Call voice real-time noise reduction method and system based on dynamic noise perception

Abstract

The invention discloses a call voice real-time noise reduction method and system based on dynamic noise perception. The method comprises the following steps: collecting an original voice signal with noise; performing framing and windowing preprocessing on the digital voice signal with noise; extracting spectrum features of the noise, and dividing the noise into steady-state noise, unsteady-state noise and burst noise according to the spectrum features of the noise; quantitatively constructing a lightweight neural network structure through model pruning; if yes, performing parameter and strategy longitude adjustment on the types of the steady-state noise, the unsteady-state noise and the burst noise to realize optimization reduction; performing post-processing of window removal and overlapping addition on the voice signal after noise reduction, and outputting the voice signal after noise reduction; according to the method, background noise in a complex noise environment is effectively removed through dynamic noise perception and a self-adaptive noise reduction strategy, and the voice quality is remarkably improved; real-time noise reduction processing is achieved under the low-delay condition through lightweight model design, and the real-time requirements of voice communication and voice control are met.

Description

Technical Field

[0001] The present invention relates to the technical field related to speech signal processing, and specifically to a method and system for real-time noise reduction of call speech based on dynamic noise perception. Background Art

[0002] In the field of speech signal processing, background noise is one of the main factors affecting speech quality and speech recognition accuracy. With the rapid development of voice communication, speech recognition, and voice control technologies, the demand for real-time noise reduction technologies is increasing. Especially in real-time voice conversations, speech signals are often interfered with by complex background noise (such as industrial noise, environmental noise, and equipment noise). Furthermore, embedded devices (such as industrial robots, smart home devices, and in-vehicle systems) typically have limited computing power and memory resources. While existing deep learning noise reduction models offer excellent performance, their high computational complexity and memory usage make them difficult to deploy directly on lightweight devices. Traditional noise reduction methods (such as spectral subtraction and Wiener filtering) are limited in their effectiveness when dealing with dynamic and non-stationary noise, making it difficult to effectively remove complex noise, resulting in degraded speech quality.

[0003] Furthermore, traditional noise reduction methods are typically optimized for specific noise types and lack the ability to adapt to dynamic noise environments. In practical applications, the type and intensity of noise can change at any time, requiring a method that can perceive noise characteristics in real time and dynamically adjust the noise reduction strategy. Summary of the Invention

[0004] In order to address the defects of the prior art, the present invention provides a method and system for real-time noise reduction of call speech based on dynamic noise perception.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: The lightweight real-time noise reduction method for call speech based on dynamic noise perception of the present invention comprises the following steps: Step 1: collect the original speech signal with noise, and convert the analog speech signal into a digital speech signal through an analog-to-digital converter; Step 2: performing frame division and windowing preprocessing on the digital speech signal with noise to obtain a digital speech signal to be processed; Step 3: Analyze the digital speech signal to be processed, extract the spectrum characteristics of the noise, and classify the noise into steady-state noise, non-steady-state noise and burst noise according to the noise spectrum characteristics; Step 4: Build a lightweight neural network structure through model pruning and quantization; and train the model with a large amount of noisy speech and clean speech to obtain a lightweight RNN model.

[0006] Step 5: The lightweight RNN model adjusts parameters and strategies for steady-state noise, non-steady-state noise, and burst noise to achieve optimized noise reduction. Step 6: Perform windowing and overlap-addition post-processing on the denoised speech signal, and output the denoised speech signal.

[0007] As a preferred technical solution of the present invention, the method of performing framing preprocessing on the noisy digital speech signal is to frame the noisy digital speech signal according to a fixed length, and there is a certain length of overlap between frames; the method of performing windowing preprocessing on the noisy digital speech signal is to add a Hamming window to each frame of the speech signal.

[0008] As a preferred technical solution of the present invention, the method of constructing a lightweight neural network structure is to improve the existing RNNJ structure. The steps of the improvement method are as follows: Step 1: Lightweight the basic structure and compress the hidden layer dimension. Use a bidirectional structure to reduce the unidirectional hidden layer dimension to 32-64 dimensions. Introduce a parameter sharing mechanism to share the weight matrix between time steps and couple the forget gate with the input gate as ; Step 2: Create a lightweight unit, remove the output gate in the traditional LSTM, and introduce a stride jump connection; Step 3: Perform gate merging to merge the input gate and the forget gate into a complementary relationship; Step 4: Perform structured pruning to implement block pruning on the four gating matrices of LSTM and remove 30%-50% of the parameters based on the sensitivity analysis of the gradient amplitude. Step 5: Use dynamic fixed-point quantization on the hidden state to obtain a lightweight neural network.

[0009] As a preferred technical solution of the present invention, the method for compressing the hidden layer dimension has the following expression: , in is the input dimension, is represented as the hidden layer dimension, where Compressed to 1 / 4 of the original design, FLOPs refers to the number of floating-point operations.

[0010] As a preferred technical solution of the present invention, the method of introducing stride jump connection is to add a stride connection item , directly introducing the hidden state of the first two steps has the following formula; , In the above formula, α is a learnable attenuation factor, W is the weight matrix input to the hidden layer, and U is the weight matrix from the hidden layer to the hidden layer, which is consistent with the traditional RNN and is responsible for the fusion of the current input and the most recent state. is the input vector Linear transformation of .

[0011] As a preferred technical solution of the present invention, the specific method of step 3 is to divide it into three independent gates after linear transformation: input gate , Forget Gate , output gate , where each gate is calculated as follows: ; ; ; In the above formula, the sigmoid function combines the input gate and the forget gate into a complementary relationship. and Instead of independent parameters; Then there is ; Among them, the above formula It represents the cell state at the current moment; It represents the cell state at the previous moment; It represents the output value of the forget gate, which is generated by the Sigmoid function. It represents the candidate state, which is generated by the activation function after the linear combination of input and hidden state. represents element-wise multiplication.

[0012] As a preferred technical solution of the present invention, the method of using dynamic fixed-point quantization for the hidden state is: , where b=4, where is the approximate value after quantization, b is the number of quantization bits, are raw floating-point values, such as hidden states in a neural network.

[0013] As a preferred technical solution of the present invention, a system is provided for implementing the above-mentioned lightweight call voice real-time noise reduction method based on dynamic noise perception, comprising a voice acquisition module, a preprocessing module, a dynamic noise perception module, a lightweight noise reduction module, a post-processing module, and an output module; The speech acquisition module is used to collect speech signals and convert the analog signals into digital signals through an analog-to-digital converter; the preprocessing module performs frame division and windowing preprocessing on the collected noisy digital speech signals; the dynamic noise perception module extracts the spectral characteristics of the noise in real time through Mel spectrum analysis, and divides the noise into types such as steady-state noise, non-steady-state noise and burst noise according to the noise spectral characteristics, and dynamically adjusts the parameters and strategies of the noise reduction model according to the noise type; the lightweight noise reduction module constructs a lightweight neural network structure through model pruning and quantization; and trains the model through a large number of noisy speech and pure speech to obtain a lightweight RNN model, which optimizes the noise reduction of steady-state noise, non-steady-state noise and burst noise types; the post-processing module performs windowing and overlap-addition post-processing on the denoised speech signal; the output module outputs the denoised speech signal and outputs the denoised speech signal to a speaker or a storage device.

[0014] The beneficial effects of the present invention are: This lightweight real-time noise reduction method for call speech based on dynamic noise perception improves the noise reduction effect: through dynamic noise perception and adaptive noise reduction strategy, it effectively removes background noise in complex noise environments and significantly improves voice quality; meets real-time requirements: through lightweight model design, it realizes real-time noise reduction processing under low-latency conditions to meet the real-time requirements of voice communication and voice control; reduces computing resource usage: adopts a lightweight neural network structure to ensure efficient operation in embedded devices with limited computing resources; enhances dynamic noise adaptability: through the dynamic noise perception module, it analyzes noise characteristics in real time and adjusts the noise reduction strategy, thereby improving the robustness and adaptability of noise reduction. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a flow chart of a lightweight real-time noise reduction method for call speech based on dynamic noise perception according to the present invention; Figure 2 This is a system block diagram of the method and system for real-time noise reduction of call speech based on dynamic noise perception of the present invention. DETAILED DESCRIPTION

[0016] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0017] Example: Figure 1 As shown, the present invention provides a method and system for real-time noise reduction of call speech based on dynamic noise perception, comprising the following steps: Step 1: collect the original speech signal with noise, and convert the analog speech signal into a digital speech signal through an analog-to-digital converter; Step 2: performing frame division and windowing preprocessing on the digital speech signal with noise to obtain a digital speech signal to be processed; Step 3: Analyze the digital speech signal to be processed, extract the spectrum characteristics of the noise, and classify the noise into steady-state noise, non-steady-state noise and burst noise according to the noise spectrum characteristics; Step 4: Build a lightweight neural network structure through model pruning and quantization; and train the model with a large amount of noisy speech and clean speech to obtain a lightweight RNN model.

[0018] Step 5: The lightweight RNN model adjusts parameters and strategies for steady-state noise, non-steady-state noise, and burst noise to achieve optimized noise reduction. Step 6: Perform windowing and overlap-addition post-processing on the denoised speech signal, and output the denoised speech signal.

[0019] At present, the mainstream noise reduction methods mainly include the following categories: Traditional signal processing methods: Spectral subtraction: Denoise is achieved by estimating the noise spectrum and subtracting it from the noisy speech spectrum. This method is computationally simple, but is less effective when dealing with non-stationary noise and is prone to introducing musical noise.

[0020] Wiener filtering: The filter is designed based on the minimum mean square error criterion. It is suitable for stationary noise environments, but has poor adaptability to dynamic noise.

[0021] Adaptive filtering: This method adjusts filter parameters through an adaptive algorithm. It is suitable for specific types of noise, but its computational complexity is high and it is difficult to apply in real-time systems.

[0022] Deep Learning Methods: Denoising methods based on deep neural networks, such as RNNoise and DNN, effectively handle complex noise environments by training models using large amounts of noisy and clean speech. However, these methods often have high computational complexity, making them difficult to implement in real time on resource-constrained embedded devices.

[0023] Hybrid method: Combining traditional signal processing methods and deep learning technology, such as the spectrum enhancement method based on deep learning, although it improves the noise reduction effect to a certain extent, it still has problems such as high computational complexity and poor real-time performance.

[0024] Deficiencies of the Existing Technology Traditional methods have limited effectiveness in dealing with complex noise environments (such as non-stationary noise and burst noise) and are difficult to meet practical application needs.

[0025] Although deep learning methods have good noise reduction effects, they usually require high computing resources and are difficult to run in real time on embedded devices.

[0026] Lack of real-time performance: Traditional methods and some deep learning methods have high computational complexity, making it difficult to achieve real-time noise reduction processing under low-latency conditions.

[0027] In real-time voice call scenarios, the real-time performance of the noise reduction algorithm is a key indicator, and existing methods are difficult to meet this requirement.

[0028] High computing resource usage: Deep learning methods usually require high computing resources and memory usage, making them difficult to deploy in resource-constrained embedded devices.

[0029] Although traditional methods have low computational complexity, their noise reduction effects are limited and cannot meet the needs of practical applications.

[0030] The present invention proposes a real-time noise reduction method and system for call speech based on dynamic noise perception. Through dynamic noise perception and adaptive noise reduction strategy, the method and system can effectively remove background noise in complex noise environments and significantly improve voice quality. The method meets real-time requirements: through lightweight model design, real-time noise reduction processing is achieved under low-latency conditions to meet the real-time requirements of voice communication and voice control. The method reduces computing resource usage: adopts lightweight neural network structure and optimization technology to ensure efficient operation in embedded devices with limited computing resources. The method enhances dynamic noise adaptability: through the dynamic noise perception module, the noise characteristics are analyzed in real time and the noise reduction strategy is adjusted, thereby improving the robustness and adaptability of noise reduction.

[0031] The present invention can operate efficiently in various application scenarios such as industrial machine communication, smart home equipment, and vehicle communication systems, and has wide application value.

[0032] The lightweight neural network structure and optimization technology are used to ensure efficient operation in embedded devices with limited computing resources, solving the problem of high computing resource usage of deep learning methods. Through lightweight model design and hardware acceleration technology, real-time noise reduction processing is achieved under low latency conditions to meet the real-time requirements of voice communication and voice control.

[0033] This method can achieve efficient, real-time noise suppression in a low-computing resource environment. It uses a dynamic noise perception module to analyze noise characteristics in real time and a lightweight noise reduction model to perform real-time noise reduction on speech signals, significantly improving speech quality.

[0034] Among them, the method of framing preprocessing the noisy digital speech signal is to frame the noisy digital speech signal according to a fixed length, and there is a certain length of overlap between frames; the method of windowing preprocessing the noisy digital speech signal is to add a Hamming window to each frame of the speech signal; the speech signal is framed according to a fixed length (20ms), and there is a certain length of overlap between frames (10ms); windowing: adding a window (such as a Hamming window) to each frame of the speech signal to reduce spectral leakage.

[0035] The method of constructing a lightweight neural network structure is to improve the existing RNNJ structure. The steps of the improvement method are as follows: Step 1: Lightweight the infrastructure and compress the hidden layer dimensions. Using a bidirectional structure, reduce the unidirectional hidden layer dimensions to 32-64 dimensions. Compression of the hidden layer dimensions directly reduces the scale of matrix operations, the number of parameters, and memory usage. Reducing the amount of computation directly reduces chip energy consumption. By reducing the model capacity (reducing the parameter degrees of freedom), a more robust feature expression is forced to be learned. Compression of the hidden layer dimensions can reduce the risk of gradient explosion / vanishing because the parameter matrix condition number improves as the dimension decreases.

[0036] The low-dimensional hidden layer makes the optimization surface smoother, making it easier for optimizers such as SGD to find the optimal solution, thereby improving training efficiency and reducing the parameter synchronization bandwidth requirement during distributed training.

[0037] Introduce a parameter sharing mechanism to share the weight matrix between time steps and couple the forget gate with the input gate as Sharing the weight matrix between time steps significantly improves parameter efficiency and keeps the number of parameters constant. Regardless of how the sequence length changes, the number of model parameters is only determined by the dimension of the weight matrix, rather than growing linearly with the sequence length. If independent weights are used for each time step, the number of parameters will become very large, which will lead to memory explosion and reduce storage and transmission costs. Weight sharing reduces the size of model files by tens to hundreds of times.

[0038] The forget gate is coupled with the input gate, so that the input gate (i_t) no longer requires an independent weight matrix (in traditional LSTM (W_{xi}, W_{hi}) is omitted), reducing the number of parameters by about 25%; the gate calculation amount after merging is reduced, and FLOPs are reduced by about 1 / 4.

[0039] Step 2: Build a lightweight unit, remove the output gate in the traditional LSTM, and introduce stride jump connections. Stride connections (k=2) can effectively model cross-frame dependencies between phonemes, reduce computational latency, and force the model to reuse historical states through stride connections, reducing repeated learning of similar patterns.

[0040] Step 3: Merge gates to combine the input gate and the forget gate into a complementary relationship. Parameter reduction: The weight matrix of the input gate is removed, reducing the total number of parameters by approximately 25%. This prevents vanishing gradients caused by closing both gates simultaneously. This alleviates the vanishing gradient problem. Enforcing (it + ft = 1) ensures that at least one gate is active (it approaches 1 when ft approaches 0), preventing simultaneous closure of the gradient propagation path. Independent input and forget gates may learn conflicting strategies (e.g., closing simultaneously). The coupled design eliminates such conflicts through mathematical constraints.

[0041] Step 4: Perform structured pruning to implement block pruning on the four gating matrices of LSTM and remove 30%-50% of the parameters based on the sensitivity analysis of the gradient amplitude. Step 5: Use dynamic fixed-point quantization for the hidden state to obtain a lightweight neural network. The real-time scaling factor reduces the quantization error caused by the input data distribution offset. The quantization error is dynamically adjusted with back propagation to alleviate the problem of gradient deviation accumulation.

[0042] The method for compressing the hidden layer dimension is as follows: , in is the input dimension, is represented as the hidden layer dimension, where The design is compressed to 1 / 4 of the original design. FLOPs represents the number of floating-point operations. Compressing the hidden layer dimensions directly reduces the scale of matrix operations, reducing the number of parameters and memory usage. Reducing computational effort directly reduces chip energy consumption. By reducing model capacity (reducing parameter degrees of freedom), it forces learning of more robust feature representations. Hidden layer dimension compression reduces the risk of exploding / vanishing gradients, as the parameter matrix condition number improves with decreasing dimensionality.

[0043] The method of introducing stride jump connection is to add a stride connection term , directly introduce the hidden state of the first two steps, with the following formula; , In the above formula, α is a learnable attenuation factor, W is the weight matrix input to the hidden layer, and U is the weight matrix from the hidden layer to the hidden layer, which is consistent with the traditional RNN and is responsible for the fusion of the current input and the most recent state. is the input vector The stride connection (k=2) can effectively model the cross-frame dependency between phonemes, reduce computational latency, and force the model to reuse historical states through the stride connection to reduce repeated learning of similar patterns.

[0044] The specific method of step 3 is to divide it into three independent gates after linear transformation: input gate , Forget Gate , output gate , where each gate is calculated as follows: ; ; ; In the above formula, the sigmoid function combines the input gate and the forget gate into a complementary relationship. and Instead of independent parameters; Then there is ; Among them, the above formula It represents the cell state at the current moment; It represents the cell state at the previous moment; It represents the output value of the forget gate, which is generated by the Sigmoid function. It represents the candidate state, which is generated by the activation function after the linear combination of input and hidden state. represents element-by-element multiplication. The weight matrix of the input gate is removed, reducing the total number of parameters by approximately 25% and preventing vanishing gradients caused by closing both gates simultaneously. This alleviates the vanishing gradient problem. Enforcing (it + ft = 1) ensures that at least one gate is active (it approaches 1 when ft approaches 0), preventing simultaneous closure of the gradient propagation path. Independent input gates and forget gates may learn conflicting strategies (e.g., closing simultaneously). The coupled design eliminates such conflicts through mathematical constraints.

[0045] Wherein, the method of using dynamic fixed-point quantization for the hidden state is: , where b=4, where is the approximate value after quantization, b is the number of quantization bits, It is a raw floating-point value, such as the hidden state in a neural network. The real-time scaling factor reduces the quantization error caused by the offset of the input data distribution. The quantization error is dynamically adjusted with backpropagation to alleviate the problem of gradient deviation accumulation.

[0046] This invention can effectively remove industrial noise and improve voice communication quality in industrial machine communications. In smart home devices, it can significantly improve the voice recognition accuracy of voice assistants and smart speakers. In vehicle communication systems, it can reduce the interference of ambient noise on voice control and intercom systems.

[0047] A system such as Figure 2As shown, it is used to implement the above-mentioned lightweight call voice real-time noise reduction method based on dynamic noise perception, including a voice acquisition module 1, a preprocessing module 2, a dynamic noise perception module 3, a lightweight noise reduction module 4, a post-processing module 5 and an output module 6; the lightweight noise reduction module also uses hardware acceleration technology to achieve real-time noise reduction processing under low-latency conditions to meet the real-time requirements of voice communication and voice control.

[0048] The speech acquisition module 1 is used to collect speech signals and convert the analog signals into digital signals through an analog-to-digital converter; the preprocessing module 2 performs framing and windowing preprocessing on the collected noisy digital speech signals; the dynamic noise perception module 3 extracts the spectral characteristics of the noise in real time through Mel spectrum analysis, and divides the noise into types such as steady-state noise, non-steady-state noise and burst noise according to the noise spectral characteristics, and dynamically adjusts the parameters and strategies of the noise reduction model according to the noise type; the lightweight noise reduction module 4 constructs a lightweight neural network structure through model pruning and quantization; and trains the model through a large number of noisy speech and pure speech to obtain a lightweight RNN model, which optimizes the noise reduction of steady-state noise, non-steady-state noise and burst noise types; the post-processing module 5 performs windowing and overlap-addition post-processing on the denoised speech signal; the output module 6 outputs the denoised speech signal and outputs the denoised speech signal to a speaker or storage device.

[0049] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A lightweight real-time noise reduction method for call speech based on dynamic noise perception, characterized by: The following steps are involved: Step 1: collect the original speech signal with noise, and convert the analog speech signal into a digital speech signal through an analog-to-digital converter; Step 2: performing frame division and windowing preprocessing on the digital speech signal with noise to obtain a digital speech signal to be processed; Step 3: Analyze the digital speech signal to be processed, extract the spectrum characteristics of the noise, and classify the noise into steady-state noise, non-steady-state noise and burst noise according to the noise spectrum characteristics; Step 4: Build a lightweight neural network structure through model pruning and quantization; and train the model with a large amount of noisy speech and clean speech to obtain a lightweight RNN model. Step 5: Use the lightweight RNN model to adjust the parameters and strategies of steady-state noise, non-steady-state noise and burst noise to achieve optimized noise reduction; Step 6: Perform windowing and overlap-addition post-processing on the denoised speech signal, and output the denoised speech signal.

2. The lightweight real-time noise reduction method for call speech based on dynamic noise perception according to claim 1 is characterized in that: The method for performing framing preprocessing on the digital speech signal with noise is to divide the digital speech signal with noise into frames according to a fixed length, and there is a certain length of overlap between frames; the method for performing windowing preprocessing on the digital speech signal with noise is to add a Hamming window to each frame of the speech signal.

3. The lightweight real-time noise reduction method for call speech based on dynamic noise perception according to claim 1 is characterized in that: The method of constructing a lightweight neural network structure is to improve the existing RNNJ structure. The improvement method steps are as follows: Step 1: Lightweight the basic structure and compress the hidden layer dimension. Use a bidirectional structure to reduce the unidirectional hidden layer dimension to 32-64 dimensions. Introduce a parameter sharing mechanism to share the weight matrix between time steps and couple the forget gate with the input gate as ; Step 2: Create a lightweight unit, remove the output gate in the traditional LSTM, and introduce a stride jump connection; Step 3: Perform gate merging to merge the input gate and the forget gate into a complementary relationship; Step 4: Perform structured pruning to implement block pruning on the four gating matrices of LSTM and remove 30%-50% of the parameters based on the sensitivity analysis of the gradient amplitude. Step 5: Use dynamic fixed-point quantization on the hidden state to obtain a lightweight neural network.

4. The lightweight real-time noise reduction method for call speech based on dynamic noise perception according to claim 3 is characterized in that: The method for compressing the hidden layer dimension is to have the following expression, ， in is the input dimension, is represented as the hidden layer dimension, where Compressed to 1 / 4 of the original design, FLOPs refers to the number of floating-point operations.

5. The lightweight real-time noise reduction method for call speech based on dynamic noise perception according to claim 3 is characterized in that: The method of introducing stride jump connection is to add stride connection item , directly introduce the hidden state of the first two steps, with the following formula, ， In the above formula, α is a learnable attenuation factor, W is the weight matrix input to the hidden layer, and U is the weight matrix from the hidden layer to the hidden layer, which is consistent with the traditional RNN and is responsible for the fusion of the current input and the most recent state. is the input vector Linear transformation of .

6. The lightweight real-time noise reduction method for call speech based on dynamic noise perception according to claim 3 is characterized in that: The specific method of step 3 is to divide it into three independent gates after linear transformation: input gate , Forget Gate , output gate , where each gate is calculated as follows: ； ； ； In the above formula, the sigmoid function combines the input gate and the forget gate into a complementary relationship. and Instead of independent parameters; Then there is ; Among them, the above formula It represents the cell state at the current moment; It represents the cell state at the previous moment; It represents the output value of the forget gate, which is generated by the Sigmoid function. It represents the candidate state, which is generated by the activation function after the linear combination of input and hidden state. represents element-wise multiplication.

7. The lightweight real-time noise reduction method for call speech based on dynamic noise perception according to claim 3 is characterized in that: The method of using dynamic fixed-point quantization for hidden states is: , where b=4, In the above formula is the approximate value after quantization, b is the number of quantization bits, are raw floating-point values, such as hidden states in a neural network.

8. A system for implementing the lightweight real-time noise reduction method for call speech based on dynamic noise perception according to any one of claims 1 to 7, characterized in that: It includes a speech acquisition module (1), a pre-processing module (2), a dynamic noise perception module (3), a lightweight noise reduction module (4), a post-processing module (5) and an output module (6); The speech acquisition module (1) is used to acquire speech signals and convert the analog signals into digital signals through an analog-to-digital converter; the pre-processing module (2) performs frame division and windowing pre-processing on the acquired digital speech signals with noise; the dynamic noise perception module (3) extracts the spectrum characteristics of the noise in real time through Mel spectrum analysis, and divides the noise into types such as steady-state noise, non-steady-state noise and burst noise according to the noise spectrum characteristics, and dynamically adjusts the parameters and strategies of the noise reduction model according to the noise type; the lightweight noise reduction module (4) constructs a lightweight neural network structure through model pruning and quantization; and trains the model through a large number of noisy speech and pure speech to obtain a lightweight RNN model, which optimizes the noise reduction of steady-state noise, non-steady-state noise and burst noise; the post-processing module (5) performs windowing and overlap-addition post-processing on the denoised speech signal; the output module (6) outputs the denoised speech signal and outputs the denoised speech signal to a speaker or a storage device.

Images