Dual microphone directional noise reduction method, device, and medium based on empirical coherence template

Abstract

The application discloses a double-microphone directional noise reduction method based on an experience coherence template, relates to the technical field of audio processing, and is applied to intelligent wearable devices. The method comprises the following steps: a target space trajectory template and an experience coherence template are set in advance; spatial direction features and spatial coherence features are extracted according to a current frame signal of a target audio signal; a first gain is calculated according to the spatial direction features and the target space trajectory template; a second gain is calculated according to the spatial coherence features and the experience coherence template; the first gain and the second gain are fused, the current frame signal of the target audio signal is subjected to noise reduction processing, and a signal subjected to noise reduction is output. In the application, the spatial coherence characteristics in a real environment are reserved in the experience coherence template, the negative number in the root sign is no longer obtained after substitution, the equation has a real number solution, and a CDR value can be calculated.

Description

Technical Field

[0001] This application relates to the field of audio processing technology, and in particular to a dual-microphone directional noise reduction method, apparatus and medium based on empirical coherence templates. Background Technology

[0002] In the field of audio processing technology, traditional beamforming is based on the theoretical geometric guide vector of dual microphones to form spatial directivity. Deep learning (Neural Network, NN) extracts spatial and frequency domain features through neural networks and performs nonlinear mapping. Traditional blind coherence-to-diffuse ratio (CDR) estimation calculates the noise reduction gain by calculating the complex spatial coherence between the audio signals collected by dual microphones and combining it with the theoretical formula of an ideal spherical diffuse sound field.

[0003] The dual microphones of smart glasses are mounted on the sides of the temples. The wearer's head acts as an acoustic barrier, causing sound waves to diffract. Simultaneously, hardware manufacturing tolerances and microphone orientation deviations result in an asymmetric sound field. Traditional beamforming, sensitive to hardware tolerances, suffers from a sharp amplification of low-frequency white noise due to discrepancies between the theoretical geometric guide vector and the actual acoustic transmission path. Deep learning, with its massive parameter count, faces computational bottlenecks and system latency on low-power edge chips. The structural difference between the actual sound field and the ideal diffused sound field prevents theoretical formulas from accurately describing the spatial coherence characteristics of the real environment. Substituting this into the CDR estimation equation yields a negative value within the square root, causing the algorithm to fail to output effective noise reduction gain and crash. Summary of the Invention

[0004] This invention provides a dual-microphone directional noise reduction method, device, and medium based on empirical coherent templates, aiming to solve the technical problem of failure of traditional blind CDR estimation under asymmetric sound field conditions.

[0005] To achieve the aforementioned objectives, the first aspect of this invention proposes a dual-microphone directional noise reduction method based on an empirical coherence template, comprising: pre-setting a target spatial trajectory template and an empirical coherence template, wherein the target spatial trajectory template is used to characterize the trajectory of the phase difference characteristics between channels as a function of frequency, and the empirical coherence template is used to characterize the coherence characteristics of a diffuse sound field; acquiring the current frame signal of a target audio signal; extracting spatial direction features and spatial coherence features based on the current frame signal of the target audio signal; calculating a first gain based on the spatial direction features and the target spatial trajectory template; calculating a second gain based on the spatial coherence features and the empirical coherence template; fusing the first gain and the second gain to perform noise reduction processing on the current frame signal of the target audio signal, and outputting a noise-reduced signal.

[0006] Further, the pre-set target spatial trajectory template includes: acquiring historical signals of the target sound source in the target direction; performing time-frequency transformation on the historical signals to obtain time-frequency domain representations of the historical signals at each frequency point; calculating sinusoidal characteristic values ​​of the inter-channel phase differences of each frame of the historical signals at each frequency point based on the time-frequency domain representations; constructing a two-dimensional statistical distribution based on the sinusoidal characteristic values ​​of the inter-channel phase differences of each frame at each frequency point; extracting the statistical center values ​​of the two-dimensional statistical distribution at each frequency point along the frequency axis; and combining the statistical center values ​​of each frequency point in frequency order to obtain the target spatial trajectory template.

[0007] Furthermore, the pre-set empirical coherence template includes: acquiring the dual-microphone historical signal under an ambient noise scene; calculating the auto-power spectrum and cross-power spectrum of the dual-microphone historical signal; calculating the first complex spatial coherence based on the auto-power spectrum and cross-power spectrum of the dual-microphone historical signal; and extracting the real part of the first complex spatial coherence as the empirical coherence template.

[0008] Furthermore, the acquisition of the current frame signal of the target audio signal includes: dividing the time-domain signal input from the dual microphones into frames with a preset frame length; using a weighted overlapping addition method and a square root Hanning window, converting the time-domain signal of each frame into a complex frequency-domain signal through Fourier transform to obtain the current frame signal of the target audio signal.

[0009] Further, the step of extracting spatial direction features and spatial coherence features based on the current frame signal of the target audio signal includes: using an exponential moving average algorithm, combined with the power spectrum state of historical frames, to smoothly update the auto-power spectrum and cross-power spectrum of the current frame signal; calculating the sinusoidal characteristic value of the inter-channel phase difference of the current frame signal as the spatial direction feature based on the smoothly updated auto-power spectrum and cross-power spectrum; and calculating the second complex spatial coherence of the current frame signal as the spatial coherence feature based on the smoothly updated auto-power spectrum and cross-power spectrum.

[0010] Further, the step of calculating the first gain based on the spatial direction feature and the target spatial trajectory template includes: calculating the absolute error distance between the spatial direction feature and the target spatial trajectory template at each frequency point; mapping the absolute error distance to a first gain using an exponential decay function based on the absolute error distance, wherein the first gain decreases as the absolute error distance increases.

[0011] Further, the step of calculating the second gain based on the spatial coherence feature and the empirical coherence template includes: obtaining the real part of the spatial coherence feature and the square of the amplitude of the spatial coherence feature; calculating the coherent diffusion energy ratio based on the real part of the spatial coherence feature, the square of the amplitude of the spatial coherence feature, and the empirical coherence template; and mapping the coherent diffusion energy ratio to a second gain using the Wiener filtering formula based on the coherent diffusion energy ratio, wherein the second gain increases as the coherent diffusion energy ratio increases.

[0012] Further, the step of fusing the first gain and the second gain to perform noise reduction processing on the current frame signal of the target audio signal and outputting the noise-reduced signal includes: multiplying the first gain and the second gain at various frequency points in the frequency domain to obtain the target mask gain; averaging the amplitude spectrum of the current frame signal of the target audio signal and extracting the phase spectrum of a preset reference channel from the current frame signal of the target audio signal, and synthesizing a basis complex spectrum based on the averaged amplitude spectrum and the phase spectrum of the preset reference channel; multiplying the target mask gain by the basis complex spectrum at each frequency point to obtain the noise-reduced frequency domain signal; and performing inverse Fourier transform and weighted overlapping addition synthesis on the noise-reduced frequency domain signal to output a time-domain noise-reduced signal.

[0013] A second aspect of the present invention provides a dual-microphone directional noise reduction device, comprising: a microphone for acquiring target audio signals; a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described dual-microphone directional noise reduction method based on empirical coherence templates.

[0014] Thirdly, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described dual-microphone directional noise reduction method based on empirical coherence templates.

[0015] The dual-microphone directional noise reduction method based on empirical coherence template provided in this application pre-sets a target spatial trajectory template and an empirical coherence template. The target spatial trajectory template is extracted from the historical signals of the actual wearer and is used to replace the theoretical steering vector that depends on hardware precision. Microphone position errors and wearing offsets are absorbed in the target spatial trajectory template. The empirical coherence template is the real part of the complex spatial coherence in the environmental noise scene and is used to replace the theoretical formula based on the assumption of an ideal spherical diffuse sound field. The spatial coherence characteristics in the real environment are retained in the empirical coherence template.

[0016] Based on this, spatial direction features and spatial coherence features are extracted from the current frame signal of the target audio signal. The spatial direction features are compared with the target spatial trajectory template to obtain a first gain, and the spatial coherence features are compared with the empirical coherence template to obtain a second gain. The first gain and the second gain are fused to perform noise reduction processing on the current frame signal. The noise reduction process needs to satisfy two conditions simultaneously: direction matching and coherence matching. Direction matching ensures that the sound source comes from the target direction, and coherence matching ensures that the sound source has high coherence features of near-field clean speech. Thus, the wearer's voice is jointly screened from the two dimensions of direction and distance. Attached Figure Description

[0017] Figure 1 This is a flowchart of the dual-microphone directional noise reduction method based on empirical coherence templates in the embodiments of this application; Figure 2 This is a frame-level feature density distribution map of the first inter-channel phase difference in the embodiments of this application; Figure 3 This is a frame-level feature density distribution diagram of the second inter-channel phase difference in this embodiment of the application; Figure 4 This is a frame-level feature density distribution map of the third inter-channel phase difference in this embodiment of the application; Figure 5 This is a structural block diagram of the dual-microphone directional noise reduction device in the embodiments of this application; Figure 6 This is a structural block diagram of the computer device in the embodiments of this application.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, modules, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any modules and all combinations of one or more associated listed items.

[0021] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0022] Reference Figure 1 This invention provides a dual-microphone directional noise reduction method based on empirical coherence templates, comprising the following steps S1-S6: S1: Pre-set the target space trajectory template and the empirical coherence template. The target space trajectory template is used to characterize the trajectory of the phase difference characteristics between channels as a function of frequency, and the empirical coherence template is used to characterize the coherence characteristics of the diffuse sound field.

[0023] During the offline calibration phase, a target spatial trajectory template and an empirical coherence template are pre-set. The target spatial trajectory template is used to characterize the trajectory of the inter-channel phase difference characteristics as a function of frequency. The establishment of the target spatial trajectory template utilizes the physical characteristics of the near-field sound source. When the sound source is close to the microphone, the direct sound dominates, and the phase difference is stably locked on a fixed geometric path. The sine value of the inter-channel phase difference (IPD) is presented as a bright and thin rising curve on the statistical density map. The background area outside the curve has almost no frames falling in. The target spatial trajectory template represents the wearer's unique directional characteristics.

[0024] Empirical coherence templates are used to characterize the coherence features of diffuse sound fields. The establishment of empirical coherence templates stems from the fact that the form of smart glasses worn close to the ear disrupts the symmetry of the sound field. Traditional theoretical formulas cannot be applied under this condition. Empirical coherence templates directly reflect the coherence features of diffused noise in the actual wearing state of the current device.

[0025] In other embodiments, the technical solution of the present invention can also be applied to other smart wearable devices. Taking TWS earphones as an example, TWS earphones are also equipped with dual microphones. The two microphones are usually distributed at the top and bottom ends of an earphone stem or on the left and right earphones, with a very small physical distance between them. In actual use, TWS earphones also face the problems of hardware manufacturing tolerances and inconsistent microphone orientation. The wearing angle also has individual deviations due to differences in the shape of the user's ear. These factors can also lead to a deviation between the theoretical guiding vector of traditional beamforming and the actual acoustic transmission path. Therefore, the empirical template calibration and dual-branch fusion architecture of the present invention are also applicable to TWS earphones.

[0026] Unlike smart glasses, TWS earbuds are worn at the ear canal opening, resulting in differences in head obstruction position and sound wave diffraction paths. When applied to TWS earbuds, the template calibration stage requires re-recording the wearer's clean voice and ambient noise under actual wearing conditions to extract the device's unique target spatial trajectory template and empirical coherence template. Subsequent online processing is the same as in the smart glasses implementation.

[0027] S2: Obtain the current frame signal of the target audio signal.

[0028] If it is in the online real-time processing stage, the current frame signal of the target audio signal is acquired in real time. The current frame signal of the target audio signal refers to the frequency domain representation of the time domain signal collected by the two microphones in the same time period. The current frame signal contains the amplitude and phase information of all sound sources in the frequency domain in the same time period. The current frame signal is the input for all subsequent feature extraction and gain calculation.

[0029] S3: Extract spatial orientation features and spatial coherence features based on the current frame signal of the target audio signal.

[0030] Spatial orientation features are features extracted from the current frame signal that are related to the direction of the sound source. These features are compared with a target spatial trajectory template to determine the sound source's location. Specifically, the spatial orientation feature is the sinusoidal characteristic value of the inter-channel phase difference (IPD) of the current frame. Spatial coherence features are features extracted from the current frame signal that are related to the degree of correlation between the two channels. These features are compared with an empirical coherence template to determine the diffuse sound field characteristics of the sound source. Specifically, the spatial coherence feature is the complex spatial coherence of the current frame.

[0031] S4: Calculate the first gain based on the spatial orientation characteristics and the target spatial trajectory template.

[0032] The sinusoidal characteristic value of the inter-channel phase difference (IPD) of the current frame is compared with the target spatial trajectory template. The degree of matching between the two values ​​reflects how close the current sound source orientation is to the direction directly in front. The higher the degree of matching, the closer the sound source orientation is to the direction directly in front, and the larger the first gain value is. The lower the degree of matching, the more the sound source orientation deviates from the direction directly in front, and the smaller the first gain value is.

[0033] S5: Calculate the second gain based on spatial coherence characteristics and empirical coherence template.

[0034] The complex spatial coherence of the current frame is compared with the empirical coherence template. The degree of matching between the two values ​​reflects the degree to which the current sound source has diffuse sound field characteristics. The higher the degree of matching, the closer the current sound source is to diffuse sound field characteristics, and the smaller the value of the second gain. The lower the degree of matching, the more the current sound source deviates from diffuse sound field characteristics, and the larger the value of the second gain.

[0035] S6: Combines the first gain and the second gain to perform noise reduction on the current frame signal of the target audio signal and outputs the noise-reduced signal.

[0036] The first gain is determined based on the degree of matching between the directional features and the target spatial trajectory template; the more positive the direction, the greater the gain. The target spatial trajectory template comes from a statistical center of real recordings of the wearer. Factors such as microphone position error, wearing offset, and head diffraction have been absorbed by the template. During online processing, the directional features of the current frame are directly compared with the template, without relying on hardware geometry to calculate the theoretical steering vector, thus bypassing hardware tolerance issues.

[0037] The second gain is determined by how closely the coherence characteristics resemble the empirical coherence template; the closer to the template, the smaller the gain, and the further away from the template, the larger the gain. The prerequisite for the second gain is that the CDR equation has a real solution. Traditional CDR equations use a theoretical sinc function, whose theoretical value comes from the assumption of an ideal diffuse sound field. This doesn't match the real asymmetric sound field of smart glasses, resulting in a negative square root after substitution, and the equation has no real solution. This solution replaces the theoretical sinc function with an empirical coherence template. The empirical coherence template comes from the real coherence part of the real noise floor, belonging to the same physical space as the coherence characteristics of the current frame. After substitution, the square root is no longer negative, the equation has a real solution, the CDR value is calculated, and the second gain is generated accordingly.

[0038] "Blending" refers to combining the first and second gains. Only time frequencies that are correctly oriented, sufficiently close, and have minimal reverberation are preserved. Time frequencies that are correctly oriented but far away or incorrectly oriented are suppressed. Only when the sound direction is absolutely correct (directly in front) and extremely pure with no reverberation (extremely close to the face) will both gains approach 1, allowing the sound to be preserved.

[0039] The directional characteristics of the interfering sound directly in front match the target spatial trajectory template. The first gain is high, but far-field sounds carry multipath components from room reflections, resulting in low coherence and a low CDR value, leading to a low second gain. The directional characteristics of the wearer's speech match the target spatial trajectory template. The first gain is high, and the near-field direct sound has high coherence but deviates from the empirical coherence template, resulting in a high CDR value, leading to a high second gain. Multiplying the first and second gains results in the low total gain of the interfering sound being suppressed, while the high total gain of the wearer is preserved. Thus, the first and second gains are determined independently from two dimensions: direction and distance. The first gain ensures correct direction, while the second gain ensures sufficient distance and minimal reverberation. After merging the two gains, sounds with correct direction but far distance are suppressed, while sounds with correct direction and close distance are preserved. Collinear interference, which cannot be distinguished by relying solely on directional information, is resolved.

[0040] As shown above, the empirical coherence template is the real part of the complex spatial coherence in the environmental noise scenario, which is used to replace the theoretical formula based on the assumption of an ideal spherical diffuse sound field. The spatial coherence characteristics in the real environment are preserved in the empirical coherence template.

[0041] The embodiments of this application extract spatial direction features and spatial coherence features from the current frame signal of the target audio signal. The spatial direction features are compared with the target spatial trajectory template to obtain a first gain, and the spatial coherence features are compared with the empirical coherence template to obtain a second gain. The first gain and the second gain are fused to perform noise reduction processing on the current frame signal. The noise reduction process needs to meet two conditions at the same time: direction matching and coherence matching. Direction matching ensures that the sound source comes from the target direction, and coherence matching ensures that the sound source has high coherence features of near-field clean speech. Thus, the wearer's speech is screened from both direction and distance dimensions, effectively distinguishing collinear interference.

[0042] In one embodiment, step S1 above includes: Acquire historical signals of the target sound source in the target direction; Perform time-frequency transformation on the historical signal to obtain the time-frequency domain representation of the historical signal at each frequency point; Based on the time-frequency domain characterization, calculate the sinusoidal characteristic value of the inter-channel phase difference at each frequency point for each frame of the historical signal; A two-dimensional statistical distribution is constructed based on the sinusoidal characteristic value of the inter-channel phase difference at each frequency point in each frame; Along the frequency axis, extract the statistical center value of the two-dimensional statistical distribution at each frequency point; The statistical center values ​​of each frequency point are combined in frequency order to obtain the target space trajectory template.

[0043] Preferably, the target sound source is the wearer's own voice, and the target direction is directly in front of the wearer. The historical signal is the wearer's voice data collected by the dual microphones during the offline phase. In a factory environment, the wearer speaks normally, and the dual microphones simultaneously record clean voice data, covering a sufficient number of voice frames.

[0044] Time-frequency transformation is short-time fourier transform (STFT). The historical signal is divided into frames of a preset frame length, and a short-time fourier transform is performed on each frame to obtain complex frequency domain data. A frequency point is a discrete position on the frequency axis after the Fourier transform. The time-frequency domain representation is the complex data of each frame at each frequency point, containing both amplitude and phase components.

[0045] Each of the two microphones has a phase value at the same frequency. Subtracting the two phase values ​​yields the inter-channel phase difference (IPD). Sineing the difference gives a sinusoidal characteristic value. A sinusoidal characteristic value is calculated for each frequency point per frame. The two microphones in the smart glasses are approximately 1.5 cm apart. When the wearer speaks, their mouth acts as a near-field sound source, and the sound reaches the two microphones with a fixed optical path difference Δr. The theoretical phase difference formula for the two microphones at each frequency f is Δφ(f) = 2πfΔr / c, where c is the speed of sound. cos(IPD) remains around 1.0 in the low-frequency range, lacking discriminative power. sin(IPD) satisfies sin(x) ≈ x near zero, and increases approximately linearly with frequency from 0.0 at 0 Hz, continuously rising in the speech frequency range from 0 to 4000 Hz. Therefore, the sinusoidal characteristic value is extracted as a basis for direction determination.

[0046] Differences in facial shape, variations in wearing angle, and the diffraction effect of sound waves bypassing the cheek cause the actual Δr to become a non-linear quantity, making it impossible to measure precisely with calipers. Templates calculated directly using theoretical formulas do not match the actual equipment.

[0047] Two-dimensional statistical distribution is a two-dimensional histogram with frequency on the horizontal axis and the sinusoidal characteristic value of the phase difference between channels on the vertical axis, called a frame-level density map. The sinusoidal characteristic value of all speech frames at each frequency point is counted according to the coordinate position of (frequency, sinusoidal characteristic value). The more frames that fall at the same coordinate position, the higher the density at that position.

[0048] When the wearer speaks in the extremely close field, their mouth is pressed tightly against the microphone, resulting in a dominant direct sound energy and an extremely high direct reverberation ratio. The sinusoidal characteristic values ​​of each frame at each frequency point are highly concentrated near the values ​​corresponding to the actual physical optical path difference, presenting a bright and slender rising curve on the density map, with almost no frames falling into the background area outside the curve.

[0049] The frequency axis is the horizontal axis of a two-dimensional statistical distribution, arranged from low to high frequencies. Each position on the frequency axis corresponds to a frequency point. Step 5 means: starting from the lowest frequency point, perform the extraction operation sequentially at each frequency point position.

[0050] For the current frequency point being processed, the sin(IPD) values ​​of all frames at that frequency point are extracted, and these values ​​form a one-dimensional distribution. Under very near-field conditions, these values ​​are highly concentrated near the values ​​corresponding to the true optical path difference, forming a single-peak distribution. A few reflected sound wave components produce noise points that deviate from the main peak.

[0051] The central value of a statistical distribution can be either the median or the peak value of maximum density. The median is the value that is in the middle after all values ​​are sorted by size. Noise points that deviate from the main peak are few in number and far away, so they do not affect the value of the middle position after sorting. Therefore, the median can eliminate occasional reflection noise. The peak value of maximum density is the value with the highest count in a one-dimensional distribution.

[0052] Starting from the lowest frequency point, the statistical center values ​​corresponding to each frequency point are arranged in ascending order of frequency and pieced together to form a curve that varies with frequency, called the target space trajectory template. This curve is a smooth rising curve with hardware acoustic fingerprint, representing the stable trajectory of sin(IPD) as a function of frequency when the wearer's mouth is in the extremely near field position directly in front.

[0053] To obtain the spatial trajectory under real physical conditions, this invention utilizes frame-level density maps of the wearer's actual speech for calibration and extraction. Based on a data-driven extraction method using frame-level density maps, the method involves recording clean speech data of the wearer in a quiet environment, performing a Short-Time Fourier Transform (STFT), filtering out effective speech frames with high signal-to-noise ratios using an energy threshold, calculating the sinusoidal characteristic value sin(IPD) of the inter-channel phase difference (IPD) of each effective speech frame at each frequency point, and constructing a two-dimensional statistical histogram. After performing STFT on the clean speech recorded by the wearer, sin(IPD) of each frame is calculated and a two-dimensional frequency domain density map is plotted. Because the wearer's mouth is extremely close to the microphone, direct sound dominates, resulting in a very high direct reverberation ratio and highly concentrated energy, forming an extremely bright, low-variance, and largely clean upward curve with a large area of ​​pure blank space in the background region on the density map. In contrast, when the interferer speaks from a distance, even if the interferer happens to be directly in front of the wearer and the theoretical path difference is similar, the direct reverberation ratio is low due to room reverberation and multipath reflections. The rising trajectory in the density map becomes extremely dim, and the surrounding area is filled with numerous scattered stray phase noise points. Along the frequency axis, the median or maximum density peak of the sin(IPD) distribution is extracted at each frequency point as the statistical center value. The median effectively eliminates occasional reflection noise. The center points extracted from each frequency point are stitched together and solidified into a target spatial trajectory template unique to the smart glasses. This template is a smooth rising curve with a hardware acoustic fingerprint.

[0054] In one embodiment, step S1 above further includes: Acquire historical signals from dual microphones in noisy environments; Calculate the auto-power spectrum and cross-power spectrum of the historical signal from the dual microphones; The first complex spatial coherence is calculated based on the auto-power spectrum and cross-power spectrum of the historical signals from the dual microphones. Extract the real part of the coherence of the first complex space as an empirical coherence template.

[0055] Preferably, this invention records historical signals from dual microphones in an ambient noise environment by extracting the real part of real data. The ambient noise environment includes people walking and talking in the room, ambient white noise, etc., and the recording time covers several seconds to several minutes, ensuring that the recording includes room reverberation characteristics and head occlusion characteristics.

[0056] Calculate the auto-power spectrum and cross-power spectrum of the historical signals from the two microphones. Let the signals from the two microphone channels be x1(t) and x2(t), respectively. The auto-power spectra Φ11(f) and Φ22(f) represent the distribution of signal energy of a single microphone channel at each frequency point, respectively, and the cross-power spectrum Φ12(f) represents the distribution of cross-energy between the two microphone channels at each frequency point.

[0057] The first complex spatial coherence is calculated based on the auto-power spectrum and cross-power spectrum of the historical signals from the dual microphones. The formula for calculating the first complex spatial coherence is: .

[0058] Where E represents the average value over a long time span, Φ11(f) is the autopower spectrum of the first microphone, Φ22(f) is the autopower spectrum of the second microphone, and Φ12(f) is the cross-power spectrum of the two microphones. This complex spatial coherence includes a real part and an imaginary part. The real part reflects the energy correlation of the diffuse sound field, and the imaginary part reflects the directional deviation caused by asymmetry factors.

[0059] Extract the real part of the coherence of the first complex space as the empirical coherence template, that is: .

[0060] In the form of smart glasses worn close to the ear, the head acts as an acoustic barrier, severely disrupting the isotropy of the sound field. The openings of the two microphones often face different directions, and these asymmetric factors cause deviations that are all reflected in the imaginary part of the coherence in complex space. Extracting the real part is equivalent to performing a symmetry-forced filtering, removing stray directional interference while fully preserving the physical characteristics of the diffuse sound field oscillating with frequency in the real environment. This real part template replaces the theoretical sinc function based on the assumption of an ideal spherical diffuse sound field in traditional methods, directly reflecting the coherence characteristics of diffuse noise in the current device under actual wearing conditions. Substituting it into the coherence-diffusion energy ratio estimation equation, no negative numbers appear within the square root, and the equation always has real solutions.

[0061] Figure 2A frame-level feature density distribution of inter-channel phase difference (sin(IPD)) when a person (bystander) makes a sound in the far field directly in front.

[0062] like Figure 2 As shown, the horizontal axis represents frequency, and the vertical axis represents the sinusoidal characteristic value of the phase difference between channels. The color intensity represents the frame density of the feature at that frequency point falling into that region. Within the 0 to 4000 Hz frequency band, a high-density trajectory line smoothly rises with frequency, but the overall brightness of this trajectory line is significantly lower. In the background area surrounding the trajectory line, large areas that should have been light-colored are filled with a large number of scattered dark noise points, indicating that the sound of far-field interference is affected by room reverberation and multipath reflections. The sinusoidal characteristic value of the phase difference between channels at each frequency point in each frame is dispersed over a wide range, resulting in low spatial convergence.

[0063] Figure 3 A frame-level feature density distribution of inter-channel phase difference (sin(IPD)) for the wearer's near-field pure sound output.

[0064] like Figure 3 As shown, the horizontal axis represents frequency, and the vertical axis represents the sinusoidal characteristic value of the phase difference between channels. The color intensity represents the frame density of features falling into that region at that time-frequency point. Within the 0 to 4000 Hz frequency band, a high-density trajectory line smoothly rises with frequency. This trajectory line is extremely bright and thin, with energy highly concentrated within a narrow ascending curve and minimal variance. The background area outside the trajectory line is large and pure white, indicating that very few frames fall into these areas, forming a data vacuum.

[0065] contrast Figure 2 and Figure 3 It can be observed that: The high consistency of spatial trajectory means that both the wearer and the interfering party directly in front of them show a similar high-density trajectory line that rises smoothly with frequency in the 0 to 4000 Hz frequency band. This indicates that as long as the sound source direction is consistent, the path difference of the direct sound to the two microphones is approximately the same. The target spatial trajectory template is extracted from this bright and rising central ridge line in the wearer's density map. The difference in spatial phase convergence and dispersion is significant. When the wearer emits sound in the very near field, the direct sound dominates, resulting in an extremely high direct-to-reverberation ratio. With no wall reflection interference, the phase difference is locked onto a fixed geometric path, leading to extremely high spatial convergence. The background area of ​​the density map appears as a large area of ​​pure white. When the interfering party emits sound in the far field, the sound undergoes multiple reflections from walls and tabletops. These multipath reflections introduce random delays and phase abrupt changes, causing a sharp decrease in spatial coherence. The spatial dispersion effect coats the background area with a large amount of stray energy. Although the trajectory lines in the two maps are in the same position, there are fundamental differences in their width, brightness, and the degree of background dispersion.

[0066] Figure 4 A target spatial trajectory template for the wearer's near-field pure sound output.

[0067] like Figure 4 As shown, when the wearer emits pure sound in the very near field, the sinusoidal characteristic value of the phase difference between channels is highly concentrated on a trajectory line that rises smoothly with frequency, and the background area presents a large area of ​​pure blank space.

[0068] Along the frequency axis, the median or maximum density peak of the sinusoidal characteristic value distribution of the inter-channel phase difference is extracted at each frequency point as the statistical center value of that frequency point. The statistical center values ​​of each frequency point are combined into a curve in frequency order, and this curve is the target spatial trajectory template. This template is a smooth rising curve that varies with frequency, carrying a hardware acoustic fingerprint, representing the stable trajectory of the sinusoidal characteristic value of the inter-channel phase difference as a function of frequency when the wearer's mouth is in a very near-field position directly in front. During online processing, the sinusoidal characteristic value of the inter-channel phase difference in the current frame is compared with the target spatial trajectory template. A higher matching degree indicates that the sound source direction is closer to directly in front, and a lower matching degree indicates that the sound source direction is more deviated from directly in front.

[0069] In some embodiments, step S2 above includes: The time-domain signal input from the dual microphones is divided into frames with a preset frame length; By employing a weighted overlapping summation method and a square root Hanning window, the time-domain signals of each frame are converted into complex frequency-domain signals through Fourier transform, thus obtaining the current frame signal of the target audio signal.

[0070] Prior to this, the preset frame length refers to the duration corresponding to the number of sampling points contained in each frame of time-domain signal. When the device is working in real time at the user end, the microphone captures the external audio mixed with noise and human voice in a streaming manner (e.g., 16ms per frame).

[0071] Framing refers to dividing continuously acquired time-domain signals into multiple short time segments of fixed duration, each segment being called a frame. Adjacent frames typically overlap to ensure smooth transitions between frames and avoid information loss at frame boundaries.

[0072] The weighted overlapping-add (WOLA) method and the square-root Hann window are used to convert the time-domain signals of each frame into complex frequency-domain signals through Fourier transform (FT) to obtain the current frame signal of the target audio signal.

[0073] Weighted overlapping summation is an inter-frame signal reconstruction technique used in framing to ensure smooth transitions between adjacent frames and avoid signal discontinuities at frame boundaries. The square root Hanning window is a window function; its shape is a raised cosine window, and its mathematical expression is w(n) = 0.5 [1]. cos(2πn / (N 1))] The square root Hanning window is the square root of the Hanning window. Its shape is the same as the Hanning window, but the amplitude is compressed. When using it, the time domain signal is first multiplied by the square root Hanning window frame by frame, and then Fourier transform is performed. The Fourier transform converts the time domain signal from the time domain to the frequency domain, and the output is a complex frequency domain signal. Each frequency point contains two information components: amplitude and phase. The complex frequency domain signal is used as the current frame signal of the target audio signal.

[0074] In some embodiments, step S3 above includes: The exponential moving average algorithm is used to smoothly update the auto-power spectrum and cross-power spectrum of the current frame signal by combining the power spectrum status of historical frames. Based on the smoothed updated self-power spectrum and cross-power spectrum, the sinusoidal eigenvalue of the inter-channel phase difference of the current frame signal is calculated as a spatial direction feature. Based on the smoothed updated auto-power spectrum and cross-power spectrum, the second complex spatial coherence of the current frame signal is calculated as a spatial coherence feature.

[0075] Preferably, the Exponential Moving Average (EMA) is a recursive smoothing algorithm, whose basic form is: Current estimated value = Smoothing factor × Current observed value + (1 (Smoothing factor) × Previous frame estimate. The characteristic of this algorithm is that it only needs to retain the estimate of the previous frame when updating each frame, resulting in extremely low memory overhead and making it suitable for streaming processing.

[0076] Self-power spectrum refers to the distribution of signal energy at various frequencies of a single microphone channel. Cross-power spectrum refers to the distribution of cross-energy at various frequencies between two microphone channels.

[0077] The power spectrum state of a historical frame refers to the estimated values ​​of the auto-power spectrum and cross-power spectrum obtained after a smooth update of the previous frame.

[0078] Smooth update refers to weighted fusion of the original self-power spectrum and cross-power spectrum of the current frame with the power spectrum state of the historical frames to obtain the updated self-power spectrum and cross-power spectrum estimates.

[0079] Inter-channel phase difference (IPD) refers to the difference in phase values ​​between two microphones at the same frequency. Phase values ​​can be extracted from complex frequency domain signals. Each microphone has a phase value at the same frequency, and the difference between these two phase values ​​yields the inter-channel phase difference (IPD). The sinusoidal characteristic value of the inter-channel phase difference is obtained by taking the sine of the inter-channel phase difference, resulting in the sin(IPD) value.

[0080] Spatial orientation features refer to the sinusoidal characteristic value of the phase difference between channels of the current frame signal, which is used to compare with the target spatial trajectory template to determine the location of the sound source.

[0081] The second complex spatial coherence refers to the complex spatial coherence calculated based on the smoothed updated auto-power spectrum and cross-power spectrum of the current frame. The formula for calculating complex spatial coherence is: Γ_x(f) = Φ_12(f) / √(Φ_11(f) · Φ_22(f)) Wherein, Φ_11(f) is the smoothed autopower spectrum of the first microphone in the current frame, Φ_22(f) is the smoothed autopower spectrum of the second microphone in the current frame, and Φ_12(f) is the smoothed cross-power spectrum in the current frame. This complex result contains real and imaginary parts, serving as a feature of spatial coherence.

[0082] In some embodiments, step S4 above includes: Calculate the absolute error distance between the spatial orientation characteristics and the target spatial trajectory template at each frequency point; Based on the absolute error distance, an exponential decay function is used to map the absolute error distance to a first gain, which decreases as the absolute error distance increases.

[0083] Preferably, the spatial directional features at each frequency point of the current frame are subtracted point by point from the values ​​of the target spatial trajectory template at the same frequency point, and the absolute value is taken, with the formula D(t,f)=|sin(IPD_{t,f}) - T_phase(f)|. The smaller the value of D(t,f), the closer the directional features of the current frame at that frequency point are to the template, and the closer the sound source direction is to directly forward. The larger the value of D(t,f), the farther the directional features of the current frame at that frequency point deviate from the template, and the more the sound source direction deviates from directly forward.

[0084] An exponential attenuation function is used for mapping, with the formula G_spatial(t,f)=exp(-γ·D(t,f)). γ is the attenuation coefficient, controlling the rate at which the gain decreases with increasing error distance. A larger value of γ results in a faster decrease in gain with increasing error distance and a sharper spatial selectivity. A smaller value of γ results in a slower decrease in gain with increasing error distance and a smoother spatial selectivity. When D(t,f)=0, G_spatial=1, and the gain is at its maximum. As D(t,f) increases, G_spatial decays exponentially, approaching 0. The mapping effect is: the more positive the direction, the closer the gain is to 1, and the frequency is preserved; the more negative the direction, the closer the gain is to 0, and the frequency is suppressed.

[0085] In one embodiment, step S5 includes: Obtain the real part of the spatial coherence feature and the square of the amplitude of the spatial coherence feature; The coherent diffusion energy ratio is calculated based on the real part of the spatial coherence feature, the square of the amplitude of the spatial coherence feature, and the empirical coherence template. Based on the coherent diffusion energy ratio, the Wiener filtering formula is used to map the coherent diffusion energy ratio to a second gain, which increases as the coherent diffusion energy ratio increases.

[0086] Preferably, the real part of the complex spatial coherence Re{Γ_x}, the square of the magnitude |Γ_x|², and the offline-calibrated empirical coherence template Γ_n of the current frame are substituted into the analytical equation to solve for the coherent diffusion energy ratio.

[0087] The coherence-to-diffuse ratio (CDR) is a physical quantity that measures the purity of a sound field. It represents the ratio of direct sound energy to diffuse sound energy in the current frame signal. A higher CDR value indicates a higher proportion of direct sound and a purer sound field. A lower CDR value indicates a higher proportion of diffuse sound and a more reverberant sound field.

[0088] In the analytical equation, Γ_x represents the short-time complex spatial coherence of the current audio frame, calculated from the current frame signal, and includes both real and imaginary parts. Re{Γ_x} represents the real part of the current complex coherence, reflecting the correlation degree of the in-phase components of the target audio signal. |Γ_x|² represents the square of the current complex coherence amplitude, i.e., the magnitude of coherence, reflecting the overall correlation strength of the dual-channel signals, with a value ranging from 0 to 1. Γ_n represents the empirical coherence template extracted in step one during offline calibration. This template is a pure real number derived from the real part of the complex spatial coherence of the real environmental noise floor.

[0089] In traditional methods, Γ_n in the above equation is replaced by the purely mathematical sinc function. The sinc function is based on the assumption of an ideal spherical diffuse sound field. However, due to head obstruction and microphone orientation deviations in smart glasses, the actual sound field is asymmetrical, leading to a mismatch between theoretical and actual values. Substituting these values ​​results in negative numbers calculated inside the square root, rendering the CDR value a meaningless complex number and causing the algorithm to crash. This solution replaces Γ_n with a template of the real part extracted from real data, eliminating negative numbers inside the square root and ensuring the equation always has a real solution.

[0090] The denominator of the equation is |Γ_x|² 1. When the coherence approaches 1 infinitely, the denominator approaches 0, leading to a division-by-zero error. In engineering implementations, a very small positive bias constant ε is introduced into the denominator term, such as 10. -8 To ensure the absolute stability of streaming computation, the CDR value is always a real number of finite size.

[0091] After calculating the CDR value, it is mapped to a purity gain between 0 and 1 using the Wiener filtering formula, which is G_cdr = max(G_min, 1.0). The formula is sqrt(μ / (CDR + 1.0)). Here, μ is an adjustable over-subtraction factor, typically ranging from 1.0 to 2.0. A larger μ value results in stronger noise reduction, and a smaller second gain for the same CDR value. G_min is the lower limit of the gain, preventing harsh distortion caused by complete spectrum breakage. When the CDR value is large, the sqrt term approaches 0, and the second gain approaches 1. When the CDR value is small, the second gain is limited to around G_min.

[0092] When someone speaks directly in front of you, their directional characteristics match the wearer's, leading Branch 1 to mistakenly interpret the direction as correct, resulting in a higher first gain. However, the sound travels through the far field, carrying multiple reflections from the room's walls and tabletop. Multipath reflections and spatial diffusion effects decrease the complex spatial coherence Γ_x of the current frame, causing it to converge towards the offline empirical coherence template Γ_n. Substituting the decreased Γ_x and Γ_n into the analytical equation yields an extremely low CDR value, and the second gain after Wiener filtering approaches zero. This branch captures the essential characteristics of far-field interference from the dimensions of distance and purity; even if the first gain misjudges the direction as correct, the second gain can effectively eliminate the interference.

[0093] In some embodiments, step S6 above includes: Multiply the first gain and the second gain at various frequency points in the frequency domain to obtain the target mask gain; The amplitude spectrum of the current frame signal of the target audio signal is averaged, and the phase spectrum of the preset reference channel is extracted from the current frame signal of the target audio signal. The basis complex spectrum is synthesized based on the averaged amplitude spectrum and the phase spectrum of the preset reference channel. Multiply the target mask gain by the basis complex spectrum at each frequency point to obtain the denoised frequency domain signal; The denoised frequency domain signal is subjected to inverse Fourier transform and weighted overlapping summation to synthesize a time-domain denoised signal.

[0094] Preferably, after the first gain G_spatial and the second gain G_cdr are calculated independently, the two gains are multiplied point-by-point at each frequency point to obtain the fused gain value at that frequency point. Simultaneously, a very small fallback constant G_min is introduced, and the maximum value between the fused gain and G_min is taken as the target mask gain G_final. G_min is used to prevent the spectrum from completely returning to zero, thus preventing harsh distortion; for example, 0.05 in one example. The physical meaning of multiplying point-by-frequency is that only frequency points that simultaneously satisfy both the correct orientation and sufficiently close distance are retained; if either condition is not met, the target gain at that frequency point will be suppressed to close to G_min.

[0095] After multiplying the two gains, only when the sound is directly in front, extremely close, and without reverberation, will both the first and second gains approach 1, the target mask gain approach 1, and this frequency point be preserved. For sounds with the correct direction but far away, the first gain approaches 1, but the second gain approaches 0, and the target gain approaches G_min, thus being suppressed. For sounds with the wrong direction, the first gain approaches 0, and regardless of the value of the second gain, the target gain approaches G_min, thus being suppressed.

[0096] The absolute values ​​of the complex spectra of the two microphones are taken to obtain their respective amplitude spectra. The two amplitude spectra are then averaged at the same frequency point to obtain the average amplitude spectrum. Simultaneously, the phase spectrum of a preset master reference microphone (the microphone closer to the speaker's mouth in the hardware layout) is extracted from the target audio signal. The average amplitude spectrum is combined with the phase spectrum of the master reference microphone to synthesize the basis complex spectrum. This operation can resist monophonic spectral fluctuations caused by early reflections, avoiding spectral dips or discontinuities in the reconstructed signal.

[0097] The target mask gain G_final is multiplied by the basis complex spectrum at each frequency point to complete the real-time removal of interference noise, resulting in a denoised frequency domain signal. The frequency domain signal is then restored to the time domain signal through inverse short-time Fourier transform, and combined with a weighted overlapping summation window, the time domain signals of each frame are seamlessly stitched into a one-dimensional monophonic time domain speech signal, which is then output to the backend voice wake-up and recognition system or the communication peer.

[0098] The beneficial effects of this embodiment include: 1. Extremely high hardware tolerance (no precision calibration required): It abandons the acoustic physics equations based on vernier caliper precision and internalizes microphone position error, aperture orientation difference and head diffraction effect into the "offline experience template", which greatly reduces the pressure on hardware manufacturing yield of wearable devices.

[0099] 2. Perfectly solves the physical blind spot of "collinear interference": Traditional algorithms cannot distinguish between the speaker directly in front and people standing to the side directly in front. This invention uses a dual-mask design. The spatial branch recognizes the "correct direction", but the CDR branch will keenly capture the "spatial diffusion (reverberation)" of people at a distance, thus achieving precise filtering across dimensions.

[0100] 3. Extremely low computational power and zero-wait streaming: The entire process does not involve matrix inversion or complex nonlinear convolutions, only using basic complex number multiplication and addition and scalar operations. The computational complexity is extremely low, making it fully compatible with the computing power-constrained edge chips of smart glasses, with an algorithm latency of only one frame (e.g., 16ms).

[0101] In one embodiment, a dual-microphone directional noise reduction device is provided, and its structural block diagram can be as follows: Figure 5 As shown, it includes: a microphone for acquiring target audio signals; a memory for storing computer programs; and a processor for executing the computer programs to implement the steps of the above-described dual-microphone directional noise reduction method based on empirical coherence templates.

[0102] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, internal memory, a network interface, and a database connected via a system bus. The processor provides computing and control capabilities. The internal memory provides an environment for the operation of the operating system and computer programs stored in a non-volatile storage medium. The non-volatile storage medium stores the operating system, computer programs, and database. The network interface of the computer device is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side method for dual-microphone directional noise reduction based on empirical coherence templates.

[0103] The computer device can also be a client, including a processor, internal memory, network interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The internal memory provides an environment for the operating system and computer programs running on non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements the client-side functions or steps of a dual-microphone directional noise reduction method based on empirical coherence templates.

[0104] An embodiment of this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a dual-microphone directional noise reduction method based on empirical coherence templates.

[0105] It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0106] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0107] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0108] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A dual-microphone directional noise reduction method based on empirical coherence templates, characterized in that, include: A target spatial trajectory template and an empirical coherence template are pre-set. The target spatial trajectory template is used to characterize the trajectory of the phase difference characteristics between channels as a function of frequency, and the empirical coherence template is used to characterize the coherence characteristics of the diffuse sound field. Acquire the current frame signal of the target audio signal, which is obtained based on at least two microphones; Based on the current frame signal of the target audio signal, extract spatial direction features and spatial coherence features; Calculate the first gain based on the spatial orientation features and the target spatial trajectory template; The second gain is calculated based on the spatial coherence characteristics and the empirical coherence template; By combining the first gain and the second gain, noise reduction processing is performed on the current frame signal of the target audio signal, and the noise-reduced signal is output.

2. The method according to claim 1, wherein the pre-set target space trajectory template comprises: Acquire historical signals of the target sound source in the target direction; The historical signal is subjected to time-frequency transformation to obtain the time-frequency domain representation of the historical signal at each frequency point; Based on the time-frequency domain characterization, calculate the sinusoidal characteristic value of the inter-channel phase difference at each frequency point for each frame of the historical signal; A two-dimensional statistical distribution is constructed based on the sinusoidal characteristic value of the inter-channel phase difference at each frequency point in each frame; Along the frequency axis, extract the statistical center value of the two-dimensional statistical distribution at each frequency point; The statistical center values ​​of each frequency point are combined in frequency order to obtain the target spatial trajectory template.

3. The method according to claim 1, wherein the pre-set experience coherence template comprises: Acquire historical signals from dual microphones in noisy environments; Calculate the auto-power spectrum and cross-power spectrum of the historical signal from the dual microphones; The first complex spatial coherence is calculated based on the auto-power spectrum and the cross-power spectrum of the historical signals from the dual microphones. Extract the real part of the first complex spatial coherence as the empirical coherence template.

4. The method according to claim 1, wherein acquiring the current frame signal of the target audio signal includes: The time-domain signal input from the dual microphones is divided into frames with a preset frame length; The weighted overlapping summation method and the square root Hanning window are used to convert the time domain signal of each frame into a complex frequency domain signal through Fourier transform, so as to obtain the current frame signal of the target audio signal.

5. The method according to claim 1, wherein extracting spatial orientation features and spatial coherence features based on the current frame signal of the target audio signal comprises: The exponential moving average algorithm is used, combined with the power spectrum status of historical frames, to smoothly update the auto-power spectrum and cross-power spectrum of the current frame signal; Based on the smoothed updated self-power spectrum and cross-power spectrum, the sinusoidal characteristic value of the inter-channel phase difference of the current frame signal is calculated as the spatial direction feature; Based on the smoothed updated self-power spectrum and cross-power spectrum, the second complex spatial coherence of the current frame signal is calculated as the spatial coherence feature.

6. The method according to claim 1, wherein calculating the first gain based on the spatial orientation features and the target spatial trajectory template comprises: Calculate the absolute error distance between the spatial orientation features and the target spatial trajectory template at each frequency point; Based on the absolute error distance, an exponential decay function is used to map the absolute error distance to a first gain, which decreases as the absolute error distance increases.

7. The method according to claim 1, wherein calculating the second gain based on the spatial coherence characteristics and the empirical coherence template comprises: Obtain the real part of the spatial coherence feature and the square of the amplitude of the spatial coherence feature; The coherent diffusion energy ratio is calculated based on the real part of the spatial coherence feature, the square of the amplitude of the spatial coherence feature, and the empirical coherence template. Based on the coherent diffusion energy ratio, the Wiener filtering formula is used to map the coherent diffusion energy ratio to a second gain, which increases as the coherent diffusion energy ratio increases.

8. The method according to claim 1, wherein fusing the first gain and the second gain to perform noise reduction processing on the current frame signal of the target audio signal, and outputting the noise-reduced signal includes: Multiply the first gain and the second gain at each frequency point in the frequency domain to obtain the target mask gain; The amplitude spectrum of the current frame signal of the target audio signal is averaged, and the phase spectrum of the preset reference channel is extracted from the current frame signal of the target audio signal. The basis complex spectrum is synthesized based on the averaged amplitude spectrum and the phase spectrum of the preset reference channel. Multiply the target mask gain by the basis complex spectrum at each frequency point to obtain the noise-reduced frequency domain signal; The denoised frequency domain signal is subjected to inverse Fourier transform and weighted overlapping summation to synthesize a time-domain denoised signal.

9. A dual-microphone directional noise reduction device, characterized in that, include: At least two microphones are used to capture the target audio signal; Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the dual-microphone directional noise reduction method based on empirical coherence templates as described in any one of claims 1 to 8.

Images