Method for reducing human voice noise in machine room based on 2D-TASNet

Abstract

The invention belongs to the technical field of machine room noise reduction, and particularly relates to a machine room internal human voice noise reduction method based on 2D-TASNet, and the method comprises the steps: preprocessing: carrying out the short-time Fourier transform of a mixed signal, and converting a one-dimensional time-domain signal into a two-dimensional time-frequency spectrum; a feature decomposition step: carrying out feature extraction on the amplitude spectrum of the time-frequency spectrum by adopting sparse non-negative matrix factorization; feature clustering: K-Means clustering is performed on a matrix obtained through SNMF decomposition, and the mass center of a pure human voice time-frequency spectrum is used as a clustering center to enhance the robustness of a separation effect; and a human voice separation step: adjusting the original TASNet to be a 2D-TASNet, training the 2D-TASNet to obtain a target mask by taking the features obtained by the K-Means clustering as input, and reconstructing the separated human voice by using the mask. The frequency characteristics of the noise and the human voice are different, the structure of the 2D-TASNet focuses on mining the potential characteristics of various time frequencies, and in this way, the respective time-frequency characteristics and nonlinear mixing of the human voice noise can be utilized, and the precision advantage of the neural network can be kept.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of machine room noise reduction, and particularly relates to a machine room indoor human voice noise reduction method based on 2D-TASNet. BACKGROUND

[0002] In the machine room environment, the continuous heat dissipation of servers produces a huge background noise, which brings challenges to the communication of technicians in daily inspection, installation, debugging and maintenance work. This continuous noise not only hinders effective team collaboration, but also can cause damage to the hearing of workers. More importantly, the nature of machine room work requires highly accurate communication, because any small misunderstanding in communication can lead to equipment failure or other serious consequences. Therefore, it is particularly important to develop a technology that can clearly separate human voice and noise in this special environment.

[0003] Traditional signal processing techniques have limitations in processing complex sound signals that are nonlinearly mixed. For example, they may not be able to effectively remove background noise without losing the quality of human voice, especially when the noise is similar in frequency to the human voice or changes frequently.

[0004] Human voice and background noise (such as server heat dissipation noise) are usually nonlinearly mixed, and traditional noise reduction algorithms are difficult to effectively separate them.

[0005] Therefore, the present application starts from the time-frequency spectrum, and uses the difference in time-frequency characteristics between human voice and noise to achieve their effective separation. Specifically, human voice usually exhibits certain regularity and sparsity in the time-frequency spectrum, while background noise is more random and scattered. Through this characteristic, the present application can design algorithms to accurately identify human voice components and separate them from complex background noise.

[0006] The two-dimensional TASNet (2D-TASNet) proposed by the present application is used for server indoor human voice noise reduction, which can accurately identify and separate complex sound signals, provide clear human voice output, and thus improve the on-site communication effect of technicians. SUMMARY

[0007] The purpose of the present application is to provide a machine room indoor human voice noise reduction method based on 2D-TASNet, which realizes server indoor human voice noise reduction based on two-dimensional TASNet (2D-TASNet), accurately identifies and separates complex sound signals, provides clear human voice output, and thus improves the on-site communication effect of technicians.

[0008] To solve the above technical problems, the technical solution adopted by the present application is as follows: A machine room indoor human voice noise reduction method based on 2D-TASNet, comprising the following steps: S1: Preprocessing: Perform a short-time Fourier transform on the mixed time-domain signal of human voice and background noise to convert the one-dimensional time-domain signal into a two-dimensional time-spectrum signal. S2: Feature decomposition: Based on the sparsity of human voice, sparse nonnegative matrix decomposition is used to extract features of the input mixed signal; S3: Feature clustering: The matrix obtained by decomposing the amplitude spectrum through SNMF needs to be clustered using K-Means; S4: Voice Separation: This is an end-to-end separation system that is modified based on the original TASNet. The input of the original TASNet is modified to a two-dimensional amplitude spectrum, and the output is modified to an ideal proportional mask. The parameters and loss function of the internal structure of the network are adjusted to adapt the input and output. The features obtained by K-Means clustering are used as the input of 2D-TASNet to train 2D-TASNet to obtain the target mask. S5: Reconstruct the separated human voice using the target mask.

[0009] Preferably, the specific process of step S1 is as follows: The time-frequency domain signal is obtained by performing a short-time Fourier transform (STFT) through a Hamming window with a window length of 256 points and a window shift of 128 points. s ( f , t ); ; in, It is a window function. f Represents frequency, s ( t () is a mixed signal of human voice and background noise.

[0010] Preferably, the specific process of feature decomposition in step S2 is as follows: S21: Take the modulus value of the two-dimensional time-frequency signal to obtain the amplitude spectrum: V( f , t )=| s ( f , t )|; S22: The input matrix V is decomposed into an r-dimensional basis matrix W and a coefficient matrix H, with constraints including non-negativity of elements and L1 sparsity constraints, which are reflected in the coefficient matrix H. An SNMF objective function with Euclidean distance is then defined. ; ; in, λ It is a regularization parameter used to control sparsity and reconstruction error; the objective function aims to... H yuanh as possible to ensure the sparsity of its result, where the dimension of V is f * t , W the dimension of V is f * r , H the dimension of V is r * t , r is the decomposition dimension.

[0011] Preferably, an iterative process is performed: S23: initialize W and H as random positive matrices, apply sparsity constraints to H, then project each row of H to be non-negative, and set L1 norm to achieve the desired sparsity; S24: project each column of W to be non-negative: ; S25: project each row of H to be non-negative, set unit L1 norm to achieve the desired sparsity: ; When the cost function converges, the loop is exited and the matrices W and H are output; otherwise, iteration continues.

[0012] Preferably, the specific process of feature clustering in step S3 is as follows: S31: the decomposed basis matrix W is a vector set composed of K basis vectors , the coefficient matrix H is composed of K coefficient vectors, and the vector set W is a vector set composed of K basis vectors, represented as follows: ; ; S32: the basis matrix W and the coefficient matrix H contain human voice and noise features distributed in different frequency bands, obtained by w k and h k the product of the submatrix F k : ; The submatrix F k is the feature belonging to the source signal, therefore, clustering is needed for w k and h k to obtain target features, and clustering results in pure features containing only human voice and no noise.

[0013] Preferably, the specific process of human voice separation in step S4 is as follows: S41: mapping the 2D input features of the amplitude spectrum of the mixed signal to a higher dimensional space through an encoder: ; S42: completing separation in the high-dimensional space through an estimator; S43: restoring the high-dimensional vector calculated by the estimator to the same dimension mask as the input of the encoder through a decoder: .

[0014] Preferably, the specific process of reconstructing the separated human voice with the target mask in step S5 is as follows: After the time-frequency spectrum point product of the mask and the mixed signal, only the time-frequency spectrum containing the target human voice is obtained, and then the inverse STFT is performed to obtain the separated human voice, and the specific formula is as follows: ; Wherein, s1(t) is the separated human voice signal, M is the estimated mask, S is the time-frequency spectrum of the mixed signal, and ISTFT is the inverse Fourier transform.

[0015] The beneficial effects of the present application include: The machine room indoor human voice noise reduction method based on 2D-TASNet provided by the present application, preprocessing: performing short-time Fourier transform (STFT) on the mixed signal, converting a one-dimensional time domain signal into a two-dimensional time-frequency spectrum; feature decomposition step: performing feature extraction on the amplitude spectrum of the time-frequency spectrum by using sparse non-negative matrix factorization (SNMF); feature clustering: performing K-Means clustering on the matrix obtained by SNMF decomposition, using the centroid of the pure human voice time-frequency spectrum as the clustering center to enhance the robustness of the separation effect; human voice separation step: adjusting the original TASNet to 2D-TASNet, using the features obtained by K-Means clustering as input, training 2D-TASNet to obtain a target mask, and reconstructing the separated human voice with the mask. The frequency characteristics of noise and human voice are different, the structure of 2D-TASNet focuses on mining the potential features of various time frequencies, and such a way can not only utilize the time-frequency characteristics and nonlinear mixing of human voice and noise, but also maintain the precision advantage of the neural network.

[0016] Firstly, the time domain signal is converted into a two-dimensional time-frequency spectrum through short-time Fourier transform, the time-frequency joint features of the signal are retained, and more abundant information basis is provided for subsequent processing. The amplitude spectrum is decomposed by sparse non-negative matrix decomposition, the sparse characteristics of human voice in the time-frequency domain are utilized, the base matrix and coefficient matrix obtained by decomposition are more consistent with the feature differences between human voice and noise through non-negative constraint and L1 sparsity constraint, and representative local features are effectively separated, laying a high-quality feature foundation for subsequent clustering and separation.

[0017] Secondly, K-Means clustering is used to cluster the matrix after SNMF decomposition to further filter out pure features containing only human voices and remove noise-related features. This feature purification process reduces the interference of noise on subsequent network training, allowing 2D-TASNet to focus more on learning the feature patterns of human voices and improve the model's accuracy in recognizing and separating human voices.

[0018] Furthermore, the native TASNet is specifically improved by adjusting the input to a two-dimensional amplitude spectrum and the output to an ideally scaled mask, while adapting the network's internal parameters and loss function to better suit the processing requirements of two-dimensional time-frequency features. This improved 2D-TASNet can fully learn the local and global correlations in the time-frequency domain, generating a more accurate target mask, thereby effectively separating human voices from mixed signals. Simultaneously, the end-to-end network design simplifies the process, reduces manual intervention, and improves separation efficiency.

[0019] Furthermore, the background noise in the computer room environment is complex (such as equipment operating noise, electromagnetic interference, etc.). This method can specifically suppress the continuous or sudden noise unique to the computer room by using the sparsity constraint of SNMF and K-Means clustering. Combined with the deep mining capability of 2D-TASNet for two-dimensional time-frequency features, it can still maintain high human voice separation quality even in computer room environments with low signal-to-noise ratio, thus improving the robustness and practicality of the method in real-world scenarios.

[0020] Finally, by performing a dot product operation between the target mask and the time spectrum of the mixed signal, and then reconstructing the human voice using the inverse short-time Fourier transform, noise can be suppressed while preserving the original spectral characteristics and temporal continuity of the human voice to the greatest extent possible. This ensures that the separated human voice is clear and intelligible, and enables effective extraction of human voice signals in computer room scenarios. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the 2D-TASNet-based architecture of the present invention. Detailed Implementation

[0022] The following is in conjunction with the appendix Figure 1 The present invention will be further described in detail below: Example 1 A method for human voice noise reduction in a computer room based on 2D-TASNet includes the following steps: S1: Preprocessing: Perform a short-time Fourier transform on the mixed time-domain signal of human voice and background noise to convert the one-dimensional time-domain signal into a two-dimensional time-spectrum signal. S2: Feature decomposition: Based on the sparsity of human voice, sparse nonnegative matrix decomposition is used to extract features of the input mixed signal; S3: Feature clustering: The matrix obtained by decomposing the amplitude spectrum through SNMF needs to be clustered using K-Means; S4: Voice Separation: This is an end-to-end separation system that is modified based on the original TASNet. The input of the original TASNet is modified to a two-dimensional amplitude spectrum, and the output is modified to an ideal proportional mask. The parameters and loss function of the internal structure of the network are adjusted to adapt the input and output. The features obtained by K-Means clustering are used as the input of 2D-TASNet to train 2D-TASNet to obtain the target mask. S5: Reconstruct the separated human voice using the target mask.

[0023] In this embodiment, the specific process of step S1 is as follows: The time-frequency domain signal is obtained by performing a short-time Fourier transform (STFT) through a Hamming window with a window length of 256 points and a window shift of 128 points. s ( f , t ); ; in, It is a window function. f Represents frequency, s ( t () is a mixed signal of human voice and background noise.

[0024] The specific process of feature decomposition in step S2 is as follows: S21: Take the modulus value of the two-dimensional time-frequency signal to obtain the amplitude spectrum: V( f , t )=| s ( f , t )|; S22: The input matrix V is decomposed into an r-dimensional basis matrix W and a coefficient matrix H, with constraints including non-negativity of elements and L1 sparsity constraints, which are reflected in the coefficient matrix H. An SNMF objective function with Euclidean distance is then defined. ; ; in, λ It is a regularization parameter used to control sparsity and reconstruction error; the objective function aims to... H yuan h The value should be as small as possible to ensure the sparsity of the results, where the dimension of V is zero. f * t , W The dimension is f * r, H The dimension is r * t , r For decomposition dimensions.

[0025] Execute the iterative process: S23: Initialize W and H as random positive matrices, apply sparsity constraints to H, then project each row of H to be non-negative, and set the L1 criterion to achieve the desired sparsity; S24: Project each column of W as nonnegative: ; S25: Project each row of H to be non-negative, and set the unit L1 standard to achieve the desired sparsity: ; When the cost function converges, exit the loop and output matrices W and H; otherwise, continue iterating.

[0026] Example 2 Based on Example 1, the specific process of feature clustering in step S3 is as follows: S31: The decomposed basis matrix W is a vector set consisting of K basis vectors. The coefficient matrix H consists of K coefficient vectors, and the vector set W consists of K basis vectors, as shown below: ; ; S32: The basis matrix W and coefficient matrix H contain human voice and noise features distributed across different frequency bands, by... w k and h k The submatrix obtained by multiplication F k : ; submatrix F k These are characteristics of the source signal; therefore, it is necessary to... w k and h k Clustering is performed to obtain the target features. The clustering results in pure features that contain only human voices and no noise.

[0027] The clustering process can be described as follows: Because the initial centers of K-Means are random, which can easily lead to unstable results, this invention uses several pure human voice amplitude spectra and calculates their average to obtain a template feature. The centroid of this template feature is used as one of the cluster centers of K-Means to improve the stability of clustering. Next, a KS clustering algorithm based on cosine similarity is applied to each vector until all vectors are clustered. By superimposing these clusters, the sub-features belonging to the human voice class are obtained, thus yielding the target features.

[0028] See Figure 1 2D-TASNet consists of three parts: Encoder: an encoder that maps 2D input features (amplitude spectrum of the mixed signal) to a higher-dimensional space. In this invention, the encoder layer is essentially a convolutional layer, whose function is to map the amplitude spectrum to a high-dimensional learnable embedding space.

[0029] Estimator: An estimator that performs separation in high-dimensional space. The estimator is a fully convolutional structure, composed of multiple convolutional groups, activation functions, and normalization functions stacked together. The core of the estimator is a Temporal Convolutional Network (TCN) module, consisting of eight exponentially increasing expansion factors and Depth-Conv blocks. As the dimension *dim* increases, the expansion factor within each depth convolutional block doubles. Each depth convolutional block has two outputs, called the residual path and the jump connection path. The residual path is used as the input to the next block, and the outputs of all jump connection paths are summed to ensure a sufficiently large temporal context window to capture the long-range correlations of the input signal.

[0030] Decoder: The purpose of a decoder is to recover a high-dimensional vector calculated by the estimator into a mask of the same dimension as the encoder input. The decoder is essentially the inverse of the encoder, that is, a deconvolutional layer with the same parameters as the encoder.

[0031] The specific process of voice separation in step S4 is as follows: S41: The encoder maps the 2D input features of the amplitude spectrum of the mixed signal to a higher-dimensional space. ; S42: Separation is achieved in high-dimensional space using an estimator; S43: The decoder recovers the high-dimensional vector calculated by the estimator into a mask of the same dimension as the encoder input. .

[0032] Preferably, the specific process of reconstructing the separated human voice using the target mask in step S5 is as follows: After performing a time-spectrum dot product of the mask and the mixed signal, we obtain the time-spectrum containing only the target human voice. Then, we perform an inverse STFT to obtain the separated human voice. The specific formula is as follows: ; Where s1(t) is the separated human voice signal, M is the estimated mask, S is the time spectrum of the mixed signal, and ISTFT is the inverse Fourier transform.

[0033] The mixed signal is preprocessed by performing a short-time Fourier transform to convert the one-dimensional time-domain signal into a two-dimensional time-frequency spectrum. The feature decomposition step involves extracting features from the amplitude spectrum of the time-frequency spectrum using sparse non-negative matrix factorization. The feature clustering step involves performing K-Means clustering on the matrix obtained from SNMF decomposition, using the centroid of the pure human voice time-frequency spectrum as the cluster center to enhance the robustness of the separation effect. The human voice separation step involves adjusting the original TASNet to a 2D-TASNet, using the features obtained from K-Means clustering as input to train the 2D-TASNet to obtain the target mask, and using this mask to reconstruct the separated human voice.

[0034] In summary, the 2D-TASNet-based human voice denoising method for computer rooms provided by this invention converts the time-domain signal into a two-dimensional time-frequency spectrum through short-time Fourier transform, preserving the joint time-frequency features of the signal and providing a richer information foundation for subsequent processing. By combining sparse non-negative matrix factorization (SNMF) to perform feature decomposition on the amplitude spectrum, and leveraging the sparsity of human voice in the time-frequency domain, non-negativity constraints and L1 sparsity constraints are applied to ensure that the resulting basis matrix and coefficient matrix better reflect the feature differences between human voice and noise, effectively separating representative local features and laying a high-quality feature foundation for subsequent clustering and separation. K-Means clustering is then used to cluster the SNMF-decomposed matrix, further filtering out pure features containing only human voice and eliminating noise-related features. This feature purification process reduces the interference of noise on subsequent network training, allowing 2D-TASNet to focus more on learning the feature patterns of human voice and improving the model's accuracy in recognizing and separating human voice.

[0035] This paper proposes targeted improvements to the native TASNet, adjusting the input to a two-dimensional amplitude spectrum and the output to an ideal proportional mask. The network's internal parameters and loss function are also adapted to better suit the processing requirements of two-dimensional time-frequency features. This improved 2D-TASNet can fully learn the local and global correlations in the time-frequency domain, generating a more accurate target mask, thus effectively separating human voices from mixed signals. Simultaneously, the end-to-end network design simplifies the process, reduces manual intervention, and improves separation efficiency. Given the complex background noise in computer room environments (such as equipment operating noise and electromagnetic interference), this method utilizes the sparsity constraints of SNMF and K-Means clustering to specifically suppress the persistent or sudden noise unique to computer rooms. Combined with the deep mining capabilities of 2D-TASNet for two-dimensional time-frequency features, it maintains high human voice separation quality even in low signal-to-noise ratio computer room environments, enhancing the robustness and practicality of the method in real-world scenarios. By performing a dot product operation between the target mask and the time spectrum of the mixed signal, and combining it with the inverse short-time Fourier transform to reconstruct the human voice, it is possible to suppress noise while preserving the original spectral characteristics and temporal continuity of the human voice to the greatest extent, ensuring that the separated human voice is clear and intelligible, and realizing the effective extraction of human voice signals in computer room scenarios.

Claims

1. A 2D-TASNet-based machine room in-scene voice noise reduction method, characterized in that, Comprising the following steps: S1： Preprocessing: short-time Fourier transform is performed on the mixed time-domain signal of human voice and background noise, converting one-dimensional time-domain signal into two-dimensional time-frequency spectrum signal; S2: feature decomposition: according to the sparsity of human voice, SNMF is used to extract the features of the input mixed signal; S3: feature clustering: after the amplitude spectrum is decomposed by SNMF, the matrix obtained by decomposition needs to be clustered by K-Means; S4: human voice separation: through an end-to-end separation system, the input of the original TASNet is modified to a two-dimensional amplitude spectrum, the output is modified to an ideal proportion mask, and the parameters and loss function of the internal structure of the network are adjusted to adapt to the input and output. The features obtained by K-Means clustering are used as the input of 2D-TASNet, and 2D-TASNet is trained to obtain the target mask; S5: reconstruct the separated human voice using the target mask.

2. The method of claim 1, wherein the method is a 2D-TASNet-based in-room voice denoising method for a machine room. The specific process of step S1 is as follows: The time-frequency domain signal is obtained after short-time Fourier transform STFT by a Hamming window with a window length of 256 points and a window shift of 128 points s ( f , t ) ； wherein is a window function, f represents the frequency, s ( t ) is a mixed signal of the human voice and the background noise.

3. The method for human voice noise reduction in a computer room based on 2D-TASNet according to claim 1, characterized in that, The specific process of feature decomposition in step S2 is as follows: S21: the two-dimensional time-frequency spectrum signal is taken as the modulus value to obtain the amplitude spectrum: V( f , t )=| s ( f , t )|； S22: the input matrix V is decomposed into an r-dimensional basis matrix W and a coefficient matrix H, with the constraint condition of non-negative elements and L1 sparsity constraint. The constraint is in the coefficient matrix H, and the SNMF objective function with Euclidean distance is set: ； ； wherein, λ is a regularization parameter to control sparsity and reconstruction error, the dimension of V is f * t , W the dimension of is f * r , H the dimension of is r * t , r is the decomposition dimension.

4. The method of claim 3, wherein the method is based on a 2D-TASNet. An iterative process is performed: S23: W and H are initialized as random positive matrices, and the sparsity constraint is applied to H, then each row of H is projected to be non-negative, and the L1 criterion is set to achieve the required sparsity; S24: each column of W is projected to be non-negative: ； S25: each row of H is projected to be non-negative, and the unit L1 specification is set to achieve the preset sparsity: ； When the cost function converges, the loop is exited, and the matrix W and H are output; otherwise, iteration continues.

5. The method of claim 1, wherein the method is based on a 2D-TASNet. The specific process of feature clustering in step S3 is as follows: S31: The decomposed basis matrix W is a vector set composed of K basis vectors The coefficient matrix H is composed of K coefficient vectors, and the vector set W is a vector set composed of K basis vectors, which is represented as follows: ； ； S32: The basis matrix W and the coefficient matrix H contain the characteristics of the human voice and noise distributed in different frequency bands, which are obtained by the product of the basis matrix W and the coefficient matrix H w k and h k the sub-matrix F k : ； sub-matrix F k is a feature belonging to the source signal, and w k and h k perform clustering to obtain target features, and the clustering results are pure features containing only human voice and no noise.

6. The method of claim 1, wherein the method is a 2D-TASNet-based in-room voice denoising method for a machine room. The specific process of human voice separation in step S4 is as follows: S41: the 2D input features of the amplitude spectrum of the mixed signal are mapped to a higher-dimensional space by the encoder: ； S42: separation is completed in the high-dimensional space by the estimator; S43: the high-dimensional vector calculated by the estimator is restored to the same dimension mask as the encoder input by the decoder: 。 7. The method of claim 1, wherein the method is based on a 2D-TASNet. The specific process of reconstructing the separated human voice using the target mask in step S5 is as follows: After the mask and the time-frequency spectrum of the mixed signal are multiplied, the time-frequency spectrum containing only the target human voice is obtained, and then the inverse STFT is performed to obtain the separated human voice. The specific formula is as follows: ； Where s1(t) is the separated human voice signal, M is the estimated mask, S is the time-frequency spectrum of the mixed signal, and ISTFT is the inverse Fourier transform.

Images