A speech noise reduction method, system, medium and device based on subspace filtering

Abstract

The application discloses a speech noise reduction method and system based on subspace filtering, a medium and equipment, and belongs to the field of software engineering and communication technology.The method is as follows: noisy speech is subjected to frame processing; autocorrelation coefficients and mel cepstrum coefficients of each frame are extracted, spliced to form a feature vector, and input into a noise estimation network to predict the noise autocorrelation coefficients of each frame; a subspace separation matrix is constructed according to the autocorrelation coefficients, noise autocorrelation coefficients and unit matrix of each frame, and eigenvalue decomposition is performed to obtain eigenvalues and a characteristic matrix; the frame data is transformed to a subspace by using the characteristic matrix to obtain subspace signals in each direction; a subspace gain matrix is calculated in combination with the eigenvalues, the subspace signals are filtered to obtain a noise reduction signal; the noise reduction signal is inversely transformed according to the characteristic matrix to obtain noise reduction speech frames, which are spliced to obtain noise reduction speech. Therefore, by implementing the application, the speech noise reduction effect can be improved, and the algorithm calculation amount and complexity can be reduced.

Description

Technical Field

[0001] This invention belongs to the fields of software engineering and communication technology, and relates to a speech noise reduction method, system, medium and device based on subspace filtering. Background Technology

[0002] Speech denoising technology, as a core means to improve speech quality and intelligibility, plays a crucial role in increasingly popular application scenarios such as speech recognition, communication calls, teleconferences, scene recording, and hearing assistance, possessing broad application prospects and significant research value. Thanks to the integration of deep learning technology, the field of speech denoising has achieved significant breakthroughs in recent years. Single-channel denoising methods, due to the widespread availability of single-microphone devices and the cost-effectiveness of their algorithm implementation, have a wider practical application base. Currently, the mainstream single-channel speech denoising technologies include three categories: traditional signal processing methods, supervised methods based on machine learning and deep learning, and the fusion of deep learning and traditional methods. Traditional signal processing methods are based on digital signal processing technology and mostly fall under the category of unsupervised learning, requiring no prior labeled data. Supervised methods based on machine learning and deep learning have emerged with the advancement of machine learning technology, significantly improving the performance ceiling of denoising models. The fusion strategy of deep learning and traditional methods explores the integration of the powerful learning capabilities of neural networks into or assists traditional denoising frameworks, becoming one of the current research hotspots.

[0003] However, traditional digital signal processing methods rely on numerous theoretical assumptions that are difficult to meet in complex and diverse production environments. This results in them being able to process only a single type of noise, performing poorly with non-stationary noise or low signal-to-noise ratio conditions, and even generating musical noise. While deep learning methods can achieve good results in various application environments, their algorithmic structure needs to be designed to be complex in order to accurately extract features from speech. They also require massive amounts of data samples and long training times to achieve good generalization, as well as a large amount of memory and computing power, making them difficult to deploy in production environments with limited hardware. Summary of the Invention

[0004] This application provides a speech denoising method, system, medium, and device based on subspace filtering, which can improve the speech denoising effect and reduce the computational load and complexity of the algorithm.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a speech denoising method based on subspace filtering, comprising:

[0006] Noisy speech is acquired and processed into frames to obtain the data for each frame.

[0007] Calculate the autocorrelation coefficient and Mel-frequency cepstral coefficient of each frame data, and concatenate them to construct the feature vector corresponding to each frame data;

[0008] Each of the aforementioned feature vectors is input into a preset noise estimation network to perform noise autocorrelation coefficient prediction processing, thereby obtaining the noise autocorrelation coefficient corresponding to the noise in each frame of data.

[0009] Based on the autocorrelation coefficients, noise autocorrelation coefficients and identity matrices, a subspace separation matrix is ​​constructed and eigenvalue decomposition is performed to obtain eigenvalue data and eigenma matrices.

[0010] Based on the feature matrix, each frame data is transformed to a subspace to obtain subspace signals in each direction. The subspace gain matrix is ​​calculated by combining the feature value data, and the subspace signals are filtered to obtain the corresponding noise-reduced signals.

[0011] Based on the feature matrix, each of the noise-reduced signals is inversely transformed to obtain each noise-reduced speech data frame, and then frame merging is performed to output the noise-reduced speech corresponding to the noisy speech.

[0012] Compared to existing technologies, the embodiments of this application have the following beneficial effects: By acquiring noisy speech and performing frame-by-frame processing, continuous speech signals are converted into independently analyzable frame-by-frame data, providing a basic processing unit for subsequent feature extraction and noise estimation; the autocorrelation coefficient and Mel-Cepstral Coefficient of each frame-by-frame data are calculated and concatenated to construct a feature vector. The autocorrelation coefficient reflects the temporal correlation of the signal, and the Mel-Cepstral Coefficient captures the spectral features that the human ear is sensitive to. The combination of the two provides multi-dimensional acoustic feature input for noise estimation; the feature vector is input into a preset noise estimation network to predict the noise autocorrelation coefficient. The network model directly learns the statistical laws of noise, avoiding the transmission of noise. Traditional methods rely on estimation biases caused by manual assumptions. This approach constructs a subspace separation matrix based on the autocorrelation coefficient, noise autocorrelation coefficient, and identity matrix, and performs eigenvalue decomposition. Mathematical transformations map noisy speech to a high-dimensional space, separating speech and noise along eigenvector directions. The framed data is then transformed to the subspace based on the eigenvalue matrix, and a gain matrix is ​​calculated for filtering. Targeted gains are applied to signals in different directions to achieve a dynamic balance between noise suppression and speech preservation. Finally, the denoised signal undergoes an inverse transformation based on the eigenvalue matrix and is combined into frames, restoring the subspace-processed signal to temporal speech, ensuring the continuity and integrity of the output signal. The overall solution combines a noise estimation network and a subspace filtering framework. This overcomes the limitations of traditional subspace methods that rely on prior noise assumptions by leveraging the network's accurate noise modeling capabilities. Furthermore, subspace eigenvalue decomposition achieves low-complexity signal separation, avoiding the high computational cost of pure deep learning models. Ultimately, this results in a synergistic improvement in speech denoising performance and algorithm efficiency in complex environments.

[0013] In some embodiments of the first aspect of this application, the noise estimation network includes an input linear layer, a multi-scale convolutional layer, an LSTM network layer, and an output linear layer; the step of inputting each feature vector into a preset noise estimation network to perform noise autocorrelation coefficient prediction processing to obtain the noise autocorrelation coefficient corresponding to the noise in each frame of data includes:

[0014] Based on the input linear layer, the feature vector is dimension-aligned and then subjected to nonlinear transformation using the tanh activation function to obtain the first feature;

[0015] Based on the multi-scale convolutional layer, the first feature is convolved to extract the noise hidden features, which are then encoded into a high-dimensional latent space to obtain the second feature.

[0016] Based on the LSTM network layer, the second feature is modeled as a temporal feature, and the noisy temporal feature is output.

[0017] Based on the output linear layer, the noise temporal characteristics are mapped to obtain the noise autocorrelation coefficient.

[0018] Compared to existing technologies, the above embodiments have the following advantages: The noise estimation network performs dimension alignment on the feature vector through the input linear layer and performs nonlinear transformation using the tanh activation function, thereby normalizing and nonlinearly mapping the original features and enhancing the feature representation capability; the multi-scale convolutional layer performs convolution operations on the first feature, extracts multi-frequency hidden features of noise through convolutional kernels of different scales and encodes them into a high-dimensional latent space, improving the ability to capture complex noise patterns; the LSTM network layer models the temporal features of the second feature, uses a recurrent structure to learn the dynamic change law of noise over time, and outputs more robust noise temporal features; the output linear layer maps the noise temporal features to obtain the noise autocorrelation coefficient, accurately transforming the high-dimensional features into the key parameters required for subspace filtering, further improving the accuracy and specificity of noise estimation.

[0019] In some embodiments of the first aspect of this application, the step of constructing a subspace separation matrix and performing eigenvalue decomposition processing based on the autocorrelation coefficients, noise autocorrelation coefficients, and identity matrix to obtain eigenvalue data and a feature matrix includes:

[0020] Construct a noisy speech autocorrelation matrix based on the autocorrelation coefficients described above;

[0021] Based on the noise autocorrelation coefficients described above, construct a noise autocorrelation matrix;

[0022] Based on the noisy speech autocorrelation matrix, the noise autocorrelation matrix, and the identity matrix, a subspace separation matrix is ​​constructed, as follows: Where ∑ represents the subspace separation matrix. R represents the noise autocorrelation matrix n The transpose matrix, R x =R y -R n R represents the autocorrelation matrix of the speech signal. y Let I represent the noisy speech autocorrelation matrix;

[0023] The subspace separation matrix is ​​subjected to eigenvalue decomposition to obtain eigenvalue data and a feature matrix; wherein, in the eigenvalue data, the eigenvalues ​​representing the direction of the speech signal are greater than zero, and the eigenvalues ​​representing the direction of the noise signal are less than zero, and the eigenvalue data and the feature matrix satisfy the following constraints:

[0024] Where Λ represents the eigenvalue data, V T This represents the transpose of the characteristic matrix V.

[0025] Compared with existing technologies, the above embodiments have the following beneficial effects: Noisy speech autocorrelation matrix and noise autocorrelation matrix are constructed based on autocorrelation coefficients to quantify the temporal statistical characteristics of the mixed signal and noise, providing a data foundation for subspace separation; a subspace separation matrix is ​​constructed by multiplying the inverse of the noise autocorrelation matrix by the noisy speech autocorrelation matrix and then subtracting the identity matrix. This matrix construction method preprocesses the noisy speech signal by "whitening" it through the inverse operation of the noise matrix, transforming the energy distribution difference between the speech signal and noise into the distinguishability of matrix features; eigenvalue decomposition is performed on the constructed subspace separation matrix, and two sets of constraints are introduced: the first set of constraints multiplies the speech autocorrelation matrix by the transpose of the feature matrix and then by the feature matrix again to obtain a diagonalized eigenvalue matrix, concentrating speech energy on the eigenvalues ​​of the main diagonal; the second set of constraints multiplies the noise autocorrelation matrix by the transpose of the feature matrix and then by the feature matrix again to obtain the identity matrix, achieving decorrelation processing of the noise signal and eliminating anisotropic noise interference. Furthermore, since the subspace separation matrix is ​​essentially the product of the inverse of the noise autocorrelation matrix and the speech autocorrelation matrix, and the inverse of the noise autocorrelation matrix has positive definite properties, when there is a speech signal in the subspace direction, the corresponding eigenvalue is positive; when there is only noise in the subspace direction (speech energy is zero), the eigenvalue naturally becomes negative or zero due to matrix operations (in reality, it is mostly negative due to noise estimation errors). Therefore, compared with existing technologies that rely on empirical thresholds or eigenvalue sorting to distinguish speech and noise directions (such as traditional subspace methods that require manually setting a "significant eigenvalue" screening threshold), this method, through matrix inversion operations and orthogonalization constraints, makes the eigenvalue sign and signal type strictly correspond, automatically separating speech and noise directions without manual intervention, reducing reliance on prior knowledge and improving the stability and accuracy of separation in complex noise environments.

[0026] In some embodiments of the first aspect of this application, transforming each framed data into a subspace according to the feature matrix to obtain subspace signals in each direction includes:

[0027] Using the order of the autocorrelation coefficient of each frame data as the frame window size, the frame data is subjected to overlapping secondary frame division processing to obtain each subframe data.

[0028] Based on the feature matrix, the subframe data is transformed into a subspace to obtain subspace signals corresponding to each direction.

[0029] Compared with existing technologies, the above embodiments have the following beneficial effects: using the order of the autocorrelation coefficient of the framed data as the framed window size, the framed data is subjected to overlapping secondary framed processing to obtain subframe data. By matching the framed granularity with the statistical characteristic dimension of the autocorrelation coefficient, the temporal resolution of the subframe data is ensured to be compatible with the matrix dimension of the subsequent subspace transformation. At the same time, the overlapping processing reduces information loss caused by the framed boundary effect. Furthermore, the subframe data is transformed to the subspace through the feature matrix to obtain the subspace signal corresponding to each direction, realizing the directional separation of noisy speech in high-dimensional space, and providing a structured data foundation for subsequent filtering processing for different signal directions.

[0030] In some embodiments of the first aspect of this application, calculating the subspace gain matrix by combining the eigenvalue data includes:

[0031] Based on the sign of each feature value in the feature value data, the direction of each noise signal is selected and the corresponding feature value is adjusted to zero;

[0032] Based on the adjusted eigenvalue data, combined with the preset gain calculation function and the Lagrange multipliers in each signal direction, the subspace gain in each signal direction is calculated; wherein, each of the Lagrange multipliers is determined by the signal-to-noise ratio in the corresponding signal direction;

[0033] Based on the gains of each subspace, the subspace gain matrix is ​​constructed as follows:

[0034] Among them, g kk This represents the subspace gain in the k-th signal direction of the subspace gain matrix. μ(v) represents the ratio of speech energy to noise energy in the k-th signal direction of the eigenvalue data. k ) represents the Lagrange multiplier corresponding to the k-th signal direction in the subspace, m represents the number of speech signal directions in the subspace, and n represents the total number of signal directions in the subspace.

[0035] Compared to existing technologies, the above embodiments have the following advantages: Noise signal directions are filtered based on the sign of each feature value in the feature value data, and the corresponding feature values ​​are adjusted to zero. Speech and noise directions are directly distinguished through mathematical properties, avoiding misjudgments caused by relying on empirical thresholds in traditional methods. Simultaneously, based on the adjusted feature value data, the subspace gain for each direction is calculated using a preset gain calculation function and a Lagrange multiplier determined by the signal-to-noise ratio (SNR). This dynamically correlates gain allocation with the SNR of the signal direction. The subspace gain in the k-th signal direction is calculated by dividing the ratio of speech energy to noise energy in that direction by the sum of this ratio and the corresponding Lagrange multiplier, achieving a balance between noise suppression and speech preservation. Finally, a subspace gain matrix is ​​constructed, which can be used to apply targeted filtering to signals in different directions using diagonal elements, ensuring that noise direction signals are completely suppressed and speech direction signals are adaptively enhanced according to the SNR.

[0036] In some embodiments of the first aspect of this application, each of the Lagrange multipliers is determined by the signal-to-noise ratio in the corresponding signal direction, including:

[0037] Based on the eigenvalue data, the signal-to-noise ratio (SNR) for each signal direction within the subspace is calculated; the calculation method is as follows:

[0038] Among them, SNR(v k ) represents the direction v of the k-th signal. k Signal-to-noise ratio;

[0039] Based on the aforementioned signal-to-noise ratios, the Lagrange multipliers corresponding to each signal direction are determined according to a preset matching rule; wherein the matching rule is as follows:

[0040]

[0041] Where, μ(v) k ) represents the Lagrange multiplier corresponding to the k-th signal direction in the subspace, max μ and min μ Let up represent the preset maximum and minimum Lagrange multipliers, respectively. snr and low snr These represent the preset maximum signal-to-noise ratio and minimum signal-to-noise ratio, respectively.

[0042] Compared with existing technologies, the above embodiments have the following beneficial effects: the signal-to-noise ratio (SNR) of each signal direction in the subspace is calculated based on the feature value data, and the signal quality of each direction is quantified by the ratio of speech energy to noise energy in that direction; further, the Lagrange multiplier is determined based on a preset matching rule, and when the SNR is between the minimum and maximum SNR, a smooth transition is achieved through a linear function to ensure the continuity of gain adjustment; when the SNR is higher than the maximum SNR, the minimum Lagrange multiplier is used to preserve the high SNR speech signal, and when the SNR is lower than the minimum SNR, the maximum Lagrange multiplier is used to enhance noise suppression. The piecewise function dynamically adapts to different SNR scenarios, avoiding over-suppression or under-suppression problems caused by a single fixed parameter.

[0043] In a second aspect, the present invention also provides a speech denoising system based on subspace filtering, comprising: a data acquisition module, a feature extraction module, a model prediction module, a subspace construction module, a denoising module, and a result output module;

[0044] The data acquisition module is used to acquire noisy speech and perform frame-by-frame processing to obtain each frame of data.

[0045] The feature extraction module is used to calculate the autocorrelation coefficient and Mel-frequency cepstral coefficient of each frame data, and to concatenate them to construct the feature vector corresponding to each frame data.

[0046] The model prediction module is used to input each feature vector into a preset noise estimation network to perform noise autocorrelation coefficient prediction processing, and obtain the noise autocorrelation coefficient corresponding to the noise in each frame of data.

[0047] The subspace construction module is used to construct a subspace separation matrix and perform eigenvalue decomposition processing based on the autocorrelation coefficients, noise autocorrelation coefficients and identity matrix to obtain eigenvalue data and eigenmatrix.

[0048] The noise reduction module is used to transform each frame data into a subspace according to the feature matrix to obtain subspace signals in each direction, and calculate the subspace gain matrix in combination with the feature value data to filter each subspace signal to obtain the corresponding noise reduction signal.

[0049] The result output module is used to perform inverse transformation on each of the noise-reduced signals according to the feature matrix to obtain each noise-reduced speech data frame, and perform frame merging processing to output the noise-reduced speech corresponding to the noisy speech.

[0050] Compared to existing technologies, the above embodiments of this application have the following beneficial effects: By acquiring noisy speech and performing frame-by-frame processing, continuous speech signals are converted into independently analyzable frame-by-frame data, providing a basic processing unit for subsequent feature extraction and noise estimation; the autocorrelation coefficient and Mel-frequency cepstral coefficient of each frame-by-frame data are calculated and concatenated to construct a feature vector. The autocorrelation coefficient reflects the temporal correlation of the signal, and the Mel-frequency cepstral coefficient captures the spectral features that the human ear is sensitive to. The combination of the two provides multi-dimensional acoustic feature input for noise estimation; the feature vector is input into a preset noise estimation network to predict the noise autocorrelation coefficient. The network model directly learns the statistical laws of noise, avoiding... Traditional methods rely on estimation biases caused by manual assumptions. This approach constructs a subspace separation matrix based on the autocorrelation coefficient, noise autocorrelation coefficient, and identity matrix, and performs eigenvalue decomposition. Through mathematical transformation, noisy speech is mapped to a high-dimensional space, achieving separation of speech and noise along eigenvector directions. The framed data is then transformed to the subspace based on the eigenvalue matrix, and a gain matrix is ​​calculated for filtering. Targeted gains are applied to signals in different directions to achieve a dynamic balance between noise suppression and speech preservation. Finally, the denoised signal undergoes an inverse transformation based on the eigenvalue matrix and is combined into frames, restoring the subspace-processed signal to time-domain speech, ensuring the continuity and integrity of the output signal. The overall solution combines a noise estimation network and a subspace filtering framework. This overcomes the limitations of traditional subspace methods that rely on prior noise assumptions by leveraging the network's accurate noise modeling capabilities. Furthermore, it achieves low-complexity signal separation through subspace eigenvalue decomposition, avoiding the high computational cost of pure deep learning models. Ultimately, this results in a synergistic improvement in speech denoising performance and algorithm efficiency in complex environments.

[0051] In some embodiments of the second aspect of this application, the noise estimation network includes an input linear layer, a multi-scale convolutional layer, an LSTM network layer, and an output linear layer;

[0052] The model prediction module includes: a linear input unit, a convolutional unit, a temporal modeling unit, and a linear output unit;

[0053] The linear input unit is used to align the feature vector in dimensions according to the input linear layer and perform nonlinear transformation processing through the tanh activation function to obtain the first feature.

[0054] The convolutional unit is used to perform a convolution operation on the first feature according to the multi-scale convolutional layer, extract the noise hidden feature, and encode it into a high-dimensional latent space to obtain the second feature;

[0055] The temporal modeling unit is used to perform temporal feature modeling on the second feature based on the LSTM network layer and output noisy temporal features;

[0056] The linear output unit is used to map the noise temporal characteristics according to the output linear layer to obtain the noise autocorrelation coefficient.

[0057] Compared to existing technologies, the above embodiments have the following advantages: The noise estimation network performs dimension alignment on the feature vector through the input linear layer and performs nonlinear transformation using the tanh activation function, thereby normalizing and nonlinearly mapping the original features and enhancing the feature representation capability; the multi-scale convolutional layer performs convolution operations on the first feature, extracts multi-frequency hidden features of noise through convolutional kernels of different scales and encodes them into a high-dimensional latent space, improving the ability to capture complex noise patterns; the LSTM network layer models the temporal features of the second feature, uses a recurrent structure to learn the dynamic change law of noise over time, and outputs more robust temporal noise features; the output linear layer maps the temporal noise features to obtain the noise autocorrelation coefficient, accurately transforming the high-dimensional features into the key parameters required for subspace filtering, further improving the accuracy and specificity of noise estimation.

[0058] Thirdly, the present invention also provides a speech noise reduction device based on subspace filtering, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, it implements the steps of any of the speech noise reduction methods based on subspace filtering of the present invention.

[0059] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the speech denoising methods based on subspace filtering of the present invention. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating a speech denoising method based on subspace filtering provided in some embodiments of the present invention.

[0061] Figure 2 This is a schematic diagram of a speech noise reduction system based on subspace filtering provided in some embodiments of the present invention.

[0062] Figure 3 : This is a structural diagram of a speech noise reduction device based on subspace filtering provided in some embodiments of the present invention.

[0063] Figure 4 This is an overall framework diagram of a speech denoising method based on subspace filtering provided in some embodiments of the present invention.

[0064] Figure 5 : This is a model structure diagram of a noise estimation network provided in some embodiments of the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Example 1:

[0067] Please refer to Figure 1 To improve speech denoising performance and reduce computational load and complexity, an embodiment of the present invention provides a speech denoising method based on subspace filtering, comprising steps S1 to S6:

[0068] Step S1: Acquire noisy speech and perform frame segmentation processing to obtain each frame data.

[0069] Step S2: Calculate the autocorrelation coefficient and Mel-frequency cepstral coefficient of each frame data, and concatenate them to construct the feature vector corresponding to each frame data.

[0070] In practice, the noisy speech is first divided into frames in the time domain with a window size (nfft) of 512 and a window sliding step (hop_len) of 256. Then, the autocorrelation coefficients R of the first 64 orders of each frame are calculated. y The 64th-order Mel-frequency cepstral coefficients (MFCCs) are then directly concatenated to obtain the corresponding feature vector, which serves as the input to the subsequent model, as shown below: Among them, M0-M 63 This represents the 64th order Mel cepstral coefficients.

[0071] Step S3: Input each of the feature vectors into a preset noise estimation network to perform noise autocorrelation coefficient prediction processing, and obtain the noise autocorrelation coefficient corresponding to the noise in each frame of data.

[0072] Furthermore, the noise estimation network includes an input linear layer, a multi-scale convolutional layer, an LSTM network layer, and an output linear layer;

[0073] Step S3 can be implemented through the following preferred embodiments, including steps S31-S34, as follows:

[0074] S31: Based on the input linear layer, the feature vector is dimension-aligned and subjected to nonlinear transformation processing using the tanh activation function to obtain the first feature;

[0075] S32: Based on the multi-scale convolutional layer, perform a convolution operation on the first feature to extract the noise hidden features and encode them into a high-dimensional latent space to obtain the second feature;

[0076] S33: Based on the LSTM network layer, perform temporal feature modeling on the second feature and output noisy temporal features;

[0077] S34: Based on the output linear layer, the noise temporal characteristics are mapped to obtain the noise autocorrelation coefficient.

[0078] In this preferred embodiment, the noise estimation network performs dimension alignment on the feature vector through an input linear layer and performs nonlinear transformation using the tanh activation function to normalize and nonlinearly map the original features, thereby enhancing the feature representation capability. A multi-scale convolutional layer performs convolution operations on the first feature, extracting multi-frequency hidden features of the noise through convolutional kernels of different scales and encoding them into a high-dimensional latent space, improving the ability to capture complex noise patterns. An LSTM network layer models the second feature based on temporal characteristics, using a recurrent structure to learn the dynamic changes of noise over time, outputting more robust temporal noise features. The output linear layer maps the temporal noise features to obtain the noise autocorrelation coefficient, accurately transforming the high-dimensional features into key parameters required for subspace filtering, further improving the accuracy and specificity of noise estimation.

[0079] like Figure 5 The diagram shows the model structure of a noise estimation network. In steps S1-S2, after extracting the corresponding features from the noisy speech, these features are input into the noise estimation network to predict the noise autocorrelation coefficient. First, the feature inputs (i.e., autocorrelation coefficient and Mel-frequency cepstral coefficient) are aligned using a linear layer and a tanh activation function. Then, multi-scale convolution is used to extract the noise hidden features and encode them into a high-dimensional latent space. Finally, an LSTM network is used to extract the temporal features of the noise and map them to the noise autocorrelation coefficient. The output of the model is shown below:

[0080] Among them, R n This represents the noise autocorrelation coefficient.

[0081] Furthermore, when training the noise estimation network, the Adam optimizer is used for training, and the loss function is the mean absolute error (MAE), as shown below:

[0082] Where x (i) It is the i-th input vector, y (i) h is the vector that is the expected output related to the training objective. θ () denotes the implicit function of the neural network representation with parameter θ. M is the total number of training vectors.

[0083] Step S4: Based on the autocorrelation coefficients, noise autocorrelation coefficients, and identity matrix, construct the subspace separation matrix and perform eigenvalue decomposition to obtain eigenvalue data and eigenma.

[0084] Furthermore, step S4 can be implemented through the following preferred embodiments, including S41-S44, as follows:

[0085] S41: Construct a noisy speech autocorrelation matrix based on the autocorrelation coefficients described above;

[0086] S42: Construct a noise autocorrelation matrix based on the noise autocorrelation coefficients described above;

[0087] S43: Based on the noisy speech autocorrelation matrix, the noise autocorrelation matrix, and the identity matrix, a subspace separation matrix is ​​constructed, represented as follows: Where ∑ represents the subspace separation matrix. R represents the noise autocorrelation matrix n The transpose matrix, R x =R y -R n R represents the autocorrelation matrix of the speech signal. y Let I represent the noisy speech autocorrelation matrix;

[0088] S44: Perform eigenvalue decomposition on the subspace separation matrix to obtain eigenvalue data and a feature matrix; wherein, in the eigenvalue data, the eigenvalues ​​representing the direction of the speech signal are greater than zero, and the eigenvalues ​​representing the direction of the noise signal are less than zero, and the eigenvalue data and the feature matrix satisfy the following constraints:

[0089] Where Λ represents the eigenvalue data, V T This represents the transpose of the characteristic matrix V.

[0090] In this preferred embodiment, a noisy speech autocorrelation matrix and a noise autocorrelation matrix are constructed based on the autocorrelation coefficient to quantify the temporal statistical characteristics of the mixed signal and noise, providing a data foundation for subspace separation. A subspace separation matrix is ​​constructed by multiplying the inverse of the noise autocorrelation matrix by the noisy speech autocorrelation matrix and then subtracting the identity matrix. This matrix construction method preprocesses the noisy speech signal by "whitening" it through the inverse operation of the noise matrix, transforming the energy distribution difference between the speech signal and noise into the distinguishability of matrix features. The constructed subspace separation matrix is ​​then subjected to eigenvalue decomposition, and two sets of constraints are introduced: the first set of constraints multiplies the speech autocorrelation matrix by the transpose of the eigenvalue matrix and then by the eigenvalue matrix again, resulting in a diagonalized eigenvalue matrix that concentrates speech energy on the eigenvalues ​​along the main diagonal; the second set of constraints multiplies the noise autocorrelation matrix by the transpose of the eigenvalue matrix and then by the eigenvalue matrix again, resulting in an identity matrix, thus achieving decorrelation processing of the noise signal and eliminating anisotropic noise interference. Furthermore, since the subspace separation matrix is ​​essentially the product of the inverse of the noise autocorrelation matrix and the speech autocorrelation matrix, and the inverse of the noise autocorrelation matrix has positive definite properties, when there is a speech signal in the subspace direction, the corresponding eigenvalue is positive; when there is only noise in the subspace direction (speech energy is zero), the eigenvalue naturally becomes negative or zero due to matrix operations (in reality, it is mostly negative due to noise estimation errors). Therefore, compared with existing technologies that rely on empirical thresholds or eigenvalue sorting to distinguish speech and noise directions (such as traditional subspace methods that require manually setting a "significant eigenvalue" screening threshold), this method, through matrix inversion operations and orthogonalization constraints, makes the eigenvalue sign and signal type strictly correspond, automatically separating speech and noise directions without manual intervention, reducing reliance on prior knowledge and improving the stability and accuracy of separation in complex noise environments.

[0091] Step S5: Based on the feature matrix, transform each frame data to a subspace to obtain subspace signals in each direction, and calculate the subspace gain matrix in combination with the feature value data. Then, filter each subspace signal to obtain the corresponding noise-reduced signal.

[0092] Furthermore, when transforming the framed data to the subspace in step S5, it can be implemented through the following preferred embodiments, including steps S51-S52, as follows:

[0093] S51: Using the order of the autocorrelation coefficient of each frame data as the frame window size, perform overlapping secondary frame processing on each frame data to obtain each subframe data.

[0094] S52: Based on the feature matrix, transform the subframe data to a subspace to obtain subspace signals corresponding to each direction.

[0095] In practice, the order of the autocorrelation coefficient (e.g., the first 64 orders) determines the order of the subspace, and they are usually the same. To meet the computational requirements of solving a 64-order subspace, the current framed data needs to be further divided into subframes with nfft = 64 and hop_len = 32 before subsequent subspace transformation and filtering are performed. Of course, depending on the actual hardware conditions, subspaces of various orders can also be constructed, such as 4-order, 16-order, 32-order, or up to 64-order subspaces.

[0096] In this preferred embodiment, the order of the autocorrelation coefficient of the framed data is used as the framed window size. The framed data is subjected to overlapping secondary framed processing to obtain subframe data. By matching the framed granularity with the statistical characteristic dimension of the autocorrelation coefficient, the temporal resolution of the subframe data is ensured to be compatible with the matrix dimension of the subsequent subspace transformation. At the same time, the overlapping processing reduces information loss caused by the framed boundary effect. Furthermore, the subframe data is transformed to the subspace through the feature matrix to obtain the subspace signal corresponding to each direction, realizing the directional separation of noisy speech in high-dimensional space, and providing a structured data foundation for subsequent filtering processing for different signal directions.

[0097] Furthermore, step S5, in calculating the subspace gain matrix, can be implemented through the following preferred embodiments, including steps S53-S54, as detailed below:

[0098] S53: Based on the sign of each feature value in the feature value data, filter out the direction of each noise signal and adjust the corresponding feature value to zero;

[0099] S54: Based on the adjusted eigenvalue data, combined with the preset gain calculation function and the Lagrange multipliers in each signal direction, calculate the subspace gain in each signal direction; wherein, each of the Lagrange multipliers is determined by the signal-to-noise ratio in the corresponding signal direction;

[0100] Based on the gains of each subspace, the subspace gain matrix is ​​constructed as follows:

[0101] Among them, g kk This represents the subspace gain in the k-th signal direction of the subspace gain matrix. μ(v) represents the ratio of speech energy to noise energy in the k-th signal direction of the eigenvalue data. k ) represents the Lagrange multiplier corresponding to the k-th signal direction in the subspace, m represents the number of speech signal directions in the subspace, and n represents the total number of signal directions in the subspace.

[0102] In this preferred embodiment, the direction of the noise signal is filtered according to the positive or negative sign of each feature value in the feature value data, and the corresponding feature value is adjusted to zero. The direction of speech and noise is directly distinguished by mathematical properties, avoiding misjudgment caused by the reliance on empirical thresholds in traditional methods. At the same time, based on the adjusted feature value data, the subspace gain of each direction is calculated by combining a preset gain calculation function and a Lagrange multiplier determined by the signal-to-noise ratio, so that the gain allocation is dynamically related to the signal-to-noise ratio of the signal direction. The subspace gain of the k-th signal direction is calculated by dividing the ratio of speech energy to noise energy in that direction by the sum of this ratio and the corresponding Lagrange multiplier, so as to achieve a balance between noise suppression and speech preservation. Finally, a subspace gain matrix is ​​constructed, which can be used to apply targeted filtering to signals in different directions through diagonal elements to ensure that the noise direction signal is completely suppressed and the speech direction signal is adaptively enhanced according to the signal-to-noise ratio.

[0103] Furthermore, in step S54, each of the Lagrange multipliers is determined by the signal-to-noise ratio in the corresponding signal direction, which can be implemented through the following preferred embodiments, including steps S541-S542, as follows:

[0104] S541: Based on the eigenvalue data, calculate the signal-to-noise ratio (SNR) for each signal direction within the subspace; wherein the calculation method is as follows:

[0105] Among them, SNR(v k ) represents the direction v of the k-th signal. k Signal-to-noise ratio;

[0106] S542: Based on the signal-to-noise ratios, determine the Lagrange multipliers corresponding to each signal direction according to a preset matching rule; wherein the matching rule is as follows:

[0107]

[0108] Where, μ(v) k ) represents the Lagrange multiplier corresponding to the k-th signal direction in the subspace, max μ and min μ Let up represent the preset maximum and minimum Lagrange multipliers, respectively. snr and low snr These represent the preset maximum signal-to-noise ratio and minimum signal-to-noise ratio, respectively.

[0109] In this preferred embodiment, the signal-to-noise ratio (SNR) of each signal direction within the subspace is calculated based on eigenvalue data. The signal quality in each direction is quantified by the ratio of speech energy to noise energy in that direction. Furthermore, the Lagrange multiplier is determined based on a preset matching rule. When the SNR is between the minimum and maximum SNR, a smooth transition is achieved through a linear function to ensure the continuity of gain adjustment. When the SNR is higher than the maximum SNR, the minimum Lagrange multiplier is used to preserve the high SNR speech signal, and when the SNR is lower than the minimum SNR, the maximum Lagrange multiplier is used to enhance noise suppression. By dynamically adapting to different SNR scenarios through a piecewise function, the over-suppression or under-suppression problem caused by a single fixed parameter is avoided.

[0110] like Figure 4 The diagram shows the overall framework of a speech denoising method based on subspace filtering. The overall formula for subspace filtering is as follows:

[0111] (equivalent to) Figure 4 (Subspace transformation + filter gain + inverse subspace transformation); where y represents the noisy signal. Let G represent the denoised signal, and G represent the subspace gain matrix, which is a diagonal matrix where the diagonal elements represent the subspace gain g along each eigenvector. kk .

[0112] In practice, when the speech signal and noise signal are uncorrelated and the autocorrelation coefficient of the noise signal is known, the noisy signal in the time domain is mapped to a high-dimensional subspace. Ideally, the clean speech signal and noise signal in this subspace are separable from each other.

[0113] Let the subspace transformation matrix P = {p1, p2, ..., p n}, then the feature matrices of the corresponding parts of the speech signal and the noise signal are respectively P clean ={p1,p2,…,p m}, P noise ={p m+1 ,p m+2 ,…,p n}, that is, P = {P clean ,P noise}, where m represents the number of speech signal directions in the subspace, n represents the total number of signal directions, clean represents speech, and noise represents noise.

[0114] To ensure the separation of speech and noise signals in the subspace, the subspace separation matrix is ​​first found. Make sure the following constraints are satisfied:

[0115] Where I represents the identity matrix, i.e., a square matrix with all elements on the main diagonal being 1 and all other elements being 0; Λ={Λ x ,Λ n} and V are the eigenvalue data and eigenvectors of the subspace separation matrix, respectively, and Λ x The eigenvalue representing the direction of the speech signal and greater than 0, Λ n The characteristic value representing the direction of the noise signal is less than 0.

[0116] Next, after transforming the subframe data to subspace, a subspace gain matrix is ​​constructed, and the gain g is... kk Represented as:

[0117] in, This represents the amplitude of the speech signal x in the k-th signal direction, which is the ratio of speech energy to noise energy in the k-th signal direction of the eigenvalue data. When the dimension k of the subspace gain matrix is ​​from m+1 to n, it takes the value 0, indicating that Λ represents the noise. n Setting all terms to 0 eliminates noise in the noise direction of the subspace. Ideally, the Lagrange multiplier μ = 0, assuming the speech signal is absolutely uncorrelated with the noise, and the noise is eliminated in the direction of the speech signal Λ. x If the signal does not contain noise, then it is a binary function with a gain of 1 / 0, i.e., frequency subtraction. However, in reality, it is assumed that there may be a correlation between the speech signal and noise, and the gain Λ in the direction of the speech signal... x It also contains a small amount of noise, so let μ > 0, then the gain g kk μ needs to be calculated based on the signal-to-noise ratio in the subspace domain, which is the Wiener filtering method.

[0118] From the expression for Σ above, we can see that the k-th signal direction v of the speech signal in the subspace k Energy on Noise signal at v k Energy on

[0119] This yields the signal-to-noise ratio in each direction (frequency band) of the subspace. Right now

[0120] Based on this signal-to-noise ratio, μ is set as follows:

[0121]

[0122] As a preferred option, the parameter can be set to low. snr =0,up snr =20,min μ =0.7,max μ =2.

[0123] Finally, after calculating and constructing the subspace gain matrix, the signals in the subspace are filtered to obtain the corresponding denoised signals.

[0124] Step S6: Based on the feature matrix, perform inverse transformation on each of the noise-reduced signals to obtain each noise-reduced speech data frame, perform frame merging processing, and output the noise-reduced speech corresponding to the noisy speech.

[0125] In summary, compared with the prior art, the above embodiments of this application have the following beneficial effects: By acquiring noisy speech and performing frame-by-frame processing, continuous speech signals are converted into independently analyzable frame-by-frame data, providing a basic processing unit for subsequent feature extraction and noise estimation; the autocorrelation coefficient and Mel-frequency cepstral coefficient of each frame-by-frame data are calculated and concatenated to construct a feature vector. The autocorrelation coefficient reflects the temporal correlation of the signal, and the Mel-frequency cepstral coefficient captures the spectral features that the human ear is sensitive to. The combination of the two provides multi-dimensional acoustic feature input for noise estimation; the feature vector is input into a preset noise estimation network to predict the noise autocorrelation coefficient, and the network model directly learns the statistical laws of noise, avoiding... This approach avoids estimation biases caused by reliance on manual assumptions in traditional methods. It constructs a subspace separation matrix based on the autocorrelation coefficient, noise autocorrelation coefficient, and identity matrix, and performs eigenvalue decomposition. Through mathematical transformation, noisy speech is mapped to a high-dimensional space, achieving separation of speech and noise along eigenvector directions. The framed data is transformed to the subspace based on the eigenvalue matrix, and a gain matrix is ​​calculated for filtering. Targeted gains are applied to signals in different directions to achieve a dynamic balance between noise suppression and speech preservation. Finally, the denoised signal undergoes an inverse transformation based on the eigenvalue matrix and is combined into frames, restoring the subspace-processed signal to time-domain speech, ensuring the continuity and integrity of the output signal. The overall scheme combines a noise estimation network and a subspace filtering framework. It leverages the network's accurate noise modeling capabilities to overcome the limitations of traditional subspace methods that rely on prior noise assumptions, while achieving low-complexity signal separation through subspace eigenvalue decomposition, avoiding the high computational cost of pure deep learning models. Ultimately, it achieves a synergistic improvement in speech denoising performance and algorithm efficiency in complex environments.

[0126] Example 2:

[0127] Please refer to Figure 2 Based on the same inventive concept, the present invention discloses a speech denoising system based on subspace filtering, comprising: a data acquisition module M1, a feature extraction module M2, a model prediction module M3, a subspace construction module M4, a denoising module M5, and a result output module M6.

[0128] The data acquisition module M1 is used to acquire noisy speech and perform frame-by-frame processing to obtain each frame of data.

[0129] The feature extraction module M2 is used to calculate the autocorrelation coefficient and Mel-frequency cepstral coefficient of each frame data, and to concatenate them to construct the feature vector corresponding to each frame data.

[0130] The model prediction module M3 is used to input each of the feature vectors into a preset noise estimation network to perform noise autocorrelation coefficient prediction processing, and obtain the noise autocorrelation coefficient corresponding to the noise in each frame of data.

[0131] Furthermore, the noise estimation network includes an input linear layer, a multi-scale convolutional layer, an LSTM network layer, and an output linear layer;

[0132] The model prediction module M3 includes: a linear input unit, a convolution unit, a temporal modeling unit, and a linear output unit;

[0133] The linear input unit is used to align the feature vector in dimensions according to the input linear layer and perform nonlinear transformation processing through the tanh activation function to obtain the first feature.

[0134] The convolutional unit is used to perform a convolution operation on the first feature according to the multi-scale convolutional layer, extract the noise hidden feature, and encode it into a high-dimensional latent space to obtain the second feature;

[0135] The temporal modeling unit is used to perform temporal feature modeling on the second feature based on the LSTM network layer and output noisy temporal features;

[0136] The linear output unit is used to map the noise temporal characteristics according to the output linear layer to obtain the noise autocorrelation coefficient.

[0137] In this preferred embodiment, the noise estimation network performs dimensional alignment on the feature vector through an input linear layer and performs nonlinear transformation using the tanh activation function to normalize and nonlinearly map the original features, thereby enhancing the feature representation capability. A multi-scale convolutional layer performs convolution operations on the first feature, extracting multi-frequency hidden features of the noise through convolutional kernels of different scales and encoding them into a high-dimensional latent space, improving the ability to capture complex noise patterns. An LSTM network layer models the second feature based on temporal characteristics, using a recurrent structure to learn the dynamic changes of noise over time, outputting more robust temporal noise features. The output linear layer maps the temporal noise features to obtain the noise autocorrelation coefficient, accurately transforming the high-dimensional features into key parameters required for subspace filtering, further improving the accuracy and specificity of noise estimation.

[0138] The subspace construction module M4 is used to construct a subspace separation matrix and perform eigenvalue decomposition processing based on the autocorrelation coefficients, noise autocorrelation coefficients and identity matrix to obtain eigenvalue data and eigenma.

[0139] Furthermore, the subspace construction module M4 includes: a first matrix construction unit, a second matrix construction unit, a split matrix construction unit, and a decomposition unit;

[0140] The first matrix construction unit is used to construct a noisy speech autocorrelation matrix based on each of the autocorrelation coefficients.

[0141] The second matrix construction unit is used to construct a noise autocorrelation matrix based on each of the noise autocorrelation coefficients;

[0142] The separation matrix construction unit is used to construct a subspace separation matrix based on the noisy speech autocorrelation matrix, the noise autocorrelation matrix, and the identity matrix, as shown below:

[0143] Where ∑ represents the subspace separation matrix. R represents the noise autocorrelation matrix n The transpose matrix, R x =R y -R n R represents the autocorrelation matrix of the speech signal. y Let I represent the noisy speech autocorrelation matrix;

[0144] The decomposition unit is used to perform eigenvalue decomposition on the subspace separation matrix to obtain eigenvalue data and a feature matrix; wherein, in the eigenvalue data, the eigenvalues ​​representing the direction of the speech signal are greater than zero, and the eigenvalues ​​representing the direction of the noise signal are less than zero, and the eigenvalue data and the feature matrix satisfy the following constraints:

[0145] Where Λ represents the eigenvalue data, V T This represents the transpose of the characteristic matrix V.

[0146] In this preferred embodiment, a noisy speech autocorrelation matrix and a noise autocorrelation matrix are constructed based on the autocorrelation coefficient to quantify the temporal statistical characteristics of the mixed signal and noise, providing a data foundation for subspace separation. A subspace separation matrix is ​​constructed by multiplying the inverse of the noise autocorrelation matrix by the noisy speech autocorrelation matrix and then subtracting the identity matrix. This matrix construction method preprocesses the noisy speech signal by "whitening" it through the inverse operation of the noise matrix, transforming the energy distribution difference between the speech signal and noise into the distinguishability of matrix features. The constructed subspace separation matrix is ​​then subjected to eigenvalue decomposition, and two sets of constraints are introduced: the first set of constraints multiplies the speech autocorrelation matrix by the transpose of the eigenvalue matrix and then by the eigenvalue matrix again, resulting in a diagonalized eigenvalue matrix that concentrates speech energy on the eigenvalues ​​along the main diagonal; the second set of constraints multiplies the noise autocorrelation matrix by the transpose of the eigenvalue matrix and then by the eigenvalue matrix again, resulting in an identity matrix, thus achieving decorrelation processing of the noise signal and eliminating anisotropic noise interference. Furthermore, since the subspace separation matrix is ​​essentially the product of the inverse of the noise autocorrelation matrix and the speech autocorrelation matrix, and the inverse of the noise autocorrelation matrix has positive definite properties, when there is a speech signal in the subspace direction, the corresponding eigenvalue is positive; when there is only noise in the subspace direction (speech energy is zero), the eigenvalue naturally becomes negative or zero due to matrix operations (in reality, it is mostly negative due to noise estimation errors). Therefore, compared with existing technologies that rely on empirical thresholds or eigenvalue sorting to distinguish speech and noise directions (such as traditional subspace methods that require manually setting a "significant eigenvalue" screening threshold), this method, through matrix inversion operations and orthogonalization constraints, makes the eigenvalue sign and signal type strictly correspond, automatically separating speech and noise directions without manual intervention, reducing reliance on prior knowledge and improving the stability and accuracy of separation in complex noise environments.

[0147] The noise reduction module M5 is used to transform each frame data into a subspace according to the feature matrix to obtain subspace signals in each direction, and calculate the subspace gain matrix in combination with the feature value data to filter each subspace signal to obtain the corresponding noise reduction signal.

[0148] Furthermore, the noise reduction module M5 includes: a secondary frame division unit and a transformation unit;

[0149] The secondary framing unit is used to perform overlapping secondary framing processing on each of the frame data, using the order of the autocorrelation coefficient of each frame data as the framing window size, to obtain each subframe data.

[0150] The transformation unit is used to transform the subframe data to a subspace according to the feature matrix to obtain subspace signals corresponding to each direction.

[0151] In this preferred embodiment, the order of the autocorrelation coefficient of the framed data is used as the framed window size. The framed data is subjected to overlapping secondary framed processing to obtain subframe data. By matching the framed granularity with the statistical characteristic dimension of the autocorrelation coefficient, the temporal resolution of the subframe data is ensured to be compatible with the matrix dimension of the subsequent subspace transformation. At the same time, the overlapping processing reduces information loss caused by the framed boundary effect. Furthermore, the subframe data is transformed to the subspace through the feature matrix to obtain the subspace signal corresponding to each direction, realizing the directional separation of noisy speech in high-dimensional space, and providing a structured data foundation for subsequent filtering processing for different signal directions.

[0152] Furthermore, the noise reduction module M5 also includes: an eigenvalue adjustment unit, a gain calculation unit, and a gain matrix construction unit;

[0153] The feature value adjustment unit is used to filter out the direction of each noise signal and adjust the corresponding feature value to zero according to the positive or negative sign of each feature value in the feature value data.

[0154] The gain calculation unit is used to calculate the subspace gain in each signal direction based on the adjusted feature value data, combined with a preset gain calculation function and the Lagrange multipliers in each signal direction; wherein each Lagrange multiplier is determined by the signal-to-noise ratio in the corresponding signal direction.

[0155] The gain matrix construction unit is used to construct the subspace gain matrix based on the gains of each subspace, as shown below:

[0156] Among them, g kk This represents the subspace gain in the k-th signal direction of the subspace gain matrix. μ(v) represents the ratio of speech energy to noise energy in the k-th signal direction of the eigenvalue data. k ) represents the Lagrange multiplier corresponding to the k-th signal direction in the subspace, m represents the number of speech signal directions in the subspace, and n represents the total number of signal directions in the subspace.

[0157] In this preferred embodiment, the direction of the noise signal is filtered according to the positive or negative sign of each feature value in the feature value data, and the corresponding feature value is adjusted to zero. The direction of speech and noise is directly distinguished by mathematical properties, avoiding misjudgment caused by the reliance on empirical thresholds in traditional methods. At the same time, based on the adjusted feature value data, the subspace gain of each direction is calculated by combining a preset gain calculation function and a Lagrange multiplier determined by the signal-to-noise ratio, so that the gain allocation is dynamically related to the signal-to-noise ratio of the signal direction. The subspace gain of the k-th signal direction is calculated by dividing the ratio of speech energy to noise energy in that direction by the sum of this ratio and the corresponding Lagrange multiplier, so as to achieve a balance between noise suppression and speech preservation. Finally, a subspace gain matrix is ​​constructed, which can be used to apply targeted filtering to signals in different directions through diagonal elements to ensure that the noise direction signal is completely suppressed and the speech direction signal is adaptively enhanced according to the signal-to-noise ratio.

[0158] Furthermore, the gain calculation unit includes: a signal-to-noise ratio calculation subunit and a matching subunit;

[0159] The signal-to-noise ratio (SNR) calculation subunit is used to calculate the SNR of each signal direction within the subspace based on the feature value data; the calculation method is as follows:

[0160] Among them, SNR(v k ) represents the direction v of the k-th signal. k Signal-to-noise ratio;

[0161] The matching subunit is used to determine the Lagrange multiplier corresponding to each signal direction based on the signal-to-noise ratio and according to a preset matching rule; wherein the matching rule is as follows:

[0162]

[0163] Where, μ(v) k ) represents the Lagrange multiplier corresponding to the k-th signal direction in the subspace, max μ and min μ Let up represent the preset maximum and minimum Lagrange multipliers, respectively. snr and low snr These represent the preset maximum signal-to-noise ratio and minimum signal-to-noise ratio, respectively.

[0164] In this preferred embodiment, the signal-to-noise ratio (SNR) of each signal direction within the subspace is calculated based on eigenvalue data. The signal quality in each direction is quantified by the ratio of speech energy to noise energy in that direction. Furthermore, the Lagrange multiplier is determined based on a preset matching rule. When the SNR is between the minimum and maximum SNR, a smooth transition is achieved through a linear function to ensure the continuity of gain adjustment. When the SNR is higher than the maximum SNR, the minimum Lagrange multiplier is used to preserve the high SNR speech signal, and when the SNR is lower than the minimum SNR, the maximum Lagrange multiplier is used to enhance noise suppression. By dynamically adapting to different SNR scenarios through a piecewise function, the over-suppression or under-suppression problem caused by a single fixed parameter is avoided.

[0165] The result output module M6 is used to perform inverse transformation on each of the noise reduction signals according to the feature matrix to obtain each noise reduction speech data frame, and perform frame merging processing to output the noise reduction speech corresponding to the noisy speech.

[0166] In summary, compared with existing technologies, the embodiments of this application have the following beneficial effects: By acquiring noisy speech and performing frame-by-frame processing, continuous speech signals are converted into independently analyzable frame-by-frame data, providing a basic processing unit for subsequent feature extraction and noise estimation; the autocorrelation coefficient and Mel-frequency cepstral coefficient of each frame-by-frame data are calculated and concatenated to construct a feature vector. The autocorrelation coefficient reflects the temporal correlation of the signal, and the Mel-frequency cepstral coefficient captures the spectral features that the human ear is sensitive to. The combination of the two provides multi-dimensional acoustic feature input for noise estimation; the feature vector is input into a preset noise estimation network to predict the noise autocorrelation coefficient. The network model directly learns the statistical laws of noise, avoiding... Traditional methods rely on estimation biases caused by manual assumptions. This approach constructs a subspace separation matrix based on the autocorrelation coefficient, noise autocorrelation coefficient, and identity matrix, and performs eigenvalue decomposition. Through mathematical transformation, noisy speech is mapped to a high-dimensional space, achieving separation of speech and noise along eigenvector directions. The framed data is then transformed to the subspace based on the eigenvalue matrix, and a gain matrix is ​​calculated for filtering. Targeted gains are applied to signals in different directions to achieve a dynamic balance between noise suppression and speech preservation. Finally, the denoised signal undergoes an inverse transformation based on the eigenvalue matrix and is combined into frames, restoring the subspace-processed signal to time-domain speech, ensuring the continuity and integrity of the output signal. The overall solution combines a noise estimation network and a subspace filtering framework. This overcomes the limitations of traditional subspace methods that rely on prior noise assumptions by leveraging the network's accurate noise modeling capabilities. Furthermore, it achieves low-complexity signal separation through subspace eigenvalue decomposition, avoiding the high computational cost of pure deep learning models. Ultimately, this results in a synergistic improvement in speech denoising performance and algorithm efficiency in complex environments.

[0167] Example 3:

[0168] Figure 3 A structural diagram of a speech noise reduction device based on subspace filtering according to this application is presented. For example... Figure 3 As shown, the speech noise reduction device based on subspace filtering may include: a processor N1, a memory N2, a data interface N3, and a communication bus N4.

[0169] Wherein: processor N1, memory N2, and data interface N3 communicate with each other through communication bus N4; data interface N3 is used for data communication with other devices such as input devices or output devices; processor N1 is used to execute program N5, which can specifically execute the relevant steps in any of the above embodiments of the speech denoising method based on subspace filtering.

[0170] Specifically, program N5 may include program code, which includes computer-executable instructions.

[0171] The processor N1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The subspace-based filtering-based speech denoising device includes one or more processors, which may be processors of the same type, such as one or more CPUs, or processors of different types, such as one or more CPUs and one or more ASICs.

[0172] Memory N2 is used to store program N5. Memory N2 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage.

[0173] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Furthermore, the embodiments in this application are not directed to any particular programming language.

[0174] Example 4:

[0175] This invention also provides a computer-readable storage medium storing at least one executable instruction that, when executed on a subspace-based speech denoising device / system, causes the subspace-based speech denoising device / system to perform one of the subspace-based speech denoising methods described in any of the above method embodiments.

[0176] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. Similarly, for the purpose of simplification and aiding understanding of one or more aspects of the invention, in the above description of exemplary embodiments of this application, various features of the embodiments are sometimes grouped together in a single embodiment, figure, or description thereof. The claims, which follow the detailed description, are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of this application.

[0177] Those skilled in the art will understand that the modules in the device of the embodiment can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiment can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components, except that at least some of such features and / or processes or units are mutually exclusive.

Claims

1. A speech denoising method based on subspace filtering, characterized in that, include: Noisy speech is acquired and processed into frames to obtain the data for each frame. Calculate the autocorrelation coefficient and Mel-frequency cepstral coefficient of each frame data, and concatenate them to construct the feature vector corresponding to each frame data; Each of the aforementioned feature vectors is input into a preset noise estimation network to perform noise autocorrelation coefficient prediction processing, thereby obtaining the noise autocorrelation coefficient corresponding to the noise in each frame of data. Based on the autocorrelation coefficients, noise autocorrelation coefficients and identity matrices, a subspace separation matrix is ​​constructed and eigenvalue decomposition is performed to obtain eigenvalue data and eigenma matrices. Based on the feature matrix, each frame data is transformed to a subspace to obtain subspace signals in each direction. The subspace gain matrix is ​​calculated by combining the feature value data, and the subspace signals are filtered to obtain the corresponding noise-reduced signals. Based on the feature matrix, each of the noise-reduced signals is inversely transformed to obtain each noise-reduced speech data frame, and then frame merging is performed to output the noise-reduced speech corresponding to the noisy speech. The noise estimation network includes an input linear layer, a multi-scale convolutional layer, an LSTM network layer, and an output linear layer. The step of inputting each feature vector into the preset noise estimation network and performing noise autocorrelation coefficient prediction processing to obtain the noise autocorrelation coefficient corresponding to the noise in each frame of data includes: Based on the input linear layer, the feature vector is dimension-aligned and then subjected to nonlinear transformation using the tanh activation function to obtain the first feature; Based on the multi-scale convolutional layer, the first feature is convolved to extract the noise hidden features, which are then encoded into a high-dimensional latent space to obtain the second feature. Based on the LSTM network layer, the second feature is modeled as a temporal feature, and the noisy temporal feature is output. Based on the output linear layer, the noise temporal characteristics are mapped to obtain the noise autocorrelation coefficient.

2. The speech denoising method based on subspace filtering as described in claim 1, characterized in that, The process of constructing a subspace separation matrix and performing eigenvalue decomposition based on the autocorrelation coefficients, noise autocorrelation coefficients, and identity matrix to obtain eigenvalue data and a feature matrix includes: Construct a noisy speech autocorrelation matrix based on the autocorrelation coefficients described above; Based on the noise autocorrelation coefficients described above, construct a noise autocorrelation matrix; Based on the noisy speech autocorrelation matrix, the noise autocorrelation matrix, and the identity matrix, a subspace separation matrix is ​​constructed, as follows: ;in, Represents the subspace separation matrix. Represents the noise autocorrelation matrix The transpose of the matrix, The autocorrelation matrix represents the speech signal. This represents the autocorrelation matrix of noisy speech. Represents the identity matrix; The subspace separation matrix is ​​subjected to eigenvalue decomposition to obtain eigenvalue data and a feature matrix; wherein, in the eigenvalue data, the eigenvalues ​​representing the direction of the speech signal are greater than zero, and the eigenvalues ​​representing the direction of the noise signal are less than zero, and the eigenvalue data and the feature matrix satisfy the following constraints: ;in, Represents eigenvalue data. Representation of the characteristic matrix The transpose of .

3. The speech denoising method based on subspace filtering as described in claim 1, characterized in that, The step of transforming each frame data into a subspace based on the feature matrix to obtain subspace signals in each direction includes: Using the order of the autocorrelation coefficient of each frame data as the frame window size, the frame data is subjected to overlapping secondary frame division processing to obtain each subframe data. Based on the feature matrix, the subframe data is transformed into a subspace to obtain subspace signals corresponding to each direction.

4. The speech denoising method based on subspace filtering as described in claim 2, characterized in that, The calculation of the subspace gain matrix by combining the eigenvalue data includes: Based on the sign of each feature value in the feature value data, the direction of each noise signal is selected and the corresponding feature value is adjusted to zero; Based on the adjusted eigenvalue data, combined with the preset gain calculation function and the Lagrange multipliers in each signal direction, the subspace gain in each signal direction is calculated; wherein, each of the Lagrange multipliers is determined by the signal-to-noise ratio in the corresponding signal direction; Based on the gains of each subspace, the subspace gain matrix is ​​constructed as follows: ;in, This represents the subspace gain in the k-th signal direction of the subspace gain matrix. This represents the ratio of speech energy to noise energy in the k-th signal direction of the eigenvalue data. Let represent the Lagrange multiplier corresponding to the k-th signal direction in the subspace, m represent the number of speech signal directions in the subspace, and n represent the total number of signal directions in the subspace.

5. The speech denoising method based on subspace filtering as described in claim 4, characterized in that, Each of the Lagrange multipliers is determined by the signal-to-noise ratio in the corresponding signal direction, including: Based on the eigenvalue data, the signal-to-noise ratio (SNR) for each signal direction within the subspace is calculated; the calculation method is as follows: ;in, Indicates the direction of the k-th signal Signal-to-noise ratio; Based on the aforementioned signal-to-noise ratios, the Lagrange multipliers corresponding to each signal direction are determined according to a preset matching rule; wherein the matching rule is as follows: ； in, This represents the Lagrange multiplier corresponding to the k-th signal direction in the subspace. and These represent the preset maximum and minimum Lagrange multipliers, respectively. and These represent the preset maximum signal-to-noise ratio and minimum signal-to-noise ratio, respectively.

6. A speech denoising system based on subspace filtering, characterized in that, include: The module includes a data acquisition module, a feature extraction module, a model prediction module, a subspace construction module, a noise reduction module, and a result output module. The data acquisition module is used to acquire noisy speech and perform frame-by-frame processing to obtain each frame of data. The feature extraction module is used to calculate the autocorrelation coefficient and Mel-frequency cepstral coefficient of each frame data, and to concatenate them to construct the feature vector corresponding to each frame data. The model prediction module is used to input each feature vector into a preset noise estimation network to perform noise autocorrelation coefficient prediction processing, and obtain the noise autocorrelation coefficient corresponding to the noise in each frame of data. The subspace construction module is used to construct a subspace separation matrix and perform eigenvalue decomposition processing based on the autocorrelation coefficients, noise autocorrelation coefficients and identity matrix to obtain eigenvalue data and eigenmatrix. The noise reduction module is used to transform each frame data into a subspace according to the feature matrix to obtain subspace signals in each direction, and calculate the subspace gain matrix in combination with the feature value data to filter each subspace signal to obtain the corresponding noise reduction signal. The result output module is used to perform inverse transformation on each of the noise reduction signals according to the feature matrix to obtain each noise reduction speech data frame, and perform frame merging processing to output the noise reduction speech corresponding to the noisy speech. The noise estimation network includes an input linear layer, a multi-scale convolutional layer, an LSTM network layer, and an output linear layer. The model prediction module includes: a linear input unit, a convolutional unit, a temporal modeling unit, and a linear output unit; The linear input unit is used to align the feature vector in dimensions according to the input linear layer and perform nonlinear transformation processing through the tanh activation function to obtain the first feature. The convolutional unit is used to perform a convolution operation on the first feature according to the multi-scale convolutional layer, extract the noise hidden feature, and encode it into a high-dimensional latent space to obtain the second feature; The temporal modeling unit is used to perform temporal feature modeling on the second feature based on the LSTM network layer and output noisy temporal features; The linear output unit is used to map the noise temporal characteristics according to the output linear layer to obtain the noise autocorrelation coefficient.

7. A speech noise reduction device based on subspace filtering, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the steps of a speech denoising method based on subspace filtering according to any one of claims 1-5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of a speech denoising method based on subspace filtering according to any one of claims 1-5.

Images