-WH)(3) Where ·* and · / represent the multiplication and division in the decomposition matrix, X is the amplitude spectrum of the noisy air leakage signal, E is the introduced noise term, H is the coefficient matrix, and T represents the transposition process. is the amplitude spectrum of the noisy leakage signal after denoising, 1 K×L ∈R K×L It is represented as a K×L matrix with all element values ​​1, b is the sparse factor of the determination coefficient matrix, and soft λ () is the soft threshold function; Among them, x is soft λ (x) is the variable, λ is the threshold; Step S4 is specifically as follows: Use the amplitude spectrum of the leak signal containing noise and the joint dictionary matrix W as the input parameters of formula (2), keep W unchanged, and use formula (2) and formula (3) to iterate and update continuously until the objective function reaches convergence and terminates; The objective function is: in, is the objective function under KL divergence, X is the amplitude spectrum of the noisy signal, is the amplitude spectrum of the noisy air leakage signal after denoising, E is the introduced noise term, H is the coefficient matrix, K is the frequency of the noisy air leakage signal, L is the time frame of the noisy air leakage signal, x∈X, e∈E, (xe) k,l is the objective function of KL divergence after adding noise term, lg represents the logarithmic operation with base 10, is the amplitude spectrum of the noisy leakage signal after denoising estimated under KL divergence, ||E|| 1 L for E 1 norm matrix, ||H|| 1 L for H 1 norm matrix, a and b are different trade-off coefficients, a is the trade-off coefficient that controls the sparsity of E and the reconstruction error, and b is the sparsity factor that determines the coefficient matrix.

2. The acoustic noise reduction method based on improved INMF according to claim 1, It is characterized in that In step S4, the improved adaptive MMSE-LSA algorithm is: By minimizing the value of formula (6), the estimated value of the leakage signal after noise reduction is obtained: Among them, |X(n,k)|=G(n,k) / Y(n,k), ξ(n,k) is the prior signal-to-noise ratio of the nth frame at the kth frequency point in the denoised air leakage signal: X(n,k) represents the spectral amplitude of the noise-reduced leakage signal at the kth component of frame n, It is represented by the absolute value of the noise-reduced leakage signal spectrum amplitude estimated at the kth component of the nth frame, and Y(n,k) is represented by the noisy leakage signal amplitude spectrum at the kth component of the nth frame; The estimate of the prior signal-to-noise ratio is expressed as: in, is the spectral amplitude of the noise-reduced leakage signal estimated at the kth component of the n-1 frame, is the noise signal amplitude spectrum estimated at the kth component of the n-1 frame, α is the weighting factor, P is the half-wave rectification function, and γ(n,k) is the posterior signal-to-noise ratio. make The difference between ξ(n,k) and ξ(n,k) is as small as possible, and the minimum mean square error method is used to estimate and ξ(n,k): Among them, MMSE means using the minimum mean square error method for the following formula, and G is the result of using the minimum mean square error estimation; Substituting formula (7) into formula (9), we get formula (10): The optimal solution of α is: In formula (11), P{ξ(n,k)-1} is used to replace the unknown number ξ(n,k) to achieve an adaptive effect.

3. The acoustic noise reduction method based on improved INMF according to claim 1, It is characterized in that In step S4, the improved adaptive MMSE-LSA algorithm is used to estimate the pure leakage signal amplitude spectrum in real time. Specifically, G in the improved adaptive MMSE-LSA algorithm is used as the gain function. in, is the gain function; pass Real-time estimation of pure air leakage audio amplitude spectrum 4. Acoustic noise reduction device based on improved INMF, It is characterized in that The device comprises: processor; a memory having stored thereon a computer program executable on the processor; Wherein, when the computer program is executed by the processor, the acoustic noise reduction method based on improved INMF as described in any one of claims 1 to 3 is implemented.

Images