Lightweight hybrid dual-channel AI denoising method and system based on embedding

Abstract

The application discloses a kind of based on embedded light weight mixed double-channel AI noise reduction method and system.Belongs to the technical field of audio signal processing, method includes: obtaining the double-path time-domain audio signal of double-microphone acquisition;Short-time Fourier transform is carried out to signal, and first, second complex spectrum is obtained;Complex spectrum is merged and input light weight mixed double-channel AI noise reduction model;Model extracts perception-related sub-band features by equivalent rectangular bandwidth sub-band decomposition, spatial feature extraction is carried out by grouping convolution layer, long-time dependence is captured by time sequence cyclic network layer and noise mask is output;Mask is multiplied with complex spectrum point by point to generate enhanced complex spectrum, and inverse short-time Fourier transform and time-domain post-filtering output final speech.The application realizes low delay, low-power real-time noise reduction on embedded device, and has excellent suppression effect on steady-state and non-steady-state noise.

Description

Technical Field

[0001] This invention relates to the field of audio signal processing technology, and in particular to a lightweight hybrid dual-channel AI noise reduction method and system based on embedded systems. Background Technology

[0002] With the widespread adoption of mobile communications, intelligent voice interaction, and wearable devices, users' demand for high-quality voice communication and interactive experiences is growing daily. In real-life and work environments, voice signals are inevitably contaminated by various noises during transmission from the sound source to the microphone. These noises are diverse, including but not limited to: steady-state noise such as a car engine or air conditioner hum, and more challenging non-steady-state noise such as conversations (human voice noise), sudden car horns (traffic noise), outdoor wind noise, and keyboard typing. The presence of noise not only severely affects the subjective clarity and comfort of calls but also seriously interferes with subsequent advanced voice processing tasks such as voice coding, voice recognition, and speaker verification, leading to decreased recognition rates and system erroneous triggering. Therefore, as the front end of the voice signal processing chain, voice noise reduction technology has become an indispensable and crucial component.

[0003] Traditional speech denoising methods are primarily based on digital signal processing (DSP) theory, including spectral subtraction, Wiener filtering, and minimum mean square error estimation. These methods estimate the power spectrum of the noise and design a gain function to attenuate the noise components in the noisy speech spectrum. The advantages of these methods are low computational cost, simple principles, and the ability to run in real-time on low-power DSP chips. However, they have a fundamental drawback: they typically assume that the noise is stationary or slowly changing and rely on voice activity detection (VAD) to update the noise estimate. In non-steady-state noise environments, the statistical characteristics of the noise change rapidly, and traditional methods cannot track the noise in a timely and accurate manner, leading to a sharp decline in denoising effectiveness—either significant noise residue (insufficient denoising) or excessive suppression of the speech signal (speech distortion).

[0004] To overcome the challenge of non-stationary noise, deep learning-based speech denoising techniques have become a research hotspot in recent years. By constructing deep neural networks (DNNs), such as convolutional neural networks (CNNs), recurrent neural networks, and their variants (LSTM, GRU), complex nonlinear mappings can be learned to derive clean speech or ideal masks from noisy speech. These methods significantly outperform traditional methods in terms of non-stationary noise suppression and speech fidelity. However, high-performance deep models are usually accompanied by massive numbers of parameters and computational operations, resulting in high computational complexity, large memory footprint, and long inference latency. This makes them difficult to deploy directly on embedded devices with extremely limited power consumption, computing power, and memory. For example, a typical TWS earphone main control chip may have a CPU frequency of only tens to hundreds of megahertz and only a few hundred KB to a few MB of memory, and has extremely high requirements for power consumption and real-time performance (usually requiring end-to-end latency <20ms).

[0005] To adapt to embedded environments, academia and industry have proposed lightweight model techniques, such as model pruning, weight quantization (e.g., INT8 quantization), and knowledge distillation. However, simple model compression often comes with significant performance loss. Furthermore, existing lightweight model structures (such as MobileNet and ShuffleNet) are primarily designed for image recognition tasks, and their depthwise separable convolution operations do not fully consider the time-frequency characteristics of speech signals, especially the perceptual characteristics of the human auditory system. Human hearing frequency perception is non-linear, exhibiting an equivalent rectangular bandwidth (ERB) characteristic—high resolution at low frequencies and low resolution at high frequencies. Most existing AI denoising models directly process the uniform STFT spectrum, ignoring this important prior knowledge of physiological perception, resulting in models failing to achieve optimal perceptual denoising performance under the same computational load.

[0006] Therefore, there is an urgent need in this field for a lightweight hybrid dual-channel AI noise reduction technology solution that can balance high noise reduction performance (especially for complex non-steady-state noise), low computational complexity, low memory usage, low latency, and conforms to the characteristics of human auditory perception, so as to achieve real-time deployment and widespread application in embedded devices such as TWS earphones and smart speakers.

[0007] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0008] The purpose of this invention is to provide a lightweight hybrid dual-channel AI noise reduction method and system based on embedded systems. This aims to address the technical problems of traditional noise reduction methods' poor suppression of non-steady-state noise and the inability of deep learning methods to run in real-time on embedded devices due to their high computational cost. Through innovative algorithm architecture, model structure optimization, and hardware-level acceleration, high-fidelity, low-latency, and low-power real-time noise reduction is achieved.

[0009] To achieve the above objectives, the present invention provides the following solution: A lightweight, embedded hybrid dual-channel AI noise reduction method includes the following steps: Step S1: Acquire the first time-domain audio signal collected by the first microphone and the second time-domain audio signal collected by the second microphone; Step S2: Perform a short-time Fourier transform on the signal to convert it from the time domain to the frequency domain, and obtain the corresponding first complex spectrum and second complex spectrum, wherein the complex spectrum contains amplitude components and phase components; Step S3: After merging the first and second complex spectra, input them into the pre-trained lightweight hybrid dual-channel AI noise reduction model; Step S4: The lightweight hybrid dual-channel AI noise reduction model first decomposes the input complex spectrum into an equivalent rectangular bandwidth subband and extracts the perceptually relevant subband features; Step S5: Input the extracted ERB subband features into the grouped convolutional layer for spatial feature extraction. The grouped convolutional layer divides the input feature map into multiple groups in the channel dimension and performs convolution operations independently in each group. Step S6: Input the feature sequence output from the grouped convolutional layer into the temporal recurrent network layer to capture the long-term dependencies of the audio signal; Step S7: The temporal recurrent network layer outputs a noise mask estimate, whereby the noise mask is used to characterize the ratio of speech components to noise components in each time-frequency unit; Step S8: Multiply the estimated noise mask pointwise with the first complex spectrum and / or the second complex spectrum to suppress noise spectrum components and preserve speech spectrum components, thereby generating an enhanced complex spectrum; Step S9: Perform an inverse short-time Fourier transform (ISTFT) on the enhanced complex spectrum to convert it back into a time-domain signal, thereby obtaining the initial reconstructed time-domain speech; Step S10: Perform post-filtering on the reconstructed time-domain speech to further smooth the spectrum and suppress residual music noise, and output the final time-domain speech signal.

[0010] Optionally, the ERB subband decomposition in step S5, wherein the mapping relationship between the ERB scale and frequency f is defined by the following formula:

[0011] Where f is the frequency in Hz; according to the mapping relationship, the entire frequency band of the complex spectrum is uniformly divided into N nonlinear ERB sub-bands, and the value of N ranges from 16 to 32.

[0012] Optionally, the noise mask in step S8 is an ideal ratio mask or a complex ideal ratio mask; when an ideal ratio mask is used, its calculation formula is:

[0013] in, t Indicates the time frame index. f Indicates frequency index, S ( t , f ) and N ( t , f ) represent the complex spectra of clean speech and noise, respectively, and β is an adjustable scaling factor ranging from 0.5 to 1.0.

[0014] Optionally, the method further includes a model lightweight optimization step: performing INT8 quantization on the trained lightweight hybrid dual-channel AI denoising model to quantize the 32-bit floating-point weights and activation values ​​in the model into INT8 integers; and performing structured or unstructured pruning to remove redundant neuron connections; the model volume after quantization and pruning is reduced to less than 1 / 4 of the original volume.

[0015] Optionally, the INT8 quantization employs either a symmetric or asymmetric quantization scheme, wherein the conversion formula for symmetric quantization is:

[0016] in, r This is the original 32-bit floating-point real number. q This is the quantized 8-bit integer, where Δ is the quantization step size. clamp This function is used to restrict a numerical value to an 8-bit integer representation. This is for rounding up to the nearest integer.

[0017] Optionally, when the method is implemented on an embedded chip, the chip's built-in single instruction multiple data stream instruction set is used to accelerate vector operations in parallel; floating-point arithmetic units are used to perform fast calculations on floating-point data; and multiplication and accumulation instructions in the digital signal processor instruction set are called to optimize the core calculations of convolutional and fully connected layers.

[0018] Optionally, the noise reduction performance indicators achieved by the method are as follows: for steady-state noise, the noise reduction depth reaches 20dB to 40dB; for non-steady-state noise, the noise reduction depth reaches 10dB to 25dB; the signal-to-noise ratio of the final output time-domain speech signal is ≥40dB, and the total harmonic distortion is less than 5%; the end-to-end delay time of the entire signal processing link does not exceed 20 milliseconds; and the average MIPS required for the algorithm to run is less than 50.

[0019] A lightweight, embedded hybrid dual-channel AI noise reduction system includes: The audio acquisition module is configured to acquire dual-channel time-domain audio signals via a dual-microphone array; The traditional preprocessing module is configured to perform traditional noise reduction algorithms and short-time Fourier transforms on the acquired dual-channel signals and output a complex spectrum; The lightweight AI inference module is equipped with a lightweight hybrid dual-channel AI noise reduction model optimized by INT8 quantization and pruning. The model includes an input layer, an ERB subband decomposition layer, a grouped convolutional layer, a temporal recurrent network layer, and a mask estimation output layer, which are used to receive complex spectra and output noise masks. The frequency domain enhancement module is configured to multiply the output noise mask element-wise with the input complex spectrum to achieve noise suppression; The temporal reconstruction and post-processing module is configured to perform inverse short-time Fourier transform and temporal post-filtering, and output the final denoised speech signal. The hardware acceleration scheduling module is configured to call the SIMD, FPU and DSP instruction sets of the embedded chip to accelerate the matrix algorithm and vector operation in the lightweight AI inference module.

[0020] Optionally, the system is integrated into true wireless stereo headphones for uplink call noise reduction; or, the system is integrated into smart speakers or in-vehicle voice assistants for far-field voice pickup and enhancement; or, the system is integrated into video conferencing microphone arrays for noise suppression and voice clarity in multi-person conference scenarios.

[0021] An embedded device, comprising: At least one microphone is used to pick up sound signals; At least one processor, wherein the processor integrates an FPU and a DSP instruction set; The memory stores computer-executable instructions and a lightweight hybrid dual-channel AI noise reduction model that has been quantized by INT8 and pruned. When the instruction is executed by the processor, the device implements the steps of the above-described embedded lightweight hybrid dual-channel AI noise reduction method.

[0022] Compared with the prior art, the present invention has the following beneficial effects: By introducing a lightweight AI model, particularly combining grouped convolutions (spatial feature extraction) and temporal recurrent networks (temporal dependency modeling), it can effectively learn and predict the time-frequency patterns of complex non-steady-state noises such as human voice, traffic noise, and wind noise. Its noise reduction depth and speech fidelity far exceed those of traditional DSP methods. Furthermore, the introduction of ERB sub-band features allows the model to focus more on frequency regions sensitive to the human ear, improving subjective listening quality.

[0023] This invention achieves real-time operation on embedded devices with extremely limited computing power, memory, and power consumption through an innovative hybrid architecture (traditional DSP preprocessing + lightweight AI model), efficient model design (ERB subband dimensionality reduction + grouped convolution + small hidden layer GRU), and extreme engineering optimization. End-to-end latency is less than 20ms, MIPS is less than 50, and model size can be reduced to the KB level, meeting the most demanding engineering requirements of products such as TWS earphones.

[0024] The mask estimated by the AI ​​model can precisely preserve the harmonic structure and temporal envelope of the speech, avoiding the speech distortion and musical noise problems common in traditional methods. The post-filtering module further cleans up any residual artifacts, resulting in clear and natural output speech.

[0025] This invention provides clear and quantifiable performance indicators, including noise reduction depth, final signal-to-noise ratio, distortion, and delay for different noise types, providing strong support for product marketing and competitor comparison. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of the embedded lightweight hybrid dual-channel AI noise reduction method provided in an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of the structure of a lightweight hybrid dual-channel AI noise reduction system based on embedded systems provided in an embodiment of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] The purpose of this invention is to provide a lightweight hybrid dual-channel AI noise reduction technology solution that can balance high noise reduction performance with low computational complexity, low memory usage, low latency, and conforms to the characteristics of human auditory perception.

[0031] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] Example 1: This embodiment details a lightweight, hybrid dual-channel AI noise reduction method based on embedded systems. For example... Figure 1 As shown, Figure 1 This is a schematic diagram of an embedded lightweight hybrid dual-channel AI noise reduction method provided in an embodiment of the present invention. The method specifically includes the following steps S1 to S11.

[0033] Step S1: Acquisition of dual-channel audio signals.

[0034] First, audio is acquired using a dual-microphone array. Specifically, a first time-domain audio signal is acquired by the first microphone, and a second time-domain audio signal is acquired by the second microphone. The two signals have spatial correlation, which can be used for subsequent noise reduction processing. In practical product applications, such as true wireless stereo (TWS) earbuds, the first microphone is typically a call microphone (located near the mouth), and the second microphone is a feedforward noise-canceling microphone (located on the top of the outer side of the earbud). The two work together to obtain richer sound field information.

[0035] Step S2: Time domain to frequency domain conversion.

[0036] A short-time Fourier transform (STFT) is performed on the signal. Through framing, windowing, and Fourier transform, the time-domain waveform signal is transformed from the time domain to the complex frequency domain, yielding the corresponding first and second complex spectra. The complex spectra contain amplitude and phase components. Preserving phase information is crucial for subsequent high-quality speech reconstruction.

[0037] Step S3: Channel merging and model input.

[0038] The first and second complex spectra are concatenated or weighted and merged along the feature dimension (channel dimension) to form a multi-channel input feature map. This input feature map is then fed into a pre-trained and lightweight hybrid dual-channel AI denoising model that has undergone lightweight optimization.

[0039] Step S4: Perception-related ERB subband decomposition.

[0040] To address the issues of low efficiency in processing uniform spectrum and failure to consider human auditory perception characteristics in traditional CNNs, this invention introduces an equivalent rectangular bandwidth subband decomposition layer at the front end of the model.

[0041] The specific calculation formula is as follows: .

[0042] Based on this mapping relationship, the entire frequency band (e.g., 0-8kHz) of the complex spectrum input in step S4 is nonlinearly divided into N sub-bands, preferably N=24. The low-frequency band is densely divided (high resolution), while the high-frequency band is sparsely divided (low resolution). This operation not only extracts perceptually relevant sub-band features but also significantly reduces the frequency dimension of the feature map, thereby reducing the computational cost of subsequent network layers.

[0043] Step S5: Grouped convolutional spatial feature extraction.

[0044] To further reduce the number of model parameters and computational complexity, this invention does not employ conventional standard convolutions and depthwise separable convolutions, but instead uses grouped convolutional layers. The ERB subband features extracted in step S5 are divided into G groups along the channel dimension, and convolution operations are performed independently within each group. This approach severs the channel connections between different groups, forcing the model to learn a more compact feature representation, thereby significantly reducing MACs (multiply-accumulate operations) while maintaining noise reduction performance.

[0045] Step S6: Temporal modeling of the temporal recurrent network.

[0046] Speech signals are typical temporal signals, and the noise characteristics of the current frame are closely related to those of past frames. The feature sequences output by the grouped convolutional layers are rearranged and then input into the temporal recurrent network (RNN) layer. Preferably, gated recurrent units (GRUs), which have fewer parameters and faster computation, are used to capture the long-term dependencies of the audio signal, especially the dynamic evolution of noise.

[0047] Step S7: Noise mask estimation.

[0048] The temporal recurrent network layer is followed by a fully connected output layer to output a noise mask estimate. This noise mask is a matrix with the same size as the input spectrum, used to characterize the ratio of speech components to noise components in each time-frequency unit.

[0049] Step S8: Frequency domain enhancement.

[0050] The noise mask (gain value between 0 and 1) estimated in step S8 is multiplied point-by-point with the first complex spectrum and / or the second complex spectrum generated in step S3. For time-frequency units dominated by noise, the mask value approaches 0 to suppress noise spectral components; for time-frequency units dominated by speech, the mask value approaches 1 to preserve speech spectral components. After multiplication and weighting, the enhanced complex spectrum is generated.

[0051] Step S9: Temporal reconstruction.

[0052] Perform an inverse short-time Fourier transform (ISTFT) on the enhanced complex spectrum. Specifically, this includes: using the original noisy phase or the enhanced phase saved in step S3, reconstructing the complex spectrum by combining it with the enhanced amplitude spectrum, and then converting the frequency domain signal back to the time domain signal through an inverse Fourier transform and an overlap-addition method to obtain the initial reconstructed time-domain speech.

[0053] Step S10: Post-filtering processing.

[0054] Because the mask estimated by the AI ​​model may be discontinuous or inaccurate, slight "musical noise" or high-frequency spikes may sometimes remain in the reconstructed speech. This invention introduces a post-filtering module (such as DSP-based Wiener filtering post-processing or a simple spectral smoothing algorithm) in the final stage to perform spectral smoothing on the initial reconstructed time-domain speech, further suppressing residual musical noise, and finally outputting a high-definition, low-distortion time-domain speech signal.

[0055] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0056] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A lightweight hybrid dual-channel AI noise reduction method based on embedded systems, characterized in that, Includes the following steps: Step S1: Acquire the first time-domain audio signal collected by the first microphone and the second time-domain audio signal collected by the second microphone; Step S2: Perform a short-time Fourier transform on the signal to convert it from the time domain to the frequency domain, and obtain the corresponding first complex spectrum and second complex spectrum, wherein the complex spectrum contains amplitude components and phase components; Step S3: After merging the first and second complex spectra, input them into the pre-trained lightweight hybrid dual-channel AI noise reduction model; Step S4: The lightweight hybrid dual-channel AI noise reduction model first decomposes the input complex spectrum into an equivalent rectangular bandwidth subband and extracts the perceptually relevant subband features; Step S5: Input the extracted ERB subband features into the grouped convolutional layer for spatial feature extraction. The grouped convolutional layer divides the input feature map into multiple groups in the channel dimension and performs convolution operations independently in each group. Step S6: Input the feature sequence output from the grouped convolutional layer into the temporal recurrent network layer to capture the long-term dependencies of the audio signal; Step S7: The temporal recurrent network layer outputs a noise mask estimate, whereby the noise mask is used to characterize the ratio of speech components to noise components in each time-frequency unit; Step S8: Multiply the estimated noise mask pointwise with the first complex spectrum and / or the second complex spectrum to suppress noise spectrum components and preserve speech spectrum components, thereby generating an enhanced complex spectrum; Step S9: Perform an inverse short-time Fourier transform on the enhanced complex spectrum to convert it back into a time-domain signal, thereby obtaining the initial reconstructed time-domain speech. Step S10: Perform post-filtering on the reconstructed time-domain speech to further smooth the spectrum and suppress residual music noise, and output the final time-domain speech signal.

2. The lightweight hybrid dual-channel AI noise reduction method based on embedded systems according to claim 1, characterized in that, The ERB subband decomposition in step S5, where the mapping relationship between the ERB scale and frequency f is defined by the following formula: Where f is the frequency in Hz; according to the mapping relationship, the entire frequency band of the complex spectrum is uniformly divided into N nonlinear ERB sub-bands, and the value of N ranges from 16 to 32.

3. The lightweight hybrid dual-channel AI noise reduction method based on embedded systems according to claim 1, characterized in that, In step S8, the noise mask is either an ideal ratio mask or a complex ideal ratio mask; when an ideal ratio mask is used, its calculation formula is: in, t Indicates the time frame index. f Indicates frequency index, S ( t , f ) and N ( t , f ) represent the complex spectra of clean speech and noise, respectively, and β is an adjustable scaling factor ranging from 0.5 to 1.

0.

4. The lightweight hybrid dual-channel AI noise reduction method based on embedded systems according to claim 1, characterized in that, The method also includes a model lightweighting optimization step: performing INT8 quantization on the trained lightweight hybrid dual-channel AI denoising model to quantize the 32-bit floating-point weights and activation values ​​in the model into INT8 integers; and performing structured or unstructured pruning to remove redundant neuron connections; the model volume after quantization and pruning is reduced to less than 1 / 4 of the original volume.

5. The lightweight hybrid dual-channel AI noise reduction method based on embedded systems according to claim 4, characterized in that, The INT8 quantization employs either a symmetric or asymmetric quantization scheme, where the conversion formula for symmetric quantization is: in, r This is the original 32-bit floating-point real number. q This is the quantized 8-bit integer, where Δ is the quantization step size. clamp This function is used to restrict a numerical value to an 8-bit integer representation. This is for rounding up to the nearest integer.

6. The lightweight hybrid dual-channel AI noise reduction method based on embedded systems according to claim 1, characterized in that, When the method is implemented on an embedded chip, it utilizes the chip's built-in single instruction multiple data stream instruction set to accelerate vector operations in parallel; it utilizes the floating-point arithmetic unit to perform fast calculations on floating-point data; and it calls the multiplication and accumulation instructions in the digital signal processor instruction set to optimize the core calculations of the convolutional layer and the fully connected layer.

7. The lightweight hybrid dual-channel AI noise reduction method based on embedded systems according to claim 1, characterized in that, The noise reduction performance achieved by the method is as follows: for steady-state noise, the noise reduction depth reaches 20dB to 40dB; for non-steady-state noise, the noise reduction depth reaches 10dB to 25dB; the signal-to-noise ratio of the final output time-domain speech signal is ≥40dB, and the total harmonic distortion is less than 5%; the end-to-end delay time of the entire signal processing link does not exceed 20 milliseconds; and the average MIPS required for the algorithm to run is less than 50.

8. A lightweight hybrid dual-channel AI noise reduction system based on embedded systems, used to implement the method described in any one of claims 1 to 7, characterized in that, include: The audio acquisition module is configured to acquire dual-channel time-domain audio signals via a dual-microphone array; The traditional preprocessing module is configured to perform traditional noise reduction algorithms and short-time Fourier transforms on the acquired dual-channel signals and output a complex spectrum; The lightweight AI inference module is equipped with a lightweight hybrid dual-channel AI noise reduction model optimized by INT8 quantization and pruning. The model includes an input layer, an ERB subband decomposition layer, a grouped convolutional layer, a temporal recurrent network layer, and a mask estimation output layer, which are used to receive complex spectra and output noise masks. The frequency domain enhancement module is configured to multiply the output noise mask element-wise with the input complex spectrum to achieve noise suppression; The temporal reconstruction and post-processing module is configured to perform inverse short-time Fourier transform and temporal post-filtering, and output the final denoised speech signal. The hardware acceleration scheduling module is configured to call the SIMD, FPU and DSP instruction sets of the embedded chip to accelerate the matrix algorithm and vector operation in the lightweight AI inference module.

9. The lightweight hybrid dual-channel AI noise reduction system based on embedded systems according to claim 8, characterized in that, The system is integrated into true wireless stereo earphones for uplink call noise reduction; or, the system is integrated into smart speakers or in-vehicle voice assistants for far-field voice pickup and enhancement; or, the system is integrated into video conferencing microphone arrays for noise suppression and voice clarity in multi-person conference scenarios.

10. An embedded device, characterized in that, include: At least one microphone is used to pick up sound signals; At least one processor, wherein the processor integrates an FPU and a DSP instruction set; The memory stores computer-executable instructions and a lightweight hybrid dual-channel AI noise reduction model that has been quantized by INT8 and pruned. When the instruction is executed by the processor, the device implements the steps of the embedded lightweight hybrid dual-channel AI noise reduction method according to any one of claims 1 to 7.

Images