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

Abstract

The application discloses a call voice real-time noise reduction method and system based on dynamic noise perception, acquires a noisy original voice signal, carries out frame division and windowing pretreatment on the digital voice signal with noise, extracts the spectral characteristics of the noise, divides the noise into steady-state noise, non-steady-state noise and burst noise according to the spectral characteristics of the noise, constructs a lightweight neural network structure through model pruning and quantization, adjusts the parameters and strategies of the types of the steady-state noise, the non-steady-state noise and the burst noise to realize optimized reduction, carries out post-processing such as window removal and overlap addition on the voice signal after noise reduction, and outputs the voice signal after noise reduction; the application effectively removes the background noise in a complex noise environment through dynamic noise perception and an adaptive noise reduction strategy, significantly improves the voice quality, meets the real-time requirement, realizes real-time noise reduction processing under low delay conditions through lightweight model design, and meets the real-time requirement of voice communication and voice control.

Description

Technical Field

[0001] This invention relates to the field of speech signal processing technology, specifically to a method and system for real-time noise reduction of call speech based on dynamic noise perception. Background Technology

[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 technology is increasing. Especially in real-time voice call scenarios, 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 superior performance, their high computational complexity and memory consumption make them difficult to deploy directly in lightweight devices. Traditional noise reduction methods (such as spectral subtraction and Wiener filtering) have limited effectiveness in handling dynamic and non-stationary noise, failing to effectively remove complex noise and leading to a decline in speech quality.

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

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for real-time noise reduction of voice calls based on dynamic noise perception.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] This invention relates to a lightweight real-time noise reduction method for call speech based on dynamic noise perception, comprising the following steps:

[0007] Step 1: Acquire the original speech signal with noise, and convert the analog speech signal into a digital speech signal through an analog-to-digital converter;

[0008] Step 2: Perform frame segmentation and windowing preprocessing on the noisy digital speech signal to obtain the digital speech signal to be processed;

[0009] Step 3: Analyze the digital speech signal to be processed, extract the spectral characteristics of the noise, and classify the noise into steady-state noise, non-steady-state noise, and burst noise based on the spectral characteristics of the noise.

[0010] Step 4: Construct a lightweight neural network structure through model pruning and quantization; and train the model with a large amount of noisy and clean speech to obtain a lightweight RNN model.

[0011] Step 5: The lightweight RNN model adjusts parameters and strategies to optimize noise reduction for steady-state noise, non-steady-state noise, and burst noise.

[0012] Step 6: Perform windowing and overlap addition post-processing on the denoised speech signal to output the denoised speech signal.

[0013] As a preferred embodiment of the present invention, the method for preprocessing noisy digital speech signals by framing is to divide the noisy digital speech signals into frames of a fixed length, with a certain overlap between the frames; the method for preprocessing noisy digital speech signals by windowing is to add a Hamming window to each frame of speech signal.

[0014] As a preferred embodiment of the present invention, the method for constructing a lightweight neural network structure involves improving the existing RNN structure. The improvement method includes the following steps:

[0015] Step 1: Lighten the basic structure by compressing the dimensions of the hidden layer. Use a bidirectional structure to reduce the dimensions of the unidirectional hidden layer to 32-64 dimensions.

[0016] A parameter sharing mechanism is introduced to share the weight matrix between time steps, and the forget gate is coupled to the input gate. ;

[0017] Step 2: Create a lightweight cell, remove the output gates in a traditional LSTM, and introduce a step-skip connection;

[0018] Step 3: Perform gating merging, merging the input gate and the forget gate into a complementary relationship;

[0019] Step 4: Perform structured pruning. Prune the four gate matrices of the LSTM in blocks and remove 30%-50% of the parameters based on the sensitivity analysis of the gradient magnitude.

[0020] Step 5: Apply dynamic fixed-point quantization to the hidden state to obtain a lightweight neural network.

[0021] As a preferred embodiment of the present invention, the method for compressing the hidden layer dimension has the following expression:

[0022] ,

[0023] in For input dimensions, Represented as the hidden layer dimension, where Compressed to 1 / 4 of the original design, FLOPs represent the number of floating-point operations.

[0024] As a preferred embodiment of the present invention, the method for introducing a step-jump connection is to add a step-jump connection item. The hidden states introduced in the first two steps have the following formula:

[0025] ,

[0026] In the above formula, α is a learnable decay factor, W is the weight matrix from input to hidden layer, and U is the weight matrix from hidden layer to hidden layer, consistent with traditional RNNs, responsible for fusing the current input with the most recent state; For input vectors Linear transformation.

[0027] As a preferred embodiment of the present invention, step 3 is specifically described as follows: after linear transformation, it is divided into three independent gates: the input gate... Forgotten Gate Output gate ,

[0028] The calculation methods for each subject are as follows:

[0029] ;

[0030] ;

[0031] ;

[0032] In the above formula, the sigmoid function combines the input gate and the forget gate into a complementary relationship. and Replace independent parameters;

[0033] Then there is ;

[0034] Among them, the above formula is provided This indicates the current state of the cell; It represents the cell state at the previous moment; This represents the output value of the forget gate, which is generated by the Sigmoid function. This represents a candidate state, generated by a linear combination of the input and the hidden state followed by an activation function. This represents element-wise multiplication.

[0035] As a preferred embodiment of the present invention, the method of using dynamic fixed-point quantization for the hidden state is; Where b=4,

[0036] In the above formula Here, b is the approximate value after quantization, and b is the number of quantization bits. These are raw floating-point values, such as the hidden states in a neural network.

[0037] As a preferred technical solution of the present invention, a system is used to implement the above-mentioned lightweight real-time noise reduction method for call speech based on dynamic noise perception, including a speech acquisition module, a preprocessing module, a dynamic noise perception module, a lightweight noise reduction module, a post-processing module, and an output module.

[0038] The voice acquisition module is used to acquire voice signals and convert analog signals into digital signals through an analog-to-digital converter. The preprocessing module performs frame segmentation and windowing preprocessing on the acquired noisy digital voice signals. The dynamic noise perception module extracts the spectral characteristics of noise in real time through Mel spectrum analysis, and classifies noise into types such as steady-state noise, non-steady-state noise, and sudden noise based on the noise spectral characteristics. The parameters and strategies of the noise reduction model are dynamically adjusted according to the noise type. The lightweight noise reduction module constructs a lightweight neural network structure through model pruning and quantization. The model is trained with a large number of noisy and clean voices to obtain a lightweight RNN model, which optimizes noise reduction for steady-state noise, non-steady-state noise, and sudden noise. The post-processing module performs windowing and overlapping post-processing on the denoised voice signal. The output module outputs the denoised voice signal to a speaker or storage device.

[0039] The beneficial effects of this invention are:

[0040] This lightweight real-time noise reduction method for call speech based on dynamic noise perception improves noise reduction performance: It effectively removes background noise in complex noisy environments through dynamic noise perception and adaptive noise reduction strategies, significantly improving speech quality; it meets real-time requirements: through lightweight model design, it achieves real-time noise reduction processing under low latency conditions, meeting the real-time needs of voice communication and voice control; it reduces computational resource consumption: the lightweight neural network structure ensures efficient operation in embedded devices with limited computing resources; and it enhances dynamic noise adaptability: the dynamic noise perception module analyzes noise characteristics in real time and adjusts the noise reduction strategy, improving the robustness and adaptability of noise reduction. Attached Figure Description

[0041] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0042] Figure 1This is a flowchart illustrating the lightweight real-time noise reduction method for call speech based on dynamic noise perception according to the present invention.

[0043] Figure 2 This is a system block diagram of the real-time noise reduction method and system for call speech based on dynamic noise perception, which is based on the present invention. Detailed Implementation

[0044] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0045] Example: Figure 1 As shown, the present invention provides a real-time noise reduction method and system for call speech based on dynamic noise perception, comprising the following steps:

[0046] Step 1: Acquire the original speech signal with noise, and convert the analog speech signal into a digital speech signal through an analog-to-digital converter;

[0047] Step 2: Perform frame segmentation and windowing preprocessing on the noisy digital speech signal to obtain the digital speech signal to be processed;

[0048] Step 3: Analyze the digital speech signal to be processed, extract the spectral characteristics of the noise, and classify the noise into steady-state noise, non-steady-state noise, and burst noise based on the spectral characteristics of the noise.

[0049] Step 4: Construct a lightweight neural network structure through model pruning and quantization; and train the model with a large amount of noisy and clean speech to obtain a lightweight RNN model.

[0050] Step 5: The lightweight RNN model adjusts parameters and strategies to optimize noise reduction for steady-state noise, non-steady-state noise, and burst noise.

[0051] Step 6: Perform windowing and overlap addition post-processing on the denoised speech signal to output the denoised speech signal.

[0052] Currently, the mainstream noise reduction methods mainly include the following categories:

[0053] Traditional signal processing methods:

[0054] Spectral subtraction: This method reduces noise by estimating the noise spectrum and subtracting it from the noisy speech spectrum. It is computationally simple, but performs poorly with non-stationary noise and is prone to introducing musical noise.

[0055] Wiener filtering: Filters designed based on the minimum mean square error criterion are suitable for stable noise environments, but have poor adaptability to dynamic noise.

[0056] Adaptive filtering: It adjusts filter parameters through adaptive algorithms and is suitable for specific types of noise, but it has high computational complexity and is difficult to apply in real-time systems.

[0057] Deep learning methods:

[0058] Denoising methods based on deep neural networks, such as RNNoise and DNN, effectively handle complex noisy environments by training the model with a large amount of noisy and clean speech. However, these methods are typically computationally complex and difficult to implement in real-time on resource-constrained embedded devices.

[0059] Hybrid methods: Combining traditional signal processing methods with deep learning techniques, such as deep learning-based spectrum enhancement methods, can improve noise reduction to some extent, but still suffer from high computational complexity and poor real-time performance.

[0060] Deficiencies of existing technology

[0061] Traditional methods have limited effectiveness in dealing with complex noise environments (such as non-stationary noise and sudden noise), making it difficult to meet the needs of practical applications.

[0062] While deep learning methods offer good noise reduction, they typically require significant computing resources, making them difficult to run in real-time on embedded devices.

[0063] Insufficient real-time performance:

[0064] 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.

[0065] In real-time voice call scenarios, the real-time performance of noise reduction algorithms is a key indicator, and existing methods are unable to meet this requirement.

[0066] High computational resource consumption:

[0067] Deep learning methods typically require significant computational and memory usage, making them difficult to deploy in resource-constrained embedded devices.

[0068] While traditional methods have lower computational complexity, their noise reduction effect is limited and cannot meet the needs of practical applications.

[0069] This 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 strategies, it effectively removes background noise in complex noisy environments, significantly improving speech quality. It meets real-time requirements: through lightweight model design, real-time noise reduction processing is achieved under low latency conditions, satisfying the real-time needs of voice communication and voice control. It reduces computational resource consumption: by employing a lightweight neural network structure and optimization techniques, it ensures efficient operation in embedded devices with limited computing resources. It enhances dynamic noise adaptability: by using a dynamic noise perception module to analyze noise characteristics in real time and adjust the noise reduction strategy, it improves the robustness and adaptability of noise reduction.

[0070] This invention can operate efficiently in various application scenarios such as industrial machine communication, smart home devices, and vehicle communication systems, and has broad application value.

[0071] By employing a lightweight neural network structure and optimization techniques, it ensures efficient operation in embedded devices with limited computing resources, solving the problem of high computing resource consumption in deep learning methods. Through lightweight model design and hardware acceleration technology, it achieves real-time noise reduction processing under low latency conditions, meeting the real-time requirements of voice communication and voice control.

[0072] This invention enables efficient and real-time noise suppression in environments with low computing resources. The method analyzes noise characteristics in real time through a dynamic noise perception module and uses a lightweight noise reduction model to perform real-time noise reduction processing on the speech signal, significantly improving speech quality.

[0073] The method for preprocessing noisy digital speech signals by framing is to divide the noisy digital speech signal into frames of a fixed length, with a certain overlap between frames; the method for preprocessing noisy digital speech signals by windowing is to add a Hamming window to each frame of speech signal; the speech signal is divided into frames of a fixed length (20ms), with a certain overlap between frames (10ms); windowing is applied to each frame of speech signal (such as a Hamming window) to reduce spectral leakage.

[0074] The method for constructing a lightweight neural network structure involves improving the existing RNN structure. The improvement method includes the following steps:

[0075] Step 1: Lightweight the basic structure by compressing the hidden layer dimension. A bidirectional structure is adopted, reducing the dimension of the unidirectional hidden layer to 32-64 dimensions. The compression of the hidden layer dimension directly reduces the scale of matrix operations, the number of parameters, and memory usage. The reduction in computation directly reduces chip power consumption. By reducing the model capacity (reducing the degree of freedom of parameters), more robust feature representations are forced to be learned. The compression of the hidden layer dimension can reduce the risk of gradient explosion / vanishing, as the condition number of the parameter matrix improves with the reduction of dimension.

[0076] Low-dimensional hidden layers make the optimization surface smoother, and optimizers such as SGD find the optimal solution more easily, thereby improving training efficiency and reducing the bandwidth requirements for parameter synchronization during distributed training.

[0077] A parameter sharing mechanism is introduced to share the weight matrix between time steps, and the forget gate is coupled to the input gate. Sharing the weight matrix between time steps significantly improves parameter efficiency. The number of parameters remains constant. Regardless of how the sequence length changes, the number of model parameters is determined only 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, leading to memory explosion. This reduces storage and transmission costs. Weight sharing reduces the model file size by tens to hundreds of times.

[0078] Furthermore, by coupling the forget gate with the input gate, the input gate (i_t) no longer needs an independent weight matrix (in traditional LSTM, (W_{xi}, W_{hi}) is omitted), reducing the number of parameters by about 25%; the combined gating computation is reduced, and FLOPs are reduced by about 1 / 4.

[0079] Step 2: Build lightweight units, remove the output gates in traditional LSTM, and introduce straddle connections; straddle connections (k=2) can effectively model cross-frame dependencies between phonemes, reduce computational latency, force the model to reuse historical states through straddle connections, and reduce repeated learning of similar patterns.

[0080] Step 3: Perform gating merging, combining the input gate and 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%, avoiding the gradient vanishing problem caused by both gates closing simultaneously; enforce (it + ft = 1) to ensure that at least one gate is active (it approaches 1 when ft approaches 0), avoiding the simultaneous closure of gradient propagation paths; independent input gates and forget gates may learn contradictory strategies (such as closing simultaneously), the coupled design eliminates such conflicts through mathematical constraints.

[0081] Step 4: Perform structured pruning. Prune the four gate matrices of the LSTM in blocks and remove 30%-50% of the parameters based on the sensitivity analysis of the gradient magnitude.

[0082] Step 5: Apply dynamic fixed-point quantization to the hidden state to obtain a lightweight 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 bias accumulation.

[0083] The method for compressing the hidden layer dimension has the following expression:

[0084] ,

[0085] in For input dimensions, Represented as the hidden layer dimension, where Compression to 1 / 4 of the original design, FLOPs represent the number of floating-point operations. Compressing the hidden layer dimension directly reduces the scale of matrix operations, the number of parameters decreases, and memory usage is reduced: reducing computation directly reduces chip power consumption; by reducing model capacity (reducing parameter degrees of freedom), it forces the learning of more robust feature representations. Hidden layer dimension compression can reduce the risk of gradient explosion / vanishing, as the condition number of the parameter matrix improves with decreasing dimension.

[0086] The method for introducing step jump connections involves adding a new step jump connection term. By directly introducing the hidden states from the first two steps, we have the following formula:

[0087] ,

[0088] In the above formula, α is a learnable decay factor, W is the weight matrix from input to hidden layer, and U is the weight matrix from hidden layer to hidden layer, consistent with traditional RNNs, responsible for fusing the current input with the most recent state; For input vectors The linear transformation; step connections (k=2) can effectively model cross-frame dependencies between phonemes, reduce computational latency, force the model to reuse historical states through step connections, and reduce repeated learning of similar patterns.

[0089] Specifically, step 3 involves dividing the input gate into three independent gates after a linear transformation: the input gate... Forgotten Gate Output gate ,

[0090] The calculation methods for each subject are as follows:

[0091] ;

[0092] ;

[0093] ;

[0094] In the above formula, the sigmoid function combines the input gate and the forget gate into a complementary relationship. and Replace independent parameters;

[0095] Then there is ;

[0096] Among them, the above formula is provided This indicates the current state of the cell; It represents the cell state at the previous moment; This represents the output value of the forget gate, which is generated by the Sigmoid function. This represents a candidate state, generated by a linear combination of the input and the hidden state followed by an activation function. This represents element-wise multiplication; the weight matrix of the input gate is removed, reducing the total number of parameters by about 25%, and preventing the gradient vanishing problem caused by both gates closing simultaneously. This mitigates the gradient vanishing problem; it forces (it + ft = 1) to ensure that at least one gate is active (it approaches 1 when ft approaches 0), avoiding the simultaneous closure of gradient propagation paths; independent input gates and forget gates may learn contradictory strategies (such as closing simultaneously), and the coupled design eliminates such conflicts through mathematical constraints.

[0097] The method of using dynamic fixed-point quantization for the hidden state is as follows: Where b=4,

[0098] In the above formula Here, b is the approximate value after quantization, and b is the number of quantization bits. The original floating-point values ​​are used, such as the hidden states 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 bias accumulation.

[0099] In industrial machine communication, this invention can effectively remove industrial noise and improve voice communication quality. 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 environmental noise on voice control and intercom systems.

[0100] A system, such as Figure 2 As shown, the method for implementing the above-mentioned lightweight real-time noise reduction method for call speech based on dynamic noise perception includes a speech 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 incorporates hardware acceleration technology to achieve real-time noise reduction processing under low latency conditions, meeting the real-time requirements of voice communication and voice control.

[0101] The voice acquisition module 1 is used to acquire voice signals and convert analog signals into digital signals through an analog-to-digital converter; the preprocessing module 2 performs frame segmentation and windowing preprocessing on the acquired noisy digital voice signals; the dynamic noise perception module 3 extracts the spectral characteristics of noise in real time through Mel spectrum analysis, and classifies noise into types such as steady-state noise, non-steady-state noise, and sudden 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 with a large number of noisy and clean voices to obtain a lightweight RNN model, which optimizes noise reduction for steady-state noise, non-steady-state noise, and sudden noise; the post-processing module 5 performs windowing and overlapping post-processing on the denoised voice signal; and the output module 6 outputs the denoised voice signal to a speaker or storage device.

[0102] 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 foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A lightweight real-time noise reduction method for call speech based on dynamic noise perception, characterized in that, Includes the following steps: Step 1: Acquire 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: Perform frame segmentation and windowing preprocessing on the noisy digital speech signal to obtain the digital speech signal to be processed; Step 3: Analyze the digital speech signal to be processed, extract the spectral characteristics of the noise, and classify the noise into steady-state noise, non-steady-state noise, and burst noise based on the spectral characteristics of the noise. Step 4: Construct a lightweight neural network structure through model pruning and quantization; and train the model with a large amount of noisy and clean speech to obtain a lightweight RNN model; Step 5: Utilize a lightweight RNN model to adjust 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 to output the denoised speech signal; The method for constructing a lightweight neural network structure involves improving the existing RNN structure. The improvement method involves the following steps: Step 1: Lighten the basic structure by compressing the dimensions of the hidden layer. Use a bidirectional structure to reduce the dimensions of the unidirectional hidden layer to 32-64 dimensions. A parameter sharing mechanism is introduced to share the weight matrix between time steps, and the forget gate is coupled to the input gate. ; Step 2: Create a lightweight cell, remove the output gates in a traditional LSTM, and introduce a step-skip connection; Step 3: Perform gating merging, merging the input gate and the forget gate into a complementary relationship; Step 4: Perform structured pruning. Prune the four gate matrices of the LSTM in blocks and remove 30%-50% of the parameters based on the sensitivity analysis of the gradient magnitude. Step 5: Apply dynamic fixed-point quantization to the hidden states to obtain a lightweight neural network. The method for compressing the hidden layer dimension has the following expression: ， in For input dimensions, Represented as the hidden layer dimension, where Compressed to 1 / 4 of the original design, FLOPs represent the number of floating-point operations.

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

3. The lightweight real-time noise reduction method for call speech based on dynamic noise perception according to claim 1, characterized in that, The method for introducing step jump connections is to add a new step jump connection term. By directly introducing the hidden states from the first two steps, we have the following formula: ， In the above formula, α is a learnable decay factor, W is the weight matrix from input to hidden layer, and U is the weight matrix from hidden layer to hidden layer, consistent with traditional RNNs, responsible for fusing the current input with the most recent state; For input vectors Linear transformation.

4. The lightweight real-time noise reduction method for call speech based on dynamic noise perception according to claim 1, characterized in that, The specific method of step 3 is to divide the input gate into 3 independent gates after linear transformation: input gate Forgotten Gate Output gate , • Calculation methods for each subject: · ； · ； ； In the above formula, the sigmoid function combines the input gate and the forget gate into a complementary relationship. and Replace independent parameters; Then there is ; Among them, the above formula is provided This indicates the current state of the cell; It represents the cell state at the previous moment; This represents the output value of the forget gate, which is generated by the Sigmoid function. This represents a candidate state, generated by a linear combination of the input and the hidden state followed by an activation function. This represents element-wise multiplication.

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

6. A system for implementing the lightweight real-time noise reduction method for call speech based on dynamic noise perception as described in any one of claims 1-5, characterized in that: It includes 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 voice acquisition module (1) is used to acquire voice signals and convert analog signals into digital signals through an analog-to-digital converter; the preprocessing module (2) performs frame segmentation and windowing preprocessing on the acquired noisy digital voice signals; the dynamic noise perception module (3) extracts the spectral characteristics of noise in real time through Mel spectrum analysis, and classifies noise into steady-state noise, non-steady-state noise and sudden 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 with a large number of noisy voices and clean voices to obtain a lightweight RNN model, which optimizes the noise reduction of steady-state noise, non-steady-state noise and sudden noise; the post-processing module (5) performs windowing and overlapping post-processing on the denoised voice signal; the output module (6) outputs the denoised voice signal and outputs the denoised voice signal to a speaker or storage device.

Images