Voice noise reduction method and system for data center scene, terminal and storage medium

Abstract

The invention relates to the technical field of voice noise reduction, in particular to a data center scene-oriented voice noise reduction method and system, a terminal and a storage medium, and the method comprises the steps: carrying out the STFT time-frequency transformation of a to-be-denoised voice, and obtaining a complex frequency spectrum of the to-be-denoised voice; adopting a multi-scale CNN convolutional neural network to perform feature extraction on the complex frequency spectrum; performing time sequence modeling on the extracted features by using a bidirectional long short-term memory (LSTM) network; voice features in the features after time sequence modeling are separated; carrying out reverberation suppression on the separated voice features; and inverse STFT reconstruction is carried out on the voice features after reverberation suppression, and the voice after noise reduction is obtained. The method is used for solving the problem of time-frequency-space three-dimensional coupling of the noise of the data center.

Description

Technical Field

[0001] The present invention relates to the technical field of speech noise reduction, and in particular to a speech noise reduction method, system, terminal and storage medium for data center scenarios. Background Art

[0002] Data centers have various noise sources, such as server fans, which form time-frequency domain coupled interference. The average speech-to-noise ratio (SNR) is lower than -5dB, seriously affecting the clarity of voice signals.

[0003] In addition, sudden changes in device load can also cause drastic changes in the noise spectrum. For example, when a GPU cluster is started, low-frequency noise increases by 12dB / oct, which is difficult to cope with using traditional noise reduction methods based on the assumption of stationarity.

[0004] In addition, reflections from the cabinet's metal surface cause the reverberation time (RT60) to reach 0.8-1.2 seconds, causing the voice signal to trail and overlap with noise, reducing voice intelligibility.

[0005] Traditional signal processing: spectral subtraction (e.g., Boll, 1979) produces musical noise for non-stationary noise, and Wiener filtering requires prior noise statistics information.

[0006] Therefore, a speech noise reduction method is needed that can effectively solve the problem of time-frequency-space three-dimensional coupling of data center noise. Summary of the Invention

[0007] In view of the above shortcomings, the present invention provides a voice noise reduction method, system, terminal and storage medium for data center scenarios to solve at least one of the above shortcomings.

[0008] In a first aspect, the present invention provides a method for speech noise reduction in a data center scenario, the method comprising: S1. Perform STFT (Short-Time Fourier Transform) time-frequency transformation on the speech to be de-noised to obtain the complex spectrum of the speech to be de-noised; S2. Using a multi-scale CNN convolutional neural network to extract features from the complex spectrum; S3, using a bidirectional long short-term memory network LSTM to perform time series modeling on the features extracted in step S2; S4, separating the speech features from the features after temporal modeling; S5. performing reverberation suppression on the separated speech features; S6. Perform inverse STFT reconstruction on the speech features after reverberation suppression to obtain the denoised speech.

[0009] In an optional embodiment, in S4, a residual additive separation module is used to separate the speech features from the features after time series modeling; In S5, a reverberation suppression unit is used to suppress reverberation of the separated speech features.

[0010] In an optional embodiment, the network architecture of the multi-scale CNN convolutional neural network includes: Input layer, used for input of the entire multi-scale CNN convolutional neural network; The first convolutional layer uses a 5×1 convolution; the input end of the 5×1 convolution is connected to the input layer, and the output end of the 5×1 convolution is connected to a ReLU activation function; A MaxPool pooling layer, connected to the output of the ReLU activation function, is used to downsample the output of the first convolutional layer; The second convolutional layer uses a 3×3 convolution; the input end of the 3×3 convolution is connected to the MaxPool pooling layer, and the output end is connected to a ReLU activation function; The global average pooling layer GlobalAveragePool is connected to the ReLU activation function connected to the output of the second convolutional layer to compress the feature map along the spatiotemporal dimension and generate a channel description vector; The Flatten layer is connected to the output of the global average pooling layer and is used to convert the description vector into a format that can be processed by the fully connected layer; The first fully connected layer Gemm has its input connected to the output of the Flatten layer, and its output is connected to a ReLU activation function. The frequency domain weight matrix B<16×64> is introduced in the first fully connected layer Gemm to achieve dimensionality reduction from 64 dimensions to 16 dimensions. The second fully connected layer Gemm has its input connected to the ReLU activation function connected to the output of the first fully connected layer Gemm, and its output is connected to a Sigmoid function. The frequency domain weight matrix B<64×16> is introduced into the first fully connected layer Gemm to achieve 16-dimensional to 64-dimensional reconstruction. The Sigmoid function is used to output the channel attention weight. The Reshape layer is connected to the output of the Sigmoid function to change the shape of the tensor; The Shape layer is connected to the ReLU activation function connected to the output of the second convolutional layer to obtain tensor shape information; The Expand layer is connected to the output of the Reshape layer and the Shape layer to copy the tensor; The feature weighting layer Multi is connected to the output of the ReLU activation function connected to the output of the second convolutional layer and the output of the Expand layer. It is used to weight the input feature map by channel and output the weighted feature map; The feature compression layer uses 1×1 convolution and is connected to the output of the feature weighting layer to map multi-scale features to a unified dimension; The output layer is a Reshape layer that uses 1×1 convolution. Its input is connected to the output of the feature compression layer to adjust the dimension of the output feature to 1×64×256.

[0011] In an optional embodiment, the input end of the bidirectional long short-term memory network LSTM is connected to the output end of the multi-scale CNN convolutional neural network; The bidirectional long short-term memory network LSTM includes: Forget Gate: , in, is a 128-dimensional bias vector, is the weight matrix 128*(128+256), is the Sigmoid activation function, represents the hidden state at the previous moment, is the input feature at the current moment; Input Gate: , in, is the input gate weight, is the weight matrix 128*(128+256), is a 128-dimensional bias vector; Candidate memory: , in, , is the weight matrix (256 is the input feature dimension); is a 128-dimensional bias vector, and tanh is a hyperbolic tangent activation function; Memory Update: , in, represents the Hadamard product (element-wise multiplication), Represents the cellular memory of the previous moment; Output gate: , in, is a 128-dimensional bias vector, is the weight matrix 128*(128+256); Hidden state: ; in, , is the hidden state at time step t; , is the cell state at time step t; Forward and backward LSTM output concatenation: , Multi-head attention mechanism: , in, ; =256 / 4=64, representing the dimension of each attention head; Output size: (T, 256).

[0012] In an optional embodiment, the output end of the bidirectional long short-term memory network LSTM is connected to the input end of the residual additive separation module; The residual additive separation module is improved based on the residual learning framework, including: Improvement 1: Use dual-branch heterogeneity, with the dual branches being the speech path and the noise path. The speech path outputs the time-frequency mask through the fully connected layer and Sigmoid , used to process speech signals; the noise path uses 5-point convolution to model the steady-state noise fundamental frequency and 3-point convolution to locate burst pulses, which are used to process noise signals; Improvement 2: Physically guided residual fusion, the fusion formula is proposed: , formula, where For voice protection items, is the noise residual term, ⊙ represents element-by-element multiplication; α is the residual weight coefficient; Improvement 3: Add pulse suppression mechanism, when abrupt energy change is detected, force the fusion formula to <0.1, and the damaged speech is reconstructed by combining the noise residual term.

[0013] In an optional embodiment, the reverberation suppression unit includes a 12th-order LPC inverse filter and a deep learning speech enhancement module; the input end of the 12th-order LPC inverse filter is used to input the separated speech features, and the output end of the 12th-order LPC inverse filter is connected to the input end of the deep learning speech enhancement module; The transfer function of the 12th-order LPC inverse filter is:

[0014] in, represents the linear prediction coefficient, which is obtained by minimizing the following objective function; z represents the Z-transform variable; is the filter transfer function; Objective function: , (2) Where, T: total duration; t: current time; : historical moment; s(t): pure speech signal; :Pure speech signals of historical moments; The deep learning speech enhancement module includes a transposed convolution layer, a multi-scale dilated convolution layer, and a feature fusion layer connected in sequence; the input end of the transposed convolution layer is the input end of the 12th-order LPC inverse filter; The transposed convolution layer is used to upsample the separated speech features; The multi-scale dilated convolution layer is used to extract features of different scales from the speech features sampled by the transposed convolution layer; The feature fusion layer is used to splice the features of different scales extracted by the multi-scale dilated convolution layer in the channel dimension to obtain and output a higher-dimensional feature map.

[0015] In an optional embodiment, the input end of the 12th-order LPC inverse filter is connected to the output end of the residual additive separation module; The training methods for the multi-scale CNN convolutional neural network, bidirectional long short-term memory network LSTM, residual additive separation module, and reverberation suppression unit include: Build the network architectures of the multi-scale CNN convolutional neural network, bidirectional long short-term memory network LSTM, residual additive separation module, and reverberation suppression unit, and establish the connection relationship between each network architecture; First, the first training set is used to train the multi-scale CNN convolutional neural network-bidirectional long short-term memory network LSTM; Then, the second training set is used to perform generator-discriminator adversarial training on the reverberation suppression unit; Then, the residual additive separation module is trained using the third training set; Then, the fourth training set is used to fine-tune the multi-scale CNN convolutional neural network, the bidirectional long short-term memory network LSTM, the residual additive separation module and the reverberation suppression unit as a whole to obtain the trained multi-scale CNN convolutional neural network, the trained bidirectional long short-term memory network LSTM, the trained residual additive separation module and the trained reverberation suppression unit.

[0016] In a second aspect, the present invention provides a speech noise reduction system, the system comprising: The first transform unit is configured to perform STFT time-frequency transform on the speech to be de-noised to obtain a complex spectrum of the speech to be de-noised; A multi-scale CNN convolutional neural network, whose input end is connected to the first transformation unit, is used to extract features from the complex spectrum using the multi-scale CNN convolutional neural network; A bidirectional long short-term memory network (LSTM) whose input is connected to the output of the multi-scale CNN convolutional neural network and is used to perform temporal modeling on the features extracted by the multi-scale CNN convolutional neural network; The residual additive separation module, whose input is connected to the bidirectional long short-term memory network (LSTM), is used to separate the speech features from the features after time series modeling; A reverberation suppression unit, used to suppress reverberation of the separated speech features; The second transformation unit is used to perform inverse STFT reconstruction on the speech features after reverberation suppression to obtain the denoised speech.

[0017] In a third aspect, the present invention provides a terminal, comprising: processor, memory, wherein The memory is used to store computer programs, The processor is used to call and run the computer program from the memory, so that the terminal executes the methods described in the above aspects.

[0018] In a fourth aspect, a computer storage medium is provided, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the methods described in the above aspects.

[0019] It can be seen from the above technical solutions that the present invention has the following advantages: The multi-scale CNN convolutional neural network (CNN) proposed in this paper uses convolution kernels of varying scales (e.g., 5×1 convolution, 3×3 convolution, etc.) to extract features from the complex spectrum, capturing the characteristics of speech and noise at different frequency and time scales. This multi-scale feature extraction approach is more adaptable to variations in non-stationary noise. Compared to traditional methods such as spectral subtraction, which poorly handle non-stationary noise and are prone to generating musical noise, this method can effectively reduce the generation of musical noise and improve the clarity of speech signals.

[0020] In the case of sudden changes in equipment load, the LSTM can ensure the robustness of the algorithm, enabling the noise reduction method to track the time-varying characteristics of noise in real time and maintain a good noise reduction effect.

[0021] Deep learning technologies such as multi-scale CNN and bidirectional LSTM are employed. Multi-scale CNN can extract multi-scale features from the complex spectrum, better capturing the characteristic differences between speech and noise at different frequency and time scales. Bidirectional LSTM can perform time series modeling on the extracted features, fully accounting for the temporal correlation of the speech signal. This ensures that the features incorporate information about speech and noise across multiple dimensions, including time, frequency, and space. This facilitates the separation of speech and noise features, avoids the introduction of noise in simple spectral subtraction, and thus improves speech quality and intelligibility.

[0022] This method uses a deep learning model to automatically learn the characteristics of speech and noise, without requiring prior knowledge of noise statistics. Multi-scale CNNs and bidirectional LSTMs can learn the inherent patterns and characteristics of speech and noise from large amounts of data, effectively suppressing noise without relying on prior information.

[0023] The present invention defines a 12th-order LPC inverse filter, suppresses early reflections by minimizing the objective function prediction error, retains direct sound, and constructs a physical priori noise reduction framework, which helps to solve the time-frequency-space three-dimensional coupling problem of data center noise.

[0024] This method supports feature fusion at different scales, leveraging information from various scales to more comprehensively describe the characteristics of speech signals and reverberation. This integrated feature representation helps subsequent processing modules more accurately separate speech and reverberation, reducing the effects of reverberation on speech signals, such as smearing and overlap, and improving speech clarity. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for the description. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0026] Figure 1 is a schematic flow chart of a method according to an embodiment of the present invention.

[0027] Figure 2 2 is a schematic diagram of the network architecture of a multi-scale CNN convolutional neural network in one embodiment of the present invention.

[0028] Figure 3 FIG. 4 is a schematic block diagram of a system according to an embodiment of the present invention.

[0029] Figure 4 A schematic diagram of the structure of a terminal provided by an embodiment of the present invention.

[0030] Figure 5 It is a structural diagram of a computer storage medium provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The present invention will be described in detail below. Specific details such as specific system structures and techniques are provided for purposes of illustration and not limitation to provide a thorough understanding of the embodiments of the present invention. However, it should be apparent to those skilled in the art that the present invention may be practiced in other embodiments without these specific details.

[0032] It should be understood that when used in the present specification, the term "comprising" indicates the presence of the described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their collections. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0033] When describing the present invention, phrases such as "one embodiment" or "some embodiments" mean that the particular features, structures, or characteristics described in the embodiment are included in one or more embodiments of the present invention. Therefore, phrases such as "in some embodiments" or "in other embodiments" that appear in different places in the present invention do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized.

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0035] Figure 1 A flow chart of a method for voice noise reduction in a data center scenario provided by an embodiment of the present invention. Figure 1 The execution subject may be a speech noise reduction system. The speech noise reduction method for data center scenarios provided by the embodiment of the present invention is executed by a computer device, and accordingly, the speech noise reduction system runs in the computer device.

[0036] Please refer to Figure 1 The method includes the following steps S1 to S6.

[0037] Step S1, performing STFT time-frequency transform on the speech to be de-noised to obtain a complex spectrum of the speech to be de-noised; Step S2: using a multi-scale CNN convolutional neural network to extract features from the complex spectrum; Step S3: Use a bidirectional long short-term memory network (LSTM) to perform time series modeling on the features extracted in step S2; Step S4: separating the speech features from the features after time series modeling; Step S5: performing reverberation suppression on the separated speech features; Step S6: Perform inverse STFT reconstruction on the speech features after reverberation suppression to obtain the denoised speech.

[0038] When the method of this embodiment is used, an STFT time-frequency transform is first performed on the speech to be denoised to obtain a complex spectrum of the speech to be denoised. A multi-scale CNN convolutional neural network is then used to extract features from the complex spectrum. A bidirectional long short-term memory network (LSTM) is then used to perform time series modeling on the extracted features. The speech features in the features after time series modeling are then separated. Reverberation suppression is then performed on the separated speech features. Finally, inverse STFT reconstruction is performed on the speech features after reverberation suppression to obtain the denoised speech.

[0039] In one embodiment, in step S5, a reverberation suppression unit is used to suppress reverberation on the separated speech features.

[0040] In another embodiment, in step S4, a residual additive separation module is used to separate the speech features from the features after time series modeling. That is, the residual additive separation module is used to separate the speech features from the features after time series modeling.

[0041] Data center noise exhibits multi-scale characteristics in the time-frequency domain. First, there are large-scale characteristics, such as the broadband continuity of low-frequency noise (0-200Hz) from server fans. Second, there are fine-grained characteristics, such as the narrowband harmonic structure of power supply whistles (2-5Hz). To address this, this paper proposes a multi-scale convolutional neural network (CNN).

[0042] Exemplarily, the network architecture of the multi-scale CNN convolutional neural network includes: Input layer, used for input of the entire multi-scale CNN convolutional neural network; The first convolutional layer uses a 5×1 convolution; the input end of the 5×1 convolution is connected to the input layer, and the output end of the 5×1 convolution is connected to a ReLU activation function; A MaxPool pooling layer, connected to the output of the ReLU activation function, is used to downsample the output of the first convolutional layer; The second convolutional layer uses a 3×3 convolution; the input end of the 3×3 convolution is connected to the MaxPool pooling layer, and the output end is connected to a ReLU activation function; The global average pooling layer GlobalAveragePool is connected to the ReLU activation function connected to the output of the second convolutional layer to compress the feature map along the spatiotemporal dimension and generate a channel description vector; The Flatten layer is connected to the output of the global average pooling layer and is used to convert the description vector into a format that can be processed by the fully connected layer; The first fully connected layer Gemm has its input connected to the output of the Flatten layer, and its output is connected to a ReLU activation function. The frequency domain weight matrix B<16×64> is introduced in the first fully connected layer Gemm to achieve dimensionality reduction from 64 dimensions to 16 dimensions. The second fully connected layer Gemm has its input connected to the ReLU activation function connected to the output of the first fully connected layer Gemm, and its output is connected to a Sigmoid function. The frequency domain weight matrix B<64×16> is introduced into the first fully connected layer Gemm to achieve 16-dimensional to 64-dimensional reconstruction. The Sigmoid function is used to output the channel attention weight. The Reshape layer is connected to the output of the Sigmoid function to change the shape of the tensor; The Shape layer is connected to the ReLU activation function connected to the output of the second convolutional layer to obtain tensor shape information; The Expand layer is connected to the output of the Reshape layer and the Shape layer to copy the tensor; The feature weighting layer Multi is connected to the output of the ReLU activation function connected to the output of the second convolutional layer and the output of the Expand layer. It is used to weight the input feature map by channel and output the weighted feature map; The feature compression layer uses 1×1 convolution and is connected to the output of the feature weighting layer to map multi-scale features to a unified dimension; The output layer is a Reshape layer that uses 1×1 convolution. Its input is connected to the output of the feature compression layer to adjust the dimension of the output feature to 1×64×256.

[0043] Specifically, if Figure 2 As shown, the multi-scale CNN convolutional neural network includes: (1) Input layer Input data: The complex spectrum of the speech to be denoised after STFT transformation (the real part and the imaginary part are divided into two channels).

[0044] Dimensions: (number of time frames T, number of frequencies 129, number of channels 2).

[0045] (2) First convolutional layer The input data first enters a 5×1 convolutional layer (weights W<32×1×5×1>). Its wide convolution kernel sweeps along the frequency axis, specifically extracting broadband, continuous features from low-frequency noise (0-200Hz) such as server fan noise. The 5-point spanning frequency design precisely matches the 20ms periodic fluctuations of the 50Hz fundamental frequency. After the ReLU activation function enhances the nonlinear representation of the noise pattern, a 2×1 max pooling layer downsamples the frequency by a factor of 2, compressing the 129 frequency points to 64. This operation preserves the noise envelope while eliminating 60% of redundant computation. The resulting feature map has an output size of [T, 64, 32], whose low resolution perfectly matches the slowly varying nature of low-frequency noise.

[0046] Conv (5×1 convolution).

[0047] Type: 2D convolution (Conv2D).

[0048] parameter: Convolution kernel size: 5×1.

[0049] Number of filters: 32.

[0050] stride: 1×1.

[0051] Padding: same.

[0052] Activation function: ReLU.

[0053] Function: Capture the wide-band continuous characteristics of low-frequency noise.

[0054] ReLU: Introduces nonlinearity and activates noise patterns in feature maps (3) MaxPool pooling layer Type: MaxPooling2D parameter: Pooling window: 2 (frequency axis) × 1 (time axis) Step size: 2x1 Output size: (T, 64, 32) (4) Second convolutional layer This path uses a 3×3 convolution kernel (weight W < 64×32×3×3>). The small kernel structure focuses on local details of high-frequency harmonics (2-5kHz) such as power supply whine. Before ReLU activation, the dilation ratio (1,2) is designed to expand the horizontal receptive field to 5 frames (the original 3 frames + the dilation gap), effectively capturing the temporal correlation of burst noise such as GPU startup (such as a 12dB / octet low-frequency surge). The output maintains the original 129-frequency resolution to ensure that the fine structure of transient noise is not blurred, generating a [T, 129, 64] feature map for subsequent fusion.

[0055] Conv(3×3 convolution): Type: 2D convolution (Conv2D).

[0056] parameter: Convolution kernel size: 3x3 Number of filters: 64 Dilation rate: (1,2) to achieve cross-timeframe perception Activation function: ReLU ReLU: Strengthen the nonlinear expression of high-frequency noise features (5) Channel attention gating GlobalAveragePooling: Global average pooling compresses the feature map along the spatiotemporal dimensions to generate a channel description vector.

[0057] , in : The spatial feature mean of channel c (compressing global information).

[0058] Flatten: Convert the vector into a format that can be processed by the fully connected layer.

[0059] Gemm (fully connected layer 1): Weight: B<16×64> (64→16 dimension reduction).

[0060] Function: Learn the nonlinear relationship between channels and extract the correlation features of key frequency bands.

[0061] ReLU: Filters redundant information.

[0062] Gemm (fully connected layer 2): Weights: B<64×16> (16→64 dimensional reconstruction).

[0063] Function: Restore channel dimension and prepare to generate attention weights.

[0064] Sigmoid: Output channel attention weight .

[0065] Shape+Reshape+Expand: Adjust the weight shape to match the feature map dimension (from 64 → 64×1×1).

[0066] The fully connected layer generates weights:

[0067] in , : It is the trainable weight matrix of the fully connected layer. The former is reduced to 1 / 4 of the original number of channels, and the latter restores the original dimension. During this period, the nonlinear relationship is learned through ReLU activation.

[0068] σ is the Sigmoid function, which maps the output to the [0,1] interval and generates the channel attention weight , its physical meaning is to quantify the contribution of the channel to the noise characterization.

[0069] Multi(feature weighting): Final feature map Weighted by channel:

[0070] in: The weighted feature map enables the network to focus on the heavily polluted frequency bands.

[0071] (6) Feature compression layer Conv (1×1 convolution) Type: 1x1 convolution (Conv2D).

[0072] parameter: Number of filters: 128.

[0073] Activation function: linear.

[0074] Purpose: Map multi-scale features to a unified dimension.

[0075] Reshape: Adjust the dimensions to 1×64×256.

[0076] Physical meaning: 64: Number of time frames (after compression).

[0077] 256: Fusion feature dimension (for LSTM time series modeling).

[0078] It can be understood that the input end of the bidirectional long short-term memory network LSTM is connected to the output end of the multi-scale CNN convolutional neural network.

[0079] The multi-scale CNN convolutional neural network in this implementation can capture noise features of different scales through convolution kernels of different sizes, and output a feature map (shape [1, 64, 256]) that is more abstract but contains rich noise information, where 64 is the number of time frames (after compression) and 256 is the feature dimension.

[0080] This convolutional neural network (CNN) processes the multi-scale characteristics of data center noise using three heterogeneous branches: the first branch focuses on the low-frequency envelope of broadband steady-state noise, the second branch locates high-frequency transient harmonics, and the third branch dynamically enhances the overlapping frequency bands of speech and noise. The CNN input is the complex spectrum after STFT transformation (size 1×1×129×2, corresponding to 129 frequency points in a dual-channel). The network ultimately outputs a fused feature of 1×64×256, which is then fed into the LSTM time series model.

[0081] In another embodiment, a bidirectional long short-term memory network (LSTM) includes: Forget Gate: , in, is a 128-dimensional bias vector, is the weight matrix 128*(128+256), is the Sigmoid activation function, represents the hidden state at the previous moment, is the input feature at the current moment; Input Gate: , in, is the input gate weight, is the weight matrix 128*(128+256), is a 128-dimensional bias vector; Candidate memory: , in, , is the weight matrix (256 is the input feature dimension); is a 128-dimensional bias vector, and tanh is a hyperbolic tangent activation function; Memory Update: , in, represents the Hadamard product (element-wise multiplication), Represents the cellular memory of the previous moment; Output Gate: , in, is a 128-dimensional bias vector, is the weight matrix 128*(128+256); Hidden state: ; in, , is the hidden state at time step t; , is the cell state at time step t; Forward and backward LSTM output concatenation: , Multi-head attention mechanism: , in, ; =256 / 4=64, representing the dimension of each attention head; Output size: (T, 256).

[0082] The bidirectional long short-term memory network LSTM in this embodiment: Input: 256-dimensional feature vector from CNN ; Output: 128-dimensional hidden state and cell status .

[0083] The bidirectional LSTM layer of this example: Number of units: 128 units in each direction, for a total of 256-dimensional output.

[0084] parameter: Forget gate bias initialization: 1.0 (to alleviate gradient disappearance).

[0085] recurrent_dropout=0.2.

[0086] 2) Layer Normalization Type: LayerNormalization; Function: Stabilize the deep network training process.

[0087] 3) Multi-head self-attention mechanism The 256-dimensional features are split into 4 heads (64 dimensions each).

[0088] In this embodiment, the bidirectional state transfer LSTM is used: the forward LSTM is used to model the historical evolution trend of noise (such as the process of increasing fan speed); the reverse LSTM is used to predict future noise mutations (such as broadband noise caused by sudden hard disk read and write).

[0089] In another embodiment, the output of the bidirectional long short-term memory network LSTM is connected to the input of the residual additive separation module; The residual additive separation module makes the following improvements based on the residual learning framework: Improvement 1: Use dual-branch heterogeneity, with the dual branches being the speech path and the noise path. The speech path outputs the time-frequency mask through the fully connected layer and Sigmoid , used to process speech signals; the noise path uses 5-point convolution to model the steady-state noise fundamental frequency and 3-point convolution to locate burst pulses, which are used to process noise signals; Improvement 2: Physically guided residual fusion, the fusion formula is proposed: , in, For voice protection items, is the noise residual term, ⊙ represents element-by-element multiplication; α is the residual weight coefficient; Improvement 3: Add pulse suppression mechanism, when abrupt energy change is detected, force the fusion formula to <0.1, and the damaged speech is reconstructed by combining the noise residual term.

[0090] Specifically, the speech path consists of a fully connected layer, a GELU activation function, and a sigmoid output. It outputs a temporal mask M(t,f)∈[0,1]. Its physical meaning is: when M(t,f)=1, the time-frequency point is likely to be speech; when M(t,f)=0, the time-frequency point is likely to be noise. For pulse bursts: The temporal abruptness of pulse noise causes its spectral energy to be significantly higher than that of surrounding areas. The speech path uses the nonlinear characteristics of the sigmoid function to automatically suppress the mask value of energy outliers (e.g., gradients >40dB / ms) to 0.

[0091] The noise path structure is: Conv1D (5 points) + ELU + LayerNorm + Conv1D (3 points). In this path, 5-point convolution captures broadband steady-state noise. The 5-point convolution kernel covers the fundamental frequency period of fan noise (50 Hz = 20 ms). The noise spectrum template is established through statistical learning. 3-point convolution locates burst pulses (points with gradients > 40 dB / ms).

[0092] The calculation formula for the output pure speech spectrum is, that is, the fusion part of the speech path and the noise path in the network structure:

[0093] in: .

[0094] The learnability coefficient α=0.6.

[0095] Steady-state noise removal process: Under the noise path, steady-state noise is learned ; In the voice path, on the noise-dominated road section , the output voice is: .

[0096] The process of eliminating burst pulses: In the detection stage, the 3-point convolution layer of the noise path outputs the pulse peak value, and the gradient detection is: , the residual complements the speech destroyed by the pulse: .

[0097] Based on the above, efficient positioning and strong suppression of high-energy, sparse pulse noise are achieved (such as attenuating hard disk pulse energy by 28dB).

[0098] Acoustic reverberation can be divided into two components: early reflections and late reverberation. Early reflections typically refer to the signals that reach the listener's ears when a sound is first reflected. These signals have a small time difference from the direct sound (usually less than 50 milliseconds), making them difficult to distinguish auditorily and primarily affecting the spatial perception of the sound. Late reverberation refers to the complex sound field formed by multiple reflections. These signals have a large time difference from the direct sound (usually greater than 50 milliseconds), creating a sense of reverberation and primarily affecting sound clarity.

[0099] In a data center environment, sound will reflect off the metal surfaces of the cabinets, causing reverberation (just like speaking in an empty room). The reverberation suppression unit is specifically designed to eliminate this reverberation, making speech cleaner and more direct.

[0100] The core idea of ​​physics-guided deep learning is to embed physical constraints (i.e., the reflection patterns of sound waves in metal cabinets (e.g., reverberation time RT60 = 0.8-1.2s)) into the training and inference process of neural networks, making their learning process conform to physical laws, thereby improving the accuracy and generalization ability of the model.

[0101] In one embodiment, the reverberation suppression unit includes a defined 12th-order LPC inverse filter and a deep learning speech enhancement module. The input end of the 12th-order LPC inverse filter is used to input the separated speech features, and the output end of the 12th-order LPC inverse filter is connected to the input end of the deep learning speech enhancement module.

[0102] The deep learning speech enhancement module includes a transposed convolution layer, a multi-scale dilated convolution layer, and a feature fusion layer connected in sequence; the input end of the transposed convolution layer is the input end of the 12th-order LPC inverse filter.

[0103] The transfer function of the 12th-order LPC inverse filter is:

[0104] in, represents the linear prediction coefficient (LPC), which is obtained by minimizing the following objective function (minimizing the prediction error (no explicit constraints, implicit acoustic physics laws)), and z represents the Z-transform variable; It is the filter transfer function; its function is to suppress early reflections (<50ms) and retain direct sound.

[0105] The transfer function clarifies the input-output relationship of the filter in the Z domain by adjusting the linear prediction coefficient , the characteristics of the filter can be changed to achieve different signal processing effects.

[0106]

[0107] Where, T: total duration; t: current time; : historical moment; s(t): pure speech signal; :Pure voice signals of historical moments.

[0108] The objective function is to minimize the prediction error to obtain the optimal LPC coefficients. In this way, the LPC inverse filter can suppress early reflections, preserve the direct sound components, and improve the clarity of speech.

[0109] By minimizing Get the best , the constraints are implicit in the physical laws of data center reverberation (metal surface RT60 = 0.8-1.2s). This output is used as the input of the deep learning module, guiding the network to focus on late reverberation suppression.

[0110] In another embodiment, the deep learning speech enhancement module includes a transposed convolution layer, a multi-scale dilated convolution layer, and a feature fusion layer connected in sequence; the input end of the transposed convolution layer is the input end of the 12th-order LPC inverse filter; The transposed convolution layer is used to upsample the separated speech features; The multi-scale dilated convolution layer is used to extract features of different scales from the speech features sampled by the transposed convolution layer; The feature fusion layer is used to splice the features of different scales extracted by the multi-scale dilated convolution layer in the channel dimension to obtain a higher-dimensional feature map and output it.

[0111] The transposed convolution layer is a commonly used neural network layer in deep learning, which achieves upsampling through deconvolution operations. In acoustic reverberation suppression, the transposed convolution layer can be used to improve temporal resolution and restore speech details. The structure of the transposed convolution layer is: .

[0112] nn.ConvTranspose1d is a neural network layer for one-dimensional transposed convolution (Transposed 1DConvolution) in PyTorch.

[0113] Multi-scale dilated convolution is an important technique in deep learning. By adjusting the dilation ratio of the convolution kernel, it expands the receptive field and captures features at different scales. In the acoustic reverberation suppression of this application, multi-scale dilated convolution is used to capture reverberation characteristics at different scales and improve the reverberation suppression effect. The multi-scale dilated convolution layer in this embodiment includes three branches, as shown in Table 1.

[0114] Table 1: Multi-scale dilated convolutional layers

[0115] It can be seen that in this embodiment, the multi-scale dilated convolution layer includes three convolutions, the convolution kernels of the three convolutions are 3×1, 5×1, and 7×1, and their receptive fields are 3 points, 9 points, and 29 points respectively; The feature fusion layer structure is: , in, 、 、 Represents the three convolution branches in the multi-scale hole convolution layer, Concat means 、 、 The output of Conv1D1×1 is concatenated, and Conv1D1×1 represents a 1×1 convolution, which is used to map the concatenated features to the target feature space.

[0116] The input of the reverberation suppression unit comes from the output features of the residual additive separation module, and the pure speech features are obtained after the final noise reduction.

[0117] After the LSTM process, the present invention extracts a noise feature sequence containing temporal context information. However, the noise features at each time point incorporate preceding and following information, facilitating subsequent noise separation and removal. Residual additive separation separates the noise into speech and noise components. Speech is then subtracted from the original audio to obtain the speech. This step filters out most steady-state and burst noise. The remaining speech is then reverberated to obtain pure speech features. The speech features are then subjected to inverse STFT reweighting to obtain noise-free speech.

[0118] In another embodiment, the input end of the 12th order LPC inverse filter is connected to the output end of the residual additive separation module. When in use, the output end of the residual additive separation module outputs the separated speech features to the 12th order LPC inverse filter.

[0119] In another embodiment, the training method of the multi-scale CNN convolutional neural network, the bidirectional long short-term memory network LSTM, the residual additive separation module and the reverberation suppression unit includes: Build the network architectures of the multi-scale CNN convolutional neural network, bidirectional long short-term memory network LSTM, residual additive separation module, and reverberation suppression unit, and establish the connection relationship between each network architecture; First, the first training set is used to train the multi-scale CNN convolutional neural network-bidirectional long short-term memory network LSTM; Then, the second training set is used to perform generator-discriminator adversarial training on the reverberation suppression unit; Then, the residual additive separation module is trained using the third training set; Then, the fourth training set is used to fine-tune the multi-scale CNN convolutional neural network, the bidirectional long short-term memory network LSTM, the residual additive separation module and the reverberation suppression unit as a whole to obtain the trained multi-scale CNN convolutional neural network, the trained bidirectional long short-term memory network LSTM, the trained residual additive separation module and the trained reverberation suppression unit.

[0120] Exemplarily, the method for constructing the first training set includes: Collect pure voice, data center air conditioning noise, and fan noise; The collected noises are labeled with the working condition labels "30%", "60%" and "100%"; The labeled air conditioner noises and fan noises are collected to obtain the first data set, which is divided into the first training set (the ratio of air conditioner noise to fan noise is 7:3) and the first validation set.

[0121] In this embodiment, the clean speech comes from LibriSpeech and engineer instruction recordings, and is converted into a time-frequency spectrum through frame segmentation, windowing, and STFT.

[0122] Among them, LibriSpeech: is a large corpus containing about 1,000 hours of English speech, whose data comes from the audiobooks of the LibriVox project.

[0123] The construction method of the second training set includes: Obtain the above collected pure voice; Collect cabinet impulse reverberation responses; Convolving the acquired clean speech with the collected cabinet impulse reverberation response to generate reverberant speech; The clean speech and its corresponding reverberant speech are collected to obtain the second training set.

[0124] The construction method of the third training set includes: Obtain the above-collected pure voice, air conditioning noise, and fan noise; Generate noisy speech using the obtained clean speech and air conditioner noise and fan noise; The noisy speech and the pre-generated progressive pulse signal are collected to form a second data set, and the second data set is divided into a training set and a test set.

[0125] The progressive pulse signal includes 80dB single pulse, continuous pulse with interval >500ms and pulse group with frequency of 10 pulses / second.

[0126] The construction method of the fourth training set includes: Gather samples from the first training set, the second training set, and the third training set; Randomly discard 30% of the frequency bands of the collected samples and use the context to complete the missing frequency bands to obtain a signal that simulates signal loss; Sine waves with a fundamental frequency of 50 / 60 Hz and harmonics were superimposed on the signals obtained for each simulated signal loss, and the amplitudes were randomly set to 10%-30% of the corresponding noise energy. ±2ms delay jitter was then applied to the resulting signal waveform to simulate a sudden change in the acoustic path. The signals from each simulated acoustic path change were then combined to form the fourth training set. Table 2 shows the data collected to construct the training set.

[0127] Table 2: Data collection

[0128] The first training set was used to train a multi-scale CNN convolutional neural network coupled with a bidirectional long short-term memory (LSTM) network. The data from the first training set was fed into the multi-scale CNN convolutional neural network coupled with a bidirectional long short-term memory (LSTM) network for training. During training, the model learned to minimize the spectral MSE loss of fan and air conditioning noise. The Adam optimizer was optimized from 1e-3 to 1e-4. Training was terminated when the validation loss did not decrease for five consecutive rounds.

[0129] Using the second training set, the reverberation suppression unit is trained with a generator-discriminator adversarial approach, including The generator receives and processes the reverberant speech in the second training set to output denoised speech ; The discriminator generates denoised speech And real clean voice Make a judgment and output the probability of authenticity ; Calculate the generator loss: ; Back propagate and update generator parameters; Calculating the discriminator loss , back-propagate and update the discriminator parameters; The training ends when the discriminator cannot distinguish effectively (accuracy < 55%).

[0130] Using the third training set, the residual additive separation module is trained, including: The data in the third training set is input into the residual additive separation module for iterative training. In each iteration, pulse samples are randomly selected and injected into the noisy speech. The output of the forward calculation residual module is , then calculate the loss , and then back-propagation is used to update the parameters until the signal-to-noise ratio in the pulse area of ​​the test set is stably improved by more than 28dB and the training is terminated.

[0131] Then, using the fourth training set, the trained multi-scale CNN convolutional neural network, bidirectional long short-term memory network LSTM, residual additive separation module and reverberation suppression unit are fine-tuned as a whole, including: loading data batch by batch → applying the perturbation combination of the current batch → forward calculation output → end-to-end loss Update all weights in reverse order. Termination conditions are: 1) Word Error Rate (WER) ≤ 5%; 2) PESQ Speech Quality Score ≥ 3.5; 3) Latency ≤ 80ms in all noise scenarios.

[0132] After fine-tuning, the final trained multi-scale CNN convolutional neural network, the trained bidirectional long short-term memory network LSTM, the trained residual additive separation module and the trained reverberation suppression unit are obtained.

[0133] like Figure 3 As shown, the present invention provides a speech noise reduction system, the system comprising: The first transform unit 201 is configured to perform STFT time-frequency transform on the speech to be de-noised to obtain a complex spectrum of the speech to be de-noised; A multi-scale CNN convolutional neural network 202, whose input end is connected to the first transformation unit, is used to extract features from the complex spectrum using a multi-scale CNN convolutional neural network; A bidirectional long short-term memory network LSTM 203, whose input end is connected to the output end of the multi-scale CNN convolutional neural network, is used to perform time series modeling on features extracted by the multi-scale CNN convolutional neural network; The residual additive separation module 204 has an input end connected to a bidirectional long short-term memory network LSTM, and is used to separate the speech features from the features after time series modeling; a reverberation suppression unit 205 for suppressing reverberation on the separated speech features; The second transform unit 206 is configured to perform inverse STFT reconstruction on the speech features after reverberation suppression to obtain the denoised speech.

[0134] The embodiment of the speech noise reduction system of this embodiment belongs to the same inventive concept as the speech noise reduction method for data center scenarios in the above-mentioned embodiments. For details not fully described in the embodiment of the speech noise reduction system, please refer to the embodiment of the speech noise reduction method for data center scenarios in the previous text.

[0135] Figure 4 This is a structural diagram of a terminal 300 provided in an embodiment of the present invention. The terminal 300 can be used to execute the method provided in an embodiment of the present invention.

[0136] The terminal 300 may include a processor 310, a memory 320, and a communication unit 330. These components communicate via one or more buses. Those skilled in the art will appreciate that the server structure shown in the figure does not limit the present invention. The server structure may be a bus structure or a star structure, and may include more or fewer components than shown, or may combine certain components or arrange the components differently.

[0137] Memory 320 can be used to store execution instructions of processor 310. Memory 320 can be implemented by any type of volatile or non-volatile storage device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk. When the execution instructions in memory 320 are executed by processor 310, terminal 300 can perform some or all of the steps in the above-described method embodiments.

[0138] The processor 310 is the control center of the storage terminal. It uses various interfaces and lines to connect various parts of the entire electronic terminal. It runs or executes software programs and / or modules stored in the memory 320, and calls data stored in the memory to perform various functions of the electronic terminal and / or process data. The processor can be composed of an integrated circuit (IC), for example, it can be composed of a single packaged IC, or it can be composed of multiple packaged ICs with the same or different functions. For example, the processor 310 can only include a central processing unit (CPU). In an embodiment of the present invention, the CPU can be a single computing core or multiple computing cores.

[0139] The communication unit 330 is configured to establish a communication channel so that the storage terminal can communicate with other terminals, receive user data sent by other terminals, or send user data to other terminals.

[0140] The present invention also provides a computer storage medium 400, such as Figure 5 As shown, computer storage medium 400 may store a program 410. When executed, program 410 may include some or all of the steps of each embodiment provided herein. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0141] In this specification, the same or similar parts between the various embodiments can be referred to each other. In particular, for the terminal embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description in the method embodiment.

[0142] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0143] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0144] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0145] The core innovation of this invention lies in building a noise reduction framework that combines physical a priori with data-driven synergy. This framework addresses the time-frequency-space coupling challenge of data center noise through the following theoretical breakthroughs: Noise decoupling theory: Establishing an additive-convolutional mixed model of speech, steady-state noise, reverberation, and burst pulses to prove component separability; Time-varying system stability: The LSTM hidden state memory mechanism ensures the robustness of the algorithm in the event of sudden changes in device load.

[0146] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A voice noise reduction method for data center scenarios, characterized in that: Methods include: S1, perform STFT time-frequency transform on the speech to be de-noised to obtain the complex spectrum of the speech to be de-noised; S2. Using a multi-scale CNN convolutional neural network to extract features from the complex spectrum; S3, using a bidirectional long short-term memory network LSTM to perform time series modeling on the features extracted in step S2; S4, separating the speech features from the features after temporal modeling; S5. performing reverberation suppression on the separated speech features; S6. Perform inverse STFT reconstruction on the speech features after reverberation suppression to obtain the denoised speech.

2. The method for data center-based speech noise reduction according to claim 1, wherein: In S4, the residual additive separation module is used to separate the speech features from the features after time series modeling; In S5, a reverberation suppression unit is used to suppress reverberation of the separated speech features.

3. The method for data center-based speech noise reduction according to claim 2, wherein: The network architecture of the multi-scale CNN convolutional neural network includes: Input layer, used for input of the entire multi-scale CNN convolutional neural network; The first convolutional layer uses a 5×1 convolution; the input end of the 5×1 convolution is connected to the input layer, and the output end of the 5×1 convolution is connected to a ReLU activation function; A MaxPool pooling layer, connected to the output of the ReLU activation function, is used to downsample the output of the first convolutional layer; The second convolutional layer uses a 3×3 convolution; the input end of the 3×3 convolution is connected to the MaxPool pooling layer, and the output end is connected to a ReLU activation function; The global average pooling layer GlobalAveragePool is connected to the ReLU activation function connected to the output of the second convolutional layer to compress the feature map along the spatiotemporal dimension and generate a channel description vector; The Flatten layer is connected to the output of the global average pooling layer and is used to convert the description vector into a format that can be processed by the fully connected layer; The first fully connected layer Gemm has its input connected to the output of the Flatten layer, and its output is connected to a ReLU activation function. The frequency domain weight matrix B<16×64> is introduced in the first fully connected layer Gemm to achieve dimensionality reduction from 64 dimensions to 16 dimensions. The second fully connected layer Gemm has its input connected to the ReLU activation function connected to the output of the first fully connected layer Gemm, and its output is connected to a Sigmoid function. The frequency domain weight matrix B<64×16> is introduced into the first fully connected layer Gemm to achieve 16-dimensional to 64-dimensional reconstruction. The Sigmoid function is used to output the channel attention weight. The Reshape layer is connected to the output of the Sigmoid function to change the shape of the tensor; The Shape layer is connected to the ReLU activation function connected to the output of the second convolutional layer to obtain tensor shape information; The Expand layer is connected to the output of the Reshape layer and the Shape layer to copy the tensor; The feature weighting layer Multi is connected to the output of the ReLU activation function connected to the output of the second convolutional layer and the output of the Expand layer. It is used to weight the input feature map by channel and output the weighted feature map; The feature compression layer uses 1×1 convolution and is connected to the output of the feature weighting layer to map multi-scale features to a unified dimension; The output layer is a Reshape layer that uses 1×1 convolution. Its input is connected to the output of the feature compression layer to adjust the dimension of the output feature to 1×64×256.

4. The method for voice noise reduction in data center scenarios according to claim 3, characterized in that: The input of the bidirectional long short-term memory network LSTM is connected to the output of the multi-scale CNN convolutional neural network; The bidirectional long short-term memory network LSTM includes: Forget Gate: ， in, is a 128-dimensional bias vector, is the weight matrix 128*(128+256), is the Sigmoid activation function, represents the hidden state at the previous moment, is the input feature at the current moment; Input Gate: ， in, is the input gate weight, is the weight matrix 128*(128+256), is a 128-dimensional bias vector; Candidate memory: ， in, , is the weight matrix (256 is the input feature dimension); is a 128-dimensional bias vector, and tanh is a hyperbolic tangent activation function; Memory Update: ， in, represents the Hadamard product (element-wise multiplication), Represents the cellular memory of the previous moment; Output gate: ， in, is a 128-dimensional bias vector, is the weight matrix 128*(128+256); Hidden state: ； in, , is the hidden state at time step t; , is the cell state at time step t; Forward and backward LSTM output concatenation: ， Multi-head attention mechanism: ， in, ; =256 / 4=64, representing the dimension of each attention head; Output size: (T, 256).

5. The method for voice noise reduction in data center scenarios according to claim 4, characterized in that: The output of the bidirectional long short-term memory network LSTM is connected to the input of the residual additive separation module; The residual additive separation module is improved based on the residual learning framework, including: Improvement 1: Use dual-branch heterogeneity, with the dual branches being the speech path and the noise path. The speech path outputs the time-frequency mask through the fully connected layer and Sigmoid , used to process speech signals; the noise path uses 5-point convolution to model the steady-state noise fundamental frequency and 3-point convolution to locate burst pulses, which are used to process noise signals; Improvement 2: Physically guided residual fusion, the fusion formula is proposed: , formula, where For voice protection items, is the noise residual term, ⊙ represents element-by-element multiplication; α is the residual weight coefficient; Improvement 3: Add pulse suppression mechanism, when abrupt energy change is detected, force the fusion formula to <0.1, and the damaged speech is reconstructed by combining the noise residual term.

6. The method for voice noise reduction in data center scenarios according to claim 5, characterized in that: The reverberation suppression unit includes a 12th-order LPC inverse filter and a deep learning speech enhancement module; the input end of the 12th-order LPC inverse filter is used to input the separated speech features, and the output end of the 12th-order LPC inverse filter is connected to the input end of the deep learning speech enhancement module; The transfer function of the 12th-order LPC inverse filter is: in, represents the linear prediction coefficient, which is obtained by minimizing the following objective function; z represents the Z-transform variable; is the filter transfer function; Objective function: Where, T: total duration; t: current time; : historical moment; s(t): pure speech signal; :Pure speech signals of historical moments; The deep learning speech enhancement module includes a transposed convolution layer, a multi-scale dilated convolution layer, and a feature fusion layer connected in sequence; the input end of the transposed convolution layer is the input end of the 12th-order LPC inverse filter; The transposed convolution layer is used to upsample the separated speech features; The multi-scale dilated convolution layer is used to extract features of different scales from the speech features sampled by the transposed convolution layer; The feature fusion layer is used to splice the features of different scales extracted by the multi-scale dilated convolution layer in the channel dimension to obtain and output a higher-dimensional feature map.

7. The method according to claim 6, characterized in that The input end of the 12th-order LPC inverse filter is connected to the output end of the residual additive separation module; The training methods for the multi-scale CNN convolutional neural network, bidirectional long short-term memory network LSTM, residual additive separation module, and reverberation suppression unit include: Build the network architectures of the multi-scale CNN convolutional neural network, bidirectional long short-term memory network LSTM, residual additive separation module, and reverberation suppression unit, and establish the connection relationship between each network architecture; First, the first training set is used to train the multi-scale CNN convolutional neural network-bidirectional long short-term memory network LSTM; Then, the second training set is used to perform generator-discriminator adversarial training on the reverberation suppression unit; Then, the residual additive separation module is trained using the third training set; Then, the fourth training set is used to fine-tune the multi-scale CNN convolutional neural network, the bidirectional long short-term memory network LSTM, the residual additive separation module and the reverberation suppression unit as a whole to obtain the trained multi-scale CNN convolutional neural network, the trained bidirectional long short-term memory network LSTM, the trained residual additive separation module and the trained reverberation suppression unit.

8. A voice noise reduction system for data center scenarios, characterized by: The system includes: The first transform unit is configured to perform STFT time-frequency transform on the speech to be de-noised to obtain a complex spectrum of the speech to be de-noised; A multi-scale CNN convolutional neural network, whose input end is connected to the first transformation unit, is used to extract features from the complex spectrum using the multi-scale CNN convolutional neural network; A bidirectional long short-term memory network (LSTM) whose input is connected to the output of the multi-scale CNN convolutional neural network and is used to perform temporal modeling on the features extracted by the multi-scale CNN convolutional neural network; The residual additive separation module, whose input is connected to the bidirectional long short-term memory network (LSTM), is used to separate the speech features from the features after time series modeling; A reverberation suppression unit, used to suppress reverberation of the separated speech features; The second transformation unit is used to perform inverse STFT reconstruction on the speech features after reverberation suppression to obtain the denoised speech.

9. A terminal, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the computer method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that The readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the computer method according to any one of claims 1 to 7.

Images