33 results about "Spatial noise" patented technology

The invention relates to the field of array signal processing, and discloses an underwater target azimuth estimation method based on spherical harmonic noise model sparse reconstruction, which comprises the following steps of: calculating array output data to obtain a sample covariance matrix, establishing a spherical harmonic noise model, expressing spatial noise power density by adopting a finite-order spherical harmonic function, determining an initial power spectrum, and estimating the azimuth of an underwater target according to the initial power spectrum. Initializing the spherical harmonic noise model to obtain a noise covariance matrix, determining an expected model covariance matrix according to the noise covariance matrix, performing peak search on the initial power spectrum to obtain an estimated orientation, and refitting based on the estimated orientation to update the noise covariance matrix and the expected model covariance matrix; the initial power spectrum is subjected to sparse reconstruction based on the expectation model covariance matrix and the sample covariance matrix before updating to obtain an updated power spectrum, when the difference between the updated power spectrum and the initial power spectrum meets the convergence condition, the peak value of the updated power spectrum is searched to determine the target orientation, and super-resolution and robust orientation estimation is achieved.

The spatial noise sub-path is used for carrying out adaptive filtering on an image block through frequency domain disturbance and texture statistics so as to effectively suppress tunnel background noise; the motion blur repairing sub-path is used for completing blur kernel estimation of the image block through feature residual reconstruction and repairing vehicle motion blur; the brightness re-adjustment sub-path is used for repairing the uneven brightness phenomenon of the image block through a piecewise gamma function; the contrast enhancement sub-path is used for enhancing the marginal definition of the image block through nonlinear reconstruction and improving the quality of the image in the tunnel weak light environment; besides, an enhanced perception module is integrated in the detection model, the enhanced perception module comprises a structure guiding normalization unit, a local brightness deviation calibration unit, a multi-head space attention unit and an edge semantic fusion unit, and the multi-head space attention unit comprises a multi-receptive-field branch subunit and a cross channel fusion subunit. The precision of the detection model is improved, and the precision of detecting the vehicle position and the vehicle appearance is further improved.

This invention discloses a lightweight watermarking method for robust image attribution based on latent space diffusion models. The method uses low-dimensional feature vectors as the watermark carrier: first, a feature watermark library is constructed; then, a feature-noise encoder and a noise-feature decoder are trained. The encoder maps the feature watermark to initial latent space noise that satisfies a standard Gaussian distribution and inputs it into the latent space diffusion model to generate an image. When attribution verification is required, the image to be verified is inverted to recover the initial latent space noise, and the corresponding feature watermark is reconstructed by the decoder. The identity is determined using the cosine similarity between the reconstructed watermark and the original watermark in the feature watermark library. This invention maintains the quality of the generated image while exhibiting strong robustness to common distortions such as JPEG compression, cropping, random discarding, blurring, and brightness variations, and significantly reduces the overhead of watermark storage and matching computation.

This application discloses a method and system for estimating the direction of arrival (DOA) of weak targets based on Riemannian manifold background suppression and adaptive sparse Bayesian learning. The method includes: acquiring time-series signals using a sensor array to construct a series of sample covariance matrices, mapping them to a point sequence on a Hermitian positive definite matrix manifold space; iteratively calculating the background interference covariance matrix using the non-Euclidean geometric properties and logarithmic shielding effect of the Riemannian metric; mapping the background interference covariance matrix back to Euclidean space, adaptively performing background subtraction based on an energy decision mechanism to reconstruct a positive definite covariance matrix to be measured; inputting the covariance matrix to be measured into a sparse Bayesian learning framework, first iteratively recovering the signal power through adaptive mesh refinement sparse Bayesian learning, then performing a closed-loop iteration of subspace noise cleaning while keeping the mesh fixed to recover the sparse spatial spectrum of the target signal; and finally, using local analytical interpolation techniques to eliminate mesh quantization errors and calculate the precise DOA of the target.

The present application relates to the technical field of electron-optical imaging, and particularly relates to a backscattered electron scintillator detector device and a detection method, which comprises a backscattered electron receiving sensor, an optical condensing group, an optical transmission group, a photoelectric conversion system and a detector protection mechanism. By using a scintillator as a sensor, a backscattered electron signal generated by an industrial electron gun electron beam is received, and then the electronic information carried by the backscattered electron is converted into an optical signal and transmitted to an area far away from the workpiece. Finally, a high signal-to-noise ratio voltage analog signal is formed by the photoelectric conversion system, avoiding the influence of spatial noise and other interference signals on weak signal transmission in the existing analog signal transmission process. For detailed topographic features, a backscattered electron image with high resolution can be displayed.

The application discloses a rPPG signal extraction network based on a CBAM attention mechanism and a bidirectional Mamba, and a non-contact physiological monitoring method and device, wherein spatial noise is filtered by introducing a CBAM attention mechanism, long time sequence dependence is modeled at low cost by using a Mamba architecture, and signal purity is improved by combining a frequency domain and a confidence gate. The rPPG signal extraction network effectively extracts weak skin color change features through an adaptive differential fusion mechanism of the fusion backbone network, realizes long time sequence global modeling through a Mamba architecture (with linear calculation complexity) in the space-time encoder, performs double cleaning on the features in the spatial domain and the frequency domain through a confidence perception gate module and a frequency domain gate module, and finally reconstructs a high-fidelity rPPG waveform through the bidirectional Mamba regression head.

The utility model relates to the technical field of noise acquisition, and discloses a near-field noise acquisition cover which comprises an acquisition cover and an expansion cover, the acquisition cover is of a conical cylinder structure, the radius of the top end of the expansion cover is smaller than that of the tail end, and the tail end of the acquisition cover is connected with the top end of the expansion cover to form a large conical cylinder structure. The outer sides of the collection cover and the expansion cover are sleeved with a first sound insulation layer and a second sound insulation layer respectively, the edges of the first sound insulation layer and the second sound insulation layer extend to the outer sides of the collection cover and the expansion cover respectively, and the first sound insulation layer covers the outer side of the top of the expansion cover and makes contact with the outer wall of the expansion cover. Compared with the prior art, the noise collecting cover is composed of the collecting cover body and the expanding cover body, the noise collecting distance can be prolonged through the expanding cover body, noise in a narrow space can be collected only through the collecting cover body, the structure can be adjusted according to use scenes, and practicability is higher.

A noise-resistant integrated communication ranging and phase measurement method comprises the following steps: generating two squared signals from the reference and measurement electrical signals of the heterodyne interferometric communication ranging integrated optical system; mixing the two squared signals with sine and cosine signals generated by an FPGA to obtain four sine and cosine mixed signals; calculating the phase of the difference frequency signal between the two complex exponential signals generated by the mixed signals using a full-phase FFT algorithm; and subtracting the phases of the two difference frequency signals to complete the phase measurement. This invention solves the problems of excessive logic resource consumption and ineffective suppression of spatial noise in FPGAs. It can suppress the severe interference of communication signal modulation on phase measurement in displacement measurement while reducing the use of logic resources in the FPGA, achieving a 2×10⁻⁶ phase ratio at a 50dB signal-to-noise ratio. ‑ 4 Phase measurement accuracy in rad.

The application provides a spatial noise annoyance modeling method based on binaural psychoacoustic parameter difference, and the method uses the binaural psychoacoustic parameter difference as a bridge, establishes a spatial noise annoyance model based on noise self annoyance by using multivariate linear regression, and realizes correction of noise self annoyance in the sound source direction, so that the annoyance of noise in different directions is more reasonably and accurately predicted, the use range of the noise annoyance model is widened, the physical mechanism of the spatial noise annoyance is explained, and the application value of the noise annoyance model is enhanced.

The invention belongs to the technical field of radio frequency fingerprint identification, and provides a radio frequency fingerprint identification method and system based on a feature latent space diffusion model in order to solve the problem that the identification rate of an existing method is reduced in a low signal-to-noise ratio environment. The method comprises the following steps: after receiving a radio frequency signal, firstly carrying out classification pre-training on a system, and updating parameters of a feature extraction module and a classification module; then, carrying out submerged space diffusion pre-training, training a noise time step prediction module and a submerged space diffusion noise prediction module based on a feature extraction and classification module, and optimizing parameters in combination with feature reconstruction, time step prediction and classification loss; and finally, systematic fine adjustment is carried out, and all modules are integrally optimized. And repeatedly training until the neural network converges, and outputting an optimal network parameter for testing. According to the method, the robustness and the accuracy of radio frequency equipment identification under the condition of low signal-to-noise ratio are effectively improved by combining feature extraction and submerged space noise suppression.

Embodiments are disclosed for spatial noise filling in multi-channel codecs. In an embodiment, a method of regenerating background noise ambience in a multi-channel codec by generating spatial hole filling noise comprises: computing noise estimates based on a primary downmix channel generated from an input audio signal representing a spatial audio scene with background noise ambience; computing spectral shaping filter coefficients based on the noise estimates; spectrally shaping the multi-channel noise signal using the spectral shaping filter coefficients and a noise distribution, the spectral shaping resulting in a diffused, multi-channel noise signal with uncorrelated channels; spatially shaping the diffused, uncorrelated multi-channel noise signal with uncorrelated channels based on a noise ambience of the spatial audio scene; and adding the spatially and spectrally shaped multi-channel noise to a multi-channel codec output to synthesize the background noise ambience of the spatial audio scene.

This invention provides a system and method for estimating and suppressing spatial noise using machine learning support. [Solution] The noise estimation and suppression method includes the steps of: estimating the probability of speech and noise for each band of the input audio signal using a machine learning classifier; estimating a set of averages of speech and noise or a set of averages and covariances of speech and noise based on the probability and microphone covariance over the bands using a directional model; estimating the average and covariance of noise power based on the probability and power spectrum using a level model; determining a first noise suppression gain based on the directional model; determining a second noise suppression gain based on the level model; selecting the first noise suppression gain or the second noise suppression gain or their sum based on the signal-to-noise ratio of the input audio signal; and scaling the time-frequency representation of the input signal by the selected noise suppression gain.

The application discloses a kind of multi-modal information fusion based on pulse neural network emotion recognition method, belong to artificial intelligence, brain-like intelligence and biomedical signal processing technical field.The purpose of the present application is to introduce mutual mode dynamic threshold adjustment and collaborative coordinate attention mechanism, improve the accuracy, robustness and energy efficiency of multi-modal emotion recognition, realize the efficient alignment and adaptive deep fusion of heterogeneous physiological signals in pulse domain based on pulse neural network multi-modal information fusion emotion recognition method.The application respectively inputs each mode signal into the pulse encoder with the same structure and parameter-configurable to extract aligned features, inputs the mutual attention output of two directions after splicing into collaborative coordinate attention module to generate joint pulse gating map, and finally performs selective gating and fusion on enhanced features.The application effectively eliminates the inherent spatial noise and time-domain artifacts in physiological electrical signals, while ensuring the sparsity of pulse features, greatly improving the information density of the fused features.

The application discloses a kind of magnetic field inversion current methods based on near-field scanning, including steps: obtaining magnetic field data in circuit, and as input;The magnetic field is extended measurement plane, and the magnetic field matrix is obtained, so that the original magnetic field matrix H is continuous on the current of boundary;The relationship between magnetic field and current in frequency domain space is constructed;Using Tikhonov regularization method to reconstruct current distribution, and using GCV method to optimize the regularization parameter required by Tikhonov regularization method;Extract the structural information of static current model, use it as a mask, and apply it to the result of radio frequency current model;Output the current in the original measurement window as the current reconstructed by the final method.The application eliminates the current artifact on the boundary by extending the measurement plane, automatically solves the current distribution by regularization, avoids the error caused by empirical selection of filter parameters, improves the accuracy of reconstructed current, and finally suppresses spatial noise by mask method to realize accurate reconstruction of radio frequency current.

In observation of a distant object with a telescope or the like, it is required to remove noise in a visible light region such as atmospheric turbulence. Although a method using SPI has been proposed as a method for removing noise due to atmospheric turbulence, the resolution of observation in the visible region is low and atmospheric turbulence has not been removed, although remote monitoring is barely possible with the current SPI.SOLUTION: In order to solve the above problem, optical signals are acquired using a DMD and a plurality of photodetectors, and in a correlation process of imaging using the optical signals, optical correlation signals are acquired from selective wavelengths of the optical signals from the plurality of photodetectors, thereby simultaneously acquiring reconstructed images in a wide wavelength band. By mounting the DMD, the plurality of photodetectors, and the correlation process of this configuration on an imaging device or a magnifying optical device, it is possible to perform imaging instead of adaptive optics for suppressing spatial noise. As a light source, a reconstructed image is acquired for a wide range of wavelengths such as visible light, infrared light, and ultraviolet light.SELECTED DRAWING: Figure 1

The application relates to the field of array signal processing and discloses an underwater target direction estimation method based on spherical harmonic noise model sparse reconstruction. Sample covariance matrices are obtained by calculating array output data, a spherical harmonic noise model is established, spatial noise power density is represented by using limited order spherical harmonic functions, an initial power spectrum is determined, the spherical harmonic noise model is initialized to obtain a noise covariance matrix, and an expected model covariance matrix is determined based on the noise covariance matrix. A peak value of the initial power spectrum is searched to obtain an estimated direction. The noise covariance matrix and the expected model covariance matrix are updated based on the estimated direction. The initial power spectrum is sparsely reconstructed based on the expected model covariance matrix before updating and the sample covariance matrix to obtain an updated power spectrum. When the difference between the updated power spectrum and the initial power spectrum meets a convergence condition, a peak value of the updated power spectrum is searched to determine a target direction. The application realizes super-resolution and robust direction estimation.

The invention belongs to the technical field of distributed artificial intelligence and edge computing, and particularly relates to a model collaborative inference system and method based on dynamic sketch verification, and the system comprises a deployment bidirectional attention constraint generator, an integrated attention alignment loss calculation unit, a built-in confidence coefficient difference entropy analyzer, a dynamic threshold regulator and a difficult lexical unit recognizer. The system further comprises an input space noise injection unit, a feature space adversarial disturbance generator and a gradient projection matrix calculation unit. According to the method, through attention weight distribution comparison and loss constraint of the edge and the cloud model, the capturing capability of the edge model on long-range semantic association is remarkably enhanced, and the bottleneck of insufficient context focusing of a traditional lightweight model is overcome; based on a negative feedback adjustment mechanism of confidence coefficient difference entropy, cloud verification strictness is adaptively adjusted, the redundancy calculation overhead is effectively reduced while reasoning accuracy is guaranteed, and dynamic balance of efficiency and precision is achieved.

The invention relates to the technical field of artificial intelligence content generation, and discloses a long-time digital human video generation method and system, electronic equipment and a storage medium, and the method comprises the steps: obtaining an input voice and a reference figure image; encoding the input speech into a speech feature sequence; encoding the reference figure image to a hidden space to obtain a reference hidden space representation, generating a hidden space noise sequence corresponding to the target video duration, and replacing a first element of the hidden space noise sequence; performing multi-step denoising on the hidden space noise sequence through a video denoising module, performing feature interaction in combination with a segmented self-attention mechanism in each step of denoising, and fusing a voice feature sequence through a voice injection module to obtain clear hidden space representation; and decoding the clear hidden space representation to obtain a long-time digital human video with synchronous mouth shape and voice and consistent character image. According to the method, minute-level and even hour-level videos can be efficiently generated without complex preprocessing, multiple scenes such as virtual anchors and online education are adapted, and the generation effect and practicability are both considered.

In an embodiment, a method comprises: receiving bands of power spectra of an input audio signal and a microphone covariance, and for each band: estimating, using a classifier, respective probabilities of speech and noise; estimating, using a directionality model, a set of means for speech and noise, or a set of means and covariances for speech and noise, based on the microphone covariance for the band and the probabilities; estimating, using a level model, a mean and covariance of noise power based on the probabilities and the power spectra; determining a first noise suppression gain based on the directionality model; determining a second noise suppression gain based on the level model; selecting the first or second noise suppression gain or their sum based on a signal-to-noise ratio of the input audio signal; and scaling a time-frequency representation of the input signal by the selected noise suppression gain.

In an embodiment, a method comprises: receiving bands of power spectra of an input audio signal and a microphone covariance, and for each band: estimating, using a classifier, respective probabilities of speech and noise; estimating, using a directionality model, a set of means for speech and noise, or a set of means and covariances for speech and noise, based on the microphone covariance for the band and the probabilities; estimating, using a level model, a mean and covariance of noise power based on the probabilities and the power spectra; determining a first noise suppression gain based on the directionality model; determining a second noise suppression gain based on the level model; selecting the first or second noise suppression gain or their sum based on a signal-to-noise ratio of the input audio signal; and scaling a time-frequency representation of the input signal by the selected noise suppression gain.

A kind of method for generating crop and weed data in farmland based on frequency domain supervision conditional diffusion model, comprising: generating structured text description of image using visual large model;Condition diffusion model is constructed and frequency domain supervision is trained, that is, two-dimensional fast Fourier transform is respectively executed on foreground and background area after decoding to RGB space, the mean square error of frequency domain real part and imaginary part is calculated, and the model weight is updated by weighted summation with hidden space noise prediction loss;According to the category and quantity specified by user, the spatial control chart of self-defined is synthesized;Text description and spatial control chart are input into model to generate image, and high-quality data set is obtained after triple intelligent screening.The texture reality and structure fidelity of generated image are improved by foreground-background regional frequency domain supervision, and the cost of agricultural data acquisition and labeling is reduced, which can provide high-reliability training data for intelligent weeding, crop monitoring and the like.

This invention provides a method for designing a causally optimal broadband acoustic metamaterial sound absorber and introduces a novel method for achieving causally optimal broadband absorption (COBA) by designing and manufacturing a metamaterial sound absorber. The method involves three key steps: (1) calculating the COBA spectrum by solving an optimization problem based on causal constraints; (2) designing a resonator array such that the mode density and resonance intensity match the COBA spectrum; and (3) manufacturing the sound absorber using a large-scale production method such as a mold molding process. The resulting acoustic metamaterial sound absorber exhibits superior broadband absorption performance compared to conventional porous sound absorbers, particularly in the low-frequency range, within a given thickness limitation. This invention represents a significant advance in the development of high-performance and space-saving noise control solutions.

Provided is a vehicle 4D millimeter-wave radar inertial odometry method and a computer-readable medium. The method includes: eliminating significant radar spatial noise through relaxation filtering; performing single-frame velocity measurement based on a selecting weight iteration-based robust estimation method, and eliminating dynamic noise; performing IMU observation integration by mechanical arrangement to obtain a predicted vehicle state; constructing velocity constraints based on single-frame velocity measurement and a predicted velocity to initially update the vehicle state; constructing a local map based on the previous point cloud frame, searching n-nearest neighbor points of each radar point in the local map at next moment, calculating a weighted distance from distribution to multi-distribution as matching constraints from point cloud to local map, and performing further updating based on iterative filtering to obtain a precise vehicle state; and updating the vehicle state at all moments by 4Dradar closed-loop detection and GICP to reduce the positioning drift.

The invention provides a frequency offset residual error and phase noise joint compensation method for a lightning stroke scene, and relates to the technical field of signal processing, and the method comprises the steps: obtaining two-dimensional time domain data corresponding to an output balance baseband I / Q of a coherent optical receiver, and sequentially carrying out the phase observation construction and frequency preprocessing of the two-dimensional time domain data, so as to obtain an observation vector sequence; based on the observation vector sequence, the weight information of each observation sample and a preset subcarrier set, constructing a weighted sample covariance matrix, performing characteristic decomposition on the weighted sample covariance matrix, obtaining information subspace, noise subspace and phase jump characteristics caused by lightning stroke, constructing a joint cost function, and calculating the phase jump characteristics of the information subspace, the noise subspace and the lightning stroke. Obtaining residual frequency offset and random phase noise at each sampling frequency point according to the optimal solution of the joint cost function; and performing frequency and phase joint compensation on the received coherent optical baseband signal according to the residual frequency offset and the random phase noise. According to the invention, suppression of inter-carrier interference caused by frequency offset and phase noise is facilitated.

The disclosure provides a method and an electronic device for inverting surrounding rock parameters of a fissure evolution underground cavern group, and relates to the technical field of engineering data processing. The method comprises the following steps: inputting first condition information and first latent space noise of the underground cavern group into a generator to output first parameter evolution information; the first parameter evolution information comprises first mechanical parameter evolution information and first fissure evolution information; calculating the time sequence displacement of the surrounding rock according to the first mechanical parameter evolution information to obtain first displacement prediction data; inputting the first displacement prediction data, first displacement observation data in the first time sequence and the first parameter evolution information into a discriminator to output comprehensive discrimination information; determining a comprehensive loss function value based on the comprehensive discrimination information to update the generator; and inputting second condition information and second latent space noise into the updated generator to output second parameter evolution information. The disclosure improves the accuracy of the inversion of the surrounding rock parameters of the underground cavern group and can reflect the dynamic evolution process of the surrounding rock.

According to one embodiment, the electronic device may include a camera including an image sensor; at least one processor including a processing circuit; and a memory for storing instructions. According to one embodiment, when the instructions are executed individually or collectively by the at least one processor, the electronic device may be configured to acquire a first image frame having a first noise level from the camera, identify a second image frame having a second noise level preceding the first image frame, reduce the spatial noise level of a motion area of ​​the first image frame based on motion information related to the difference between the first image frame and the second image frame, acquire a third image frame using the motion area of ​​the first image frame where the spatial noise level of the first image frame has been reduced and at least a portion of the first image frame, and acquire a fourth image frame having a third noise level lower than the first noise level based on at least a portion of the third image frame and at least a portion of the second image frame.

The application discloses a medical semantic image synthesis method and system based on a graph neural network, and comprises the following steps: S1, acquiring a continuous medical image semantic mask slice sequence, constructing a spatial adjacency graph by using adjacent slices in the slice sequence, and generating an adjacency matrix; S2, aggregating multi-scale context features of adjacent slices by using a graph neural network, and generating a slice feature map fusing a spatial dependency relationship; S3, injecting a three-dimensional spatial noise tensor aligned with a physical coordinate of volume data, eliminating discontinuity between slices by three-dimensional smoothing, and generating a fused feature map; S4, inputting the fused feature map into a deconvolution decoder modulated by spatial self-adaptive normalization, and generating a synthesized medical image sequence; and S5, discriminating the synthesized medical image sequence by using a gray scale-texture joint classifier; the application generates a synthesized medical image sequence by constructing a three-dimensional graph neural network fusing spatial context and injecting three-dimensional spatial noise modulated by a residual and smoothed.

The invention discloses a space active noise control and sound field optimization method for privacy office furniture, and the method comprises the steps: separating a mixed acoustic signal through employing a collection microphone array, and obtaining an environment noise reference signal and a user voice target signal; based on the environment noise reference signal, introducing a perception weighting adaptive filtering algorithm of a psychological acoustic model, and generating a first group of control signals for offsetting the environment noise; based on a user voice target signal, solving an optimization problem with multiple constraints, and generating a second group of control signals for constructing a light and dark privacy sound field; based on a collaborative strategy of model prediction control, a dynamic prediction model is established, the weight of noise reduction and privacy protection in a cost function is adjusted according to a voice activity detection result, an optimal control sequence of hardware physical constraint is solved, and a secondary sound source array is uniformly driven. According to the scheme of the invention, through an integrated cooperative control framework, the optimal balance between noise reduction and privacy is realized, and the subjective noise reduction experience and the robustness of the system in a dynamic environment are improved.

One or more embodiments of the invention provide an in-vehicle noise determination method and device, a storage medium and a program product. The method comprises the following steps: acquiring a background noise signal associated with a target seat in a vehicle; under the condition that a target seat of the vehicle is provided with active noise reduction equipment, noise reduction effect distribution data corresponding to the target seat are obtained, and the noise reduction effect distribution data represent the noise reduction effects of the active noise reduction equipment on the background noise signals in multiple seat areas of the target seat; according to the background noise signal and the noise reduction effect distribution data, an estimated noise signal used for estimating noise sensed by a passenger of the target seat is determined, and then the real noise level sensed by the ears of the passenger in the actual working state of the active noise reduction system can be accurately estimated. The defect of noise estimation misalignment caused by spatial noise reduction difference due to neglecting of an active noise reduction system in the prior art is effectively overcome.

The invention provides a microphone array voice noise reduction method and related equipment, and relates to the technical field of voice noise reduction. According to the invention, after multiple paths of to-be-denoised voice signals collected by a target microphone array are converted into multiple paths of original frequency domain signals through signal framing processing and short-time Fourier transform processing, a voice denoising model is called to carry out preliminary denoising processing on the multiple paths of original frequency domain signals; according to a spatial noise suppression matrix of a target microphone array for multi-band sound, amplitude spectrum self-adaptive correction is carried out on multiple paths of preliminary noise reduction frequency domain signals obtained through preliminary noise reduction processing; and then, through short-time inverse Fourier transform processing and signal frame combination processing, multiple paths of target frequency domain signals obtained through amplitude spectrum correction processing are converted into multiple paths of target voice signals, so that voice noise reduction is realized through cooperation between a voice noise reduction mechanism based on a deep learning technology and an audio noise differential suppression mechanism based on microphone spatial position driving. And the noise reduction effect consistency of multiple paths of voice signals is improved.