CN121237113A - A noise reduction recording method for array microphones based on cascaded noise reduction and blind source separation - Google Patents
A noise reduction recording method for array microphones based on cascaded noise reduction and blind source separationInfo
- Publication number
- CN121237113A CN121237113A CN202511338173.9A CN202511338173A CN121237113A CN 121237113 A CN121237113 A CN 121237113A CN 202511338173 A CN202511338173 A CN 202511338173A CN 121237113 A CN121237113 A CN 121237113A
- Authority
- CN
- China
- Prior art keywords
- voice
- signal
- algorithm
- noise
- signals
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Circuit For Audible Band Transducer (AREA)
Abstract
本发明涉及一种基于级联降噪与盲源分离的阵列麦克风降噪录音方法，属于语音信号处理与录音技术领域，该方法包括：配置多通道阵列麦克风，保障通道信号幅值与相位一致，采集原始多通道语音信号；采用加权核函数盲源分离算法，提取信号信噪比分布特征并赋予核函数权重，再通过独立分量分析模型解耦得到目标语音与干扰信号分量；执行目标导向型自适应级联降噪，依次经定向拾音锁定主讲人语音、降噪、滤除混响、语音掩码神经网络结合远场拾音算法增强远场目标语音；通过语音特征感知无损编码处理目标语音并存储。本发明实现多通道信号稳定采集、混合信号精准分离及噪声混响分层抑制，显著提升远场语音信噪比与清晰度，适用于单人主讲或多人对话场景。
This invention relates to a noise reduction and recording method for array microphones based on cascaded noise reduction and blind source separation, belonging to the field of speech signal processing and recording technology. The method includes: configuring a multi-channel array microphone to ensure consistency in the amplitude and phase of the channel signals, and acquiring the original multi-channel speech signals; employing a weighted kernel function blind source separation algorithm to extract the signal-to-noise ratio distribution characteristics of the signals and assign weights to the kernel function, then decoupling the target speech and interference signal components through an independent component analysis model; performing target-oriented adaptive cascaded noise reduction, sequentially using directional pickup to lock onto the speaker's voice, noise reduction, reverberation filtering, and a speech masking neural network combined with a far-field pickup algorithm to enhance the far-field target speech; and processing and storing the target speech through lossless encoding based on speech feature perception. This invention achieves stable acquisition of multi-channel signals, accurate separation of mixed signals, and layered suppression of noise and reverberation, significantly improving the far-field speech signal-to-noise ratio and clarity, and is suitable for single-person presentations or multi-person dialogue scenarios.
Description
Technical Field
The invention belongs to the technical field of voice signal processing and recording, and particularly relates to a noise reduction recording method of an array microphone based on cascade noise reduction and blind source separation.
Background
With the development of voice signal processing technology, the array microphone has multichannel signal acquisition capability, and is widely applied to scenes such as far-field recording, conference recording and multi-person conversation acquisition, and the core requirement is to realize high-definition voice recording in a complex environment, so that the signal-to-noise ratio and the intelligibility of voice signals are ensured. However, the existing array microphone noise reduction recording technology still has various technical bottlenecks, which are difficult to meet the high quality requirements in practical application, and the specific disadvantages are as follows:
The consistency of multichannel signals and the integrity of recording are difficult to guarantee. Meanwhile, partial recording schemes are not provided with effective frame loss prevention and sound breaking monitoring and compensating mechanisms, and under the condition of signal transmission load fluctuation or environmental interference, the phenomena of frame loss, sound disconnection or sound breaking of recording often occur, so that the recording content is incomplete.
The mixed signal separation accuracy is insufficient. In the prior art, a blind source separation algorithm is a common means for realizing separation of voice and interference signals, but the traditional blind source separation mostly adopts fixed kernel function parameters, and the signal-to-noise ratio distribution characteristics of original signals are not combined to dynamically adjust weights, namely in a low signal-to-noise ratio area (such as an environment noise intensive scene), the target voice and the interference signals are difficult to effectively distinguish due to lack of targeted weight reinforcement, so that a great amount of noise remains in the separated target voice or voice components are mixed in the interference signals, and the basic effect of subsequent noise reduction processing is influenced.
Noise reduction processing limitations are significant and far-field speech quality is poor. The existing noise reduction scheme is multi-focusing single-type noise suppression or can only respectively process stable noise (such as air conditioner bottom noise and fan noise) and non-stable noise (such as sudden door closing sound and clapping sound), but lacks cascading collaborative design of noise suppression, reverberation filtering and far field enhancement, on one hand, the problem that the stable noise and the non-stable noise are difficult to simultaneously suppress, one noise is easy to suppress, and the other noise is easy to remain is solved, on the other hand, for the voice reverberation tail in an open environment, the traditional filtering algorithm mainly adopts fixed window parameters, reverberation delay cannot be adaptively matched, incomplete reverberation filtering or damage to voice details are easy to be caused, in addition, in a far field scene, the existing enhancement algorithm is mainly dependent on simple gain adjustment, residual noise and target voice are not accurately distinguished by combining voice characteristics, so that the far field voice definition is low, the intelligibility is poor, and high-definition recording requirements in a single main speaking scene or a multi-person conversation scene are difficult to meet.
The code storage is not adequately matched with the voice characteristics. The coding and storage of the existing recording scheme mostly adopts a general lossless coding strategy, is not optimized for the characteristics of the pitch period, harmonic distribution and the like of the voice signal, namely, the voice energy concentration area (such as vowel pronunciation section) is not subjected to enhanced fidelity processing, so that key voice details can be lost, and the mute or weak sound section is not subjected to redundant compression, so that storage resource waste is easily caused, and the recording tone quality and storage efficiency are difficult to balance.
In summary, the existing array microphone noise reduction recording technology has defects in aspects of multi-channel signal stability, mixed signal separation precision, noise reduction comprehensiveness, coding suitability and the like, cannot meet the high-definition recording requirements in far field, multi-noise and multi-scene, and needs a technical scheme capable of breaking through the bottleneck.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a noise reduction recording method of an array microphone based on cascade noise reduction and blind source separation;
the aim of the invention can be achieved by the following technical scheme:
The array microphone noise reduction recording method based on cascade noise reduction and blind source separation is characterized by comprising the following steps of:
S1, configuring a multichannel array microphone, wherein the multichannel array microphone maintains the amplitude and phase collaborative consistency of channel recording signals through a collaborative signal processing algorithm, monitors the channel recording process through an anti-frame-loss sound breaking mechanism, and collects original multichannel voice signals in the environment;
S2, preprocessing the original multi-channel voice signal based on a weighted kernel function blind source separation algorithm, firstly extracting signal-to-noise ratio distribution characteristics, giving kernel function weight, and then decoupling the original multi-channel voice signal through an independent component analysis model to obtain a target voice signal component and an interference signal component;
S3, executing a target-oriented adaptive cascade noise reduction model on the target voice signal component, namely firstly locking a voice signal sent by a speaker through a directional pickup technology based on a preset adjustable pickup range parameter, then adopting an adaptive multi-mode noise reduction algorithm to inhibit stable noise and unsteady noise in the voice signal, then adopting an acoustic signal adaptive processing algorithm to filter voice reverberation tail in the voice signal output after noise inhibition, finally combining a voice mask neural network and a far-field pickup algorithm, distinguishing residual noise and target voice in the voice signal output after reverberation filtering through the voice mask neural network, calculating energy expectation of the target voice, and enhancing the target voice in a far-field environment based on the far-field pickup algorithm;
And S4, adopting a voice characteristic perception lossless coding algorithm to the target voice signal, namely extracting the pitch period and harmonic distribution characteristics of the target voice signal, adopting a high-fidelity coding strategy to a voice energy concentration area, and storing the coded target voice signal to generate a noise reduction recording process file.
The collaborative signal processing algorithm comprises a dynamic amplitude compensation sub-algorithm and a phase calibration sub-algorithm, wherein the dynamic amplitude compensation sub-algorithm collects amplitude differences of channel recording signals in real time, firstly analyzes reasons for the differences to generate dynamic proportionality coefficients, adjusts signal gains of amplitude channels through the dynamic proportionality coefficients, presets reference channels, calculates phase delay amounts of other channels and the reference channels in real time based on phases of the reference channels, generates reverse compensation signals according to the phase delay amounts, and is overlapped to corresponding channels.
The frame loss prevention and sound breaking mechanism comprises a signal buffer pool and a frame loss detection and retransmission sub-algorithm, wherein a buffer space is preset in the signal buffer pool, real-time recording signals in continuous time are temporarily stored in the buffer space, the frame loss detection and retransmission sub-algorithm monitors frame loss conditions in the signal transmission process in real time, and judges the frame loss conditions by comparing signal integrity marks before and after transmission, when the frame loss phenomenon is detected to influence recording continuity, a retransmission request is immediately triggered, and non-frame loss signals stored before a frame loss period are called from the signal buffer pool and are supplemented to a transmission link.
The weighting kernel function blind source separation algorithm comprises the steps of carrying out multidimensional analysis on signal-to-noise ratio distribution characteristics of an original multichannel voice signal, monitoring fluctuation frequency and amplitude of energy of the original multichannel voice signal in a time domain, calculating energy ratio of the original multichannel voice signal to a noise signal in a frequency domain, judging through fusion of time domain and frequency domain characteristics, dividing a low signal-to-noise ratio area and a high signal-to-noise ratio area, calculating kernel function weights by adopting a nonlinear weighting function, dynamically adjusting weight values of the functions in combination with the noise energy ratio in the corresponding area, and carrying out differential weight distribution based on actual characteristics of the original multichannel voice signal.
The independent component analysis model is integrated with a gradient descent optimization sub-algorithm, wherein the gradient descent optimization sub-algorithm takes mutual information of signals after separation reduction as an objective function, reduces the relevance between a target voice signal component and an interference signal component through optimization, adjusts parameters of the independent component analysis model according to a currently calculated mutual information value during iteration, gradually optimizes separation effects, and stops iteration when the mutual information difference obtained by two adjacent iteration calculation is reduced to a preset threshold value.
The directional pickup technology is combined with a spatial spectrum estimation algorithm, wherein the spatial spectrum estimation algorithm firstly collects phase information of channel voice signals, calculates an attitude angle of a source of the channel voice signals through comparing phase differences among channels, compares the attitude angle with an azimuth angle range corresponding to the preset adjustable pickup range, judges that the target speaker voice signals are obtained when the attitude angle is in the azimuth angle range, and locks the target speaker voice signals through enhancing channel signal gain.
The self-adaptive multi-mode noise reduction algorithm comprises a noise type classification sub-algorithm, wherein the noise type classification sub-algorithm firstly extracts time domain and frequency domain characteristics of a voice signal, generates a characteristic analysis result, divides noise into stable noise and non-stable noise according to the characteristic analysis result, adopts a frequency domain notch filtering strategy to the stable noise, sets a filtering notch in a frequency band in which noise is concentrated, weakens the stable noise of the frequency band, adopts a time domain threshold suppression strategy to the non-stable noise, sets a signal amplitude threshold, judges the voice signal to be the non-stable noise when the voice signal exceeds the threshold, and suppresses the voice signal through a threshold filter.
The acoustic signal self-adaptive processing algorithm comprises a reverberation time delay estimation sub-algorithm and a dynamic filtering sub-algorithm, wherein the reverberation time delay estimation sub-algorithm firstly identifies a direct wave and a first reflected wave in the voice signal, the direct wave directly propagates to a microphone from a sound source, the first reflected wave propagates to the microphone after primary reflection, the direct wave is delayed, the reverberation time delay is obtained by calculating the time difference between the direct wave and the first reflected wave, and the dynamic filtering sub-algorithm adjusts the length of a filtering window according to the reverberation time delay, the window length needs to cover the duration of the reverberation tail, and the reverberation tail is completely filtered.
The voice masking neural network adopts a framework combining a two-way long-short-term memory network layer and an attention mechanism, wherein the two-way long-term memory network layer can simultaneously extract historical context characteristics and future context characteristics of a voice signal, extract time sequence rules of the voice signal from the beginning to the current moment in the forward direction, extract time sequence rules of the voice signal from the current moment to the end in the backward direction, capture the time sequence characteristics of the voice signal through two-way feature fusion, the attention mechanism focuses on a frequency band in which human voice is mainly distributed, and the voice masking neural network prioritizes the frequency band when distinguishing residual noise from the human voice by strengthening characteristic weights of the frequency band.
The far-field pickup algorithm comprises a voice energy compensation sub-algorithm and a frequency equalization sub-algorithm, wherein the voice energy compensation sub-algorithm is based on the attenuation law of far-field voice transmission, namely the law that voice energy is gradually weakened along with the increase of transmission distance, firstly analyzes the sound source distance in the current recording environment, then calculates the energy attenuation degree according to the sound source distance, generates a corresponding energy compensation coefficient, improves the overall energy of the far-field voice through the energy compensation coefficient, and identifies a high-frequency region in the far-field voice according to the characteristic that the attenuation of a high-frequency band signal in far-field transmission is obvious, and adjusts the signal intensity of the high-frequency region through gain.
The voice feature perception lossless coding algorithm comprises a feature-driven coding redundancy optimization sub-algorithm, wherein the feature-driven coding redundancy optimization sub-algorithm comprises the steps of firstly analyzing energy distribution features of a target voice signal, identifying a voice energy concentration region and a mute region, wherein the voice energy concentration region is a core bearing part of voice content, the mute region is free of effective voice and only comprises weak background noise, a high-fidelity coding strategy is adopted for the voice energy concentration region, detail features of the signal are reserved, and a coding redundancy compression strategy is adopted for the mute region to remove redundant information in the signal.
The voice quality detection feedback step comprises the steps of detecting the target voice signal quality by adopting a voice quality perception evaluation algorithm, comprehensively judging the voice quality grade by analyzing the definition, naturalness and noise residual quantity indexes of the target voice, feeding the quality evaluation result back to a noise reduction link when the target voice signal quality does not reach a preset standard, repeatedly executing noise reduction operation until the detected voice quality reaches the preset standard, and entering the coding storage step.
The beneficial effects of the invention are as follows:
And the consistency and recording integrity of the multichannel signals are obviously improved. The invention solves the problems of amplitude deviation and phase delay of each channel caused by hardware difference and transmission loss through a collaborative signal processing algorithm, ensures that the multichannel recording signals are highly collaborative in amplitude and phase, provides high quality basis for subsequent signal processing, simultaneously monitors the recording process in real time by an anti-frame-loss sound breaking mechanism, avoids frame loss, disconnection and sound breaking phenomena caused by signal transmission fluctuation or environmental interference, ensures complete and non-missing recording content, and solves the problem of pain point that 'original signal distortion influences subsequent processing' in the prior art.
The separation precision of the mixed signal is greatly improved. The invention adopts a weighted kernel function blind source separation algorithm to break through the limitation of the traditional fixed kernel function parameters, dynamically endows kernel function weights by extracting the signal-to-noise ratio distribution characteristics of original signals, namely strengthening weights in a noise-intensive low signal-to-noise ratio area, accurately distinguishes target voice and interference signals, efficiently decouples mixed signals by combining an independent component analysis model, effectively reduces noise residues of separated target voice, simultaneously avoids mixing voice components in the interference signals, lays a foundation of 'high-purity target signals' for subsequent noise reduction treatment, and improves the separation precision remarkably compared with the traditional blind source separation technology.
And double optimization of noise reduction comprehensiveness and far-field voice quality is realized. The invention constructs a cooperative processing system of 'directional pickup-noise suppression-reverberation filtering-far field enhancement' through a target-oriented adaptive cascade noise reduction model, wherein the directional pickup technology locks the voice of a speaker based on an adjustable range, reduces irrelevant interference from space dimension, an adaptive multi-mode noise reduction algorithm can simultaneously and efficiently suppress steady-state noise and non-steady-state noise, avoids the problem of 'suppressing single noise and remaining another noise', an acoustic signal adaptive processing algorithm can match a reverberation time delay dynamic adjustment parameter, thoroughly filters reverberation tail without damaging voice details, a voice mask neural network is combined with a far-field pickup algorithm, can accurately distinguish residual noise from target voice, remarkably improves the definition and intelligibility of far-field voice through energy expectation calculation and gain optimization, perfectly adapts to scenes such as single speaker, multi-person conversation and the like, and has greatly improved far-field voice signal to noise ratio and intelligibility compared with the prior art.
And the sound quality and the storage efficiency of the recording are balanced. The voice characteristic perception lossless coding algorithm breaks through the limitation of the traditional general coding, optimizes coding strategies aiming at the characteristics of the pitch period, harmonic distribution and the like of voice signals, adopts high-fidelity coding on voice energy concentrated areas, furthest reserves key voice details such as vowels, consonants and the like, avoids tone quality loss, performs coding redundancy compression on mute or weak voice segments, reduces storage resource occupation, reduces storage cost while guaranteeing high-definition recording quality, and realizes optimal balance of tone quality-efficiency.
In conclusion, the invention solves the bottleneck of the prior art in the aspects of multi-channel signal stability, mixed signal separation precision, noise reduction comprehensiveness, coding suitability and the like, can stably output high-definition recording in far field, multi-noise and multi-scene, and has extremely strong practicability and popularization value.
Drawings
The present invention is further described below with reference to the accompanying drawings for the convenience of understanding by those skilled in the art.
FIG. 1 is a schematic flow chart of a method for noise reduction and recording of an array microphone based on cascade noise reduction and blind source separation;
Fig. 2 is a diagram of a voice enhancement effect of an array microphone noise reduction recording method based on cascade noise reduction and blind source separation according to the present invention.
Detailed Description
In order to further describe the technical means and effects adopted by the invention for achieving the preset aim, the following detailed description is given below of the specific implementation, structure, characteristics and effects according to the invention with reference to the attached drawings and the preferred embodiment.
Referring to fig. 1-2, a method for noise reduction and recording of an array microphone based on cascade noise reduction and blind source separation includes the steps of:
The array microphone noise reduction recording method based on cascade noise reduction and blind source separation is characterized by comprising the following steps of:
S1, configuring a multichannel array microphone, wherein the multichannel array microphone maintains the amplitude and phase collaborative consistency of channel recording signals through a collaborative signal processing algorithm, monitors the channel recording process through an anti-frame-loss sound breaking mechanism, and collects original multichannel voice signals in the environment;
S2, preprocessing the original multi-channel voice signal based on a weighted kernel function blind source separation algorithm, firstly extracting signal-to-noise ratio distribution characteristics, giving kernel function weight, and then decoupling the original multi-channel voice signal through an independent component analysis model to obtain a target voice signal component and an interference signal component;
S3, executing a target-oriented adaptive cascade noise reduction model on the target voice signal component, namely firstly locking a voice signal sent by a speaker through a directional pickup technology based on a preset adjustable pickup range parameter, then adopting an adaptive multi-mode noise reduction algorithm to inhibit stable noise and unsteady noise in the voice signal, then adopting an acoustic signal adaptive processing algorithm to filter voice reverberation tail in the voice signal output after noise inhibition, finally combining a voice mask neural network and a far-field pickup algorithm, distinguishing residual noise and target voice in the voice signal output after reverberation filtering through the voice mask neural network, calculating energy expectation of the target voice, and enhancing the target voice in a far-field environment based on the far-field pickup algorithm;
And S4, adopting a voice characteristic perception lossless coding algorithm to the target voice signal, namely extracting the pitch period and harmonic distribution characteristics of the target voice signal, adopting a high-fidelity coding strategy to a voice energy concentration area, and storing the coded target voice signal to generate a noise reduction recording process file.
The collaborative signal processing algorithm comprises a dynamic amplitude compensation sub-algorithm and a phase calibration sub-algorithm, wherein the dynamic amplitude compensation sub-algorithm collects amplitude differences of channel recording signals in real time, firstly analyzes reasons for the differences to generate dynamic proportionality coefficients, adjusts signal gains of amplitude channels through the dynamic proportionality coefficients, presets reference channels, calculates phase delay amounts of other channels and the reference channels in real time based on phases of the reference channels, generates reverse compensation signals according to the phase delay amounts, and is overlapped to corresponding channels.
After the multichannel array microphone is started, recording signals of each channel in 5-10 continuous sampling periods are acquired in real time, the signals of each channel are led into amplitude analysis, the average amplitude of the signals of each channel is compared, if the amplitude of one channel is continuously lower than that of other channels and fluctuation is stable, the difference is judged to be hardware sensitivity difference, if the amplitude fluctuation changes along with transmission time, the difference is judged to be transmission link loss, a dynamic proportionality coefficient is generated according to the difference reason, the fixed proportionality coefficient is calculated according to the hardware sensitivity difference and the target amplitude/current amplitude (the amplitude deviation of each channel is ensured to be reduced to a negligible range after adjustment), the transmission link loss is generated according to the real-time loss rate reverse deduction, the compensation coefficient is added to a signal amplifying circuit of the corresponding channel, and the signal gain of the low-amplitude channel is gradually adjusted until the amplitude of each channel keeps stable and consistent in 3 continuous sampling periods.
The frame loss prevention and sound breaking mechanism comprises a signal buffer pool and a frame loss detection and retransmission sub-algorithm, wherein a buffer space is preset in the signal buffer pool, real-time recording signals in continuous time are temporarily stored in the buffer space, the frame loss detection and retransmission sub-algorithm monitors frame loss conditions in the signal transmission process in real time, and judges the frame loss conditions by comparing signal integrity marks before and after transmission, when the frame loss phenomenon is detected to influence recording continuity, a retransmission request is immediately triggered, and non-frame loss signals stored before a frame loss period are called from the signal buffer pool and are supplemented to a transmission link.
The weighting kernel function blind source separation algorithm comprises the steps of carrying out multidimensional analysis on signal-to-noise ratio distribution characteristics of an original multichannel voice signal, monitoring fluctuation frequency and amplitude of energy of the original multichannel voice signal in a time domain, calculating energy ratio of the original multichannel voice signal to a noise signal in a frequency domain, judging through fusion of time domain and frequency domain characteristics, dividing a low signal-to-noise ratio area and a high signal-to-noise ratio area, calculating kernel function weights by adopting a nonlinear weighting function, dynamically adjusting weight values of the functions in combination with the noise energy ratio in the corresponding area, and carrying out differential weight distribution based on actual characteristics of the original multichannel voice signal.
The signal-to-noise ratio characteristic multidimensional analysis stage comprises the steps of dividing an original multichannel voice signal into one frame according to 50ms, calculating the energy fluctuation frequency (the number of times that the energy exceeds the average value by 2 times per second) and the fluctuation amplitude (the difference value between the maximum energy and the minimum energy) of each frame signal, namely, frames with high fluctuation frequency and large amplitude, judging the frames as low signal-to-noise ratio region candidates, carrying out Fourier transformation on each frame signal, dividing a plurality of frequency bands (covering human voice and common noise frequency bands), calculating the frequency band with the ratio of voice energy (based on voice frequency band characteristic recognition) to noise energy (non-voice frequency band energy) being smaller than 1, judging the frequency band as low signal-to-noise ratio, fusing the time domain and frequency domain analysis results, and determining the overlapping region of the time domain candidates and the frequency domain low ratio as the high signal-to-noise ratio region.
The nonlinear weighting function calculation and weight giving stage comprises selecting S-shaped nonlinear weighting function, inputting the function as 'noise energy duty ratio in region' (noise energy/(voice energy+noise energy)), outputting as weight value, dynamically adjusting weight, outputting high weight value (ensuring that the signal in the region gets more calculation resources during separation) if the noise energy duty ratio exceeds 60% for the low signal-to-noise ratio region, outputting low weight value (avoiding over processing) if the voice energy duty ratio exceeds 70% for the high signal-to-noise ratio region, and applying weight, binding the calculated weight value with kernel function parameters of the corresponding region, adjusting calculation coefficient of the signals in each region according to the weight value in matrix operation of a blind source separation algorithm, and strengthening signal separation gradient of the low signal-to-noise ratio region.
The independent component analysis model is integrated with a gradient descent optimization sub-algorithm, wherein the gradient descent optimization sub-algorithm takes mutual information of signals after separation reduction as an objective function, reduces the relevance between a target voice signal component and an interference signal component through optimization, adjusts parameters of the independent component analysis model according to a currently calculated mutual information value during iteration, gradually optimizes separation effects, and stops iteration when the mutual information difference obtained by two adjacent iteration calculation is reduced to a preset threshold value.
The directional pickup technology is combined with a spatial spectrum estimation algorithm, wherein the spatial spectrum estimation algorithm firstly collects phase information of channel voice signals, calculates an attitude angle of a source of the channel voice signals through comparing phase differences among channels, compares the attitude angle with an azimuth angle range corresponding to the preset adjustable pickup range, judges that the target speaker voice signals are obtained when the attitude angle is in the azimuth angle range, and locks the target speaker voice signals through enhancing channel signal gain.
The self-adaptive multi-mode noise reduction algorithm comprises a noise type classification sub-algorithm, wherein the noise type classification sub-algorithm firstly extracts time domain and frequency domain characteristics of a voice signal, generates a characteristic analysis result, divides noise into stable noise and non-stable noise according to the characteristic analysis result, adopts a frequency domain notch filtering strategy to the stable noise, sets a filtering notch in a frequency band in which noise is concentrated, weakens the stable noise of the frequency band, adopts a time domain threshold suppression strategy to the non-stable noise, sets a signal amplitude threshold, judges the voice signal to be the non-stable noise when the voice signal exceeds the threshold, and suppresses the voice signal through a threshold filter.
The time domain feature extraction and analysis comprises dividing the voice signal after directional pickup into one frame according to 10ms, calculating the amplitude peak value and the peak duration of each frame of signal, and judging the voice signal as unsteady noise feature (such as sudden door closing sound) if the amplitude peak value of a certain frame of signal suddenly exceeds more than 3 times of the average amplitude of the previous 5 frames and the peak duration is shorter than 200 ms. And extracting and analyzing frequency domain characteristics, namely carrying out Fourier transformation on each frame of signal, counting energy values in each frequency band, calculating the energy fluctuation amplitude of continuous 10 frames of signals in the same frequency band, and judging the frequency band as the steady-state noise characteristics if the energy fluctuation amplitude of continuous 10 frames in a certain frequency band is less than 10 percent and the frequency band corresponds to common steady-state noise frequency (such as the low frequency band of air conditioner base noise). Noise classification and targeted suppression, namely, starting a frequency domain notch filter module for signals judged to be stationary noise, setting a notch point in a frequency band in which noise is concentrated, weakening noise energy of the frequency band through a filter circuit, and reserving voice signals of other frequency bands, starting time domain threshold suppression for signals judged to be non-stationary noise, setting a dynamic threshold (determined based on average amplitude of voice signals of the previous 5 frames), temporarily reducing gain of the frame signals through a threshold filter when the amplitude of the signals exceeds a threshold, suppressing burst noise, updating threshold parameters once every 500ms, and avoiding influencing normal voice.
The acoustic signal self-adaptive processing algorithm comprises a reverberation time delay estimation sub-algorithm and a dynamic filtering sub-algorithm, wherein the reverberation time delay estimation sub-algorithm firstly identifies a direct wave and a first reflected wave in the voice signal, the direct wave directly propagates to a microphone from a sound source, the first reflected wave propagates to the microphone after primary reflection, the direct wave is delayed, the reverberation time delay is obtained by calculating the time difference between the direct wave and the first reflected wave, and the dynamic filtering sub-algorithm adjusts the length of a filtering window according to the reverberation time delay, the window length needs to cover the duration of the reverberation tail, and the reverberation tail is completely filtered.
The voice masking neural network adopts a framework combining a two-way long-short-term memory network layer and an attention mechanism, wherein the two-way long-term memory network layer can simultaneously extract historical context characteristics and future context characteristics of a voice signal, extract time sequence rules of the voice signal from the beginning to the current moment in the forward direction, extract time sequence rules of the voice signal from the current moment to the end in the backward direction, capture the time sequence characteristics of the voice signal through two-way feature fusion, the attention mechanism focuses on a frequency band in which human voice is mainly distributed, and the voice masking neural network prioritizes the frequency band when distinguishing residual noise from the human voice by strengthening characteristic weights of the frequency band.
Building a network architecture:
an input layer, which is used for receiving the voice signals after reverberation filtering and converting the signals into Mel Frequency Cepstrum Coefficient (MFCC) feature vectors (covering key features of voice);
setting 2 layers of Bi-LSTM units, extracting historical context characteristics (such as the association of previous frame voice and current frame) from the beginning end to the end of a signal by the forward LSTM unit, extracting future context characteristics (such as the association of the next frame voice and the current frame) from the end to the beginning end of the signal by the backward LSTM unit, and obtaining complete time sequence characteristics by splicing and fusing the outputs of the two layers of units;
Focusing on a frequency band of the main distribution of human voice (the frequency band is critical to the voice intelligibility), giving higher weight to the feature vector of the frequency band through a weight matrix (the weight of other frequency bands is reduced), and strengthening the expression of key features;
and the output layer outputs a voice mask (a binary mask for distinguishing target voice from residual noise) through the full connection layer, wherein 1 represents the target voice and 0 represents the residual noise).
Network training and reasoning process:
The training phase comprises the steps of inputting a mixed signal into a network by adopting a mixed data set containing pure voice and various residual noise, and iteratively optimizing network parameters by taking the minimum cross entropy of a mask outputted by the network and a real mask (artificial labeling) as a target until a model converges;
and in the reasoning stage, the voice signals subjected to reverberation filtering are input into a trained network, a voice mask is output, the target voice components are reserved through point-by-point multiplication of the mask and the original signals, and the residual noise components are restrained.
The far-field pickup algorithm comprises a voice energy compensation sub-algorithm and a frequency equalization sub-algorithm, wherein the voice energy compensation sub-algorithm is based on the attenuation law of far-field voice transmission, namely the law that voice energy is gradually weakened along with the increase of transmission distance, firstly analyzes the sound source distance in the current recording environment, then calculates the energy attenuation degree according to the sound source distance, generates a corresponding energy compensation coefficient, improves the overall energy of the far-field voice through the energy compensation coefficient, and identifies a high-frequency region in the far-field voice according to the characteristic that the attenuation of a high-frequency band signal in far-field transmission is obvious, and adjusts the signal intensity of the high-frequency region through gain.
The voice feature perception lossless coding algorithm comprises a feature-driven coding redundancy optimization sub-algorithm, wherein the feature-driven coding redundancy optimization sub-algorithm comprises the steps of firstly analyzing energy distribution features of a target voice signal, identifying a voice energy concentration region and a mute region, wherein the voice energy concentration region is a core bearing part of voice content, the mute region is free of effective voice and only comprises weak background noise, a high-fidelity coding strategy is adopted for the voice energy concentration region, detail features of the signal are reserved, and a coding redundancy compression strategy is adopted for the mute region to remove redundant information in the signal.
The voice quality detection feedback step comprises the steps of detecting the target voice signal quality by adopting a voice quality perception evaluation algorithm, comprehensively judging the voice quality grade by analyzing the definition, naturalness and noise residual quantity indexes of the target voice, feeding the quality evaluation result back to a noise reduction link when the target voice signal quality does not reach a preset standard, repeatedly executing noise reduction operation until the detected voice quality reaches the preset standard, and entering the coding storage step.
In this embodiment, the method is applied to a medium conference room with 15m×10m× 3m, and the scene features are as follows:
Noise environments including stationary noise (ceiling air conditioner operation noise, sound pressure level 45 dB), non-stationary noise (occasional door closing sound, seat dragging sound, peak sound pressure level 65 dB);
Reverberation characteristics: the conference room wall is ordinary latex paint, floor carpeting, reverberation time (RT 60) about 0.5 seconds;
The recording requirement is that 3 participants are respectively positioned at the front 3m (a speaker A), the front left 4m (a participant B) and the front right 4.5m (a participant C) of the microphone array, and the 3-person voice is required to be recorded clearly without noise and reverberation interference;
Hardware configuration:
4 channels (channels 1-4), the sampling rate is 48kHz, the bit depth is 24 bits, and the channel spacing is 15cm;
the signal processing terminal is provided with a processor, supports floating point operation and has 8GB memory;
And the storage equipment is SSD solid state disk with capacity of 1TB.
Specific implementation steps
Multichannel array microphone configuration and original signal acquisition
After a microphone is started, a dynamic amplitude compensation sub-algorithm collects signals of 8 continuous sampling periods (20 ms per cycle) of each channel, wherein the amplitude of each channel 1-3 is stabilized at 0.8V, the amplitude of each channel 4 is only 0.5V due to low hardware sensitivity, and the difference of the hardware sensitivity is judged;
the phase calibration sub-algorithm selects a channel 2 as a reference channel (the phase fluctuation frequency is 0.2Hz, the fluctuation amplitude is 0.1rad, and the phase difference between the channels 1-4 and the channel 2 is the most stable), and integrates the calculated phase difference, wherein the accumulated delay amount of the channel 1 is 0.3rad, the accumulated delay amount of the channel 3 is 0.25rad, the accumulated delay amount of the channel 4 is 0.4rad, a reverse compensation signal (-0.3 rad, -0.25rad, -0.4 rad) is generated, the phase difference of each channel after compensation is stabilized within 0.05 rad;
The configuration of an anti-lost frame and sound breaking mechanism comprises the steps of presetting the capacity of a buffer pool to be 2 seconds (the single maximum transmission delay is 0.8 seconds and 2 times of delay time), adopting 'first-in first-out' storage, dividing a section of signal every 10ms, and adding an integrity mark containing 'signal length 10ms+CRC32 check code';
In the recording process, when the fluctuation of the transmission load is monitored, 1 single-segment signal frame loss (identification missing) occurs in the channel 3, a retransmission request is immediately triggered, and a corresponding 10ms signal supplement is called from a buffer pool without disconnection or sound breaking;
Raw signal acquisition, namely, continuously acquiring the raw multi-channel voice signals for 10 minutes, storing the signals in WAV format, and enabling the single-channel data volume to be about 5.76GB (48 kHz multiplied by 24 bits multiplied by 600 s).
Weighted kernel function blind source separation preprocessing
Signal-to-noise ratio feature analysis:
Time domain analysis, namely framing an original signal according to 50ms, and calculating energy fluctuation of each frame, wherein the frequency of 2 times of energy per second of an air conditioner noise section (without voice) exceeds the average value by 15 times, the fluctuation amplitude is 0.6V, and the frame is judged to be a low signal-to-noise ratio candidate frame, and the frequency of 3 times of fluctuation of each second of a voice section, the fluctuation amplitude is 0.2V, so that the frame is a high signal-to-noise ratio candidate frame;
performing frequency domain analysis, namely performing Fourier transformation on each frame, dividing into three sections of 20-200Hz (low frequency), 200-3400Hz (voice frequency band) and 3400-20000Hz (high frequency), wherein air conditioning noise is concentrated at 20-200Hz, and the frequency band voice energy/noise energy=0.6 (< 1) is judged to be a low signal-to-noise ratio frequency band;
The region division, namely superposing the time domain and frequency domain results, determining a '20-200 Hz frequency band plus time domain fluctuation frequent frame' as a low signal-to-noise ratio region, and the rest as a high signal-to-noise ratio region;
The weight calculation and separation are carried out by selecting S-shaped nonlinear weighting function, the noise energy ratio of the low signal-to-noise ratio area is 72%, the output weight is 1.8, the voice energy ratio of the high signal-to-noise ratio area is 85%, the output weight is 0.9, and the initial separation matrix of the independent component analysis model Learning step size η=0.2 (initial mutual information) After 6 iterations, the mutual information differenceStopping iteration (less than or equal to 0.01), and separating to obtain 3 paths of target voice signal components (corresponding to A, B, C) and 1 path of interference signal components (including air conditioning noise, door closing sound and reverberation).
Target-oriented adaptive cascading noise reduction execution
The directional pickup is locked, namely an azimuth interval corresponding to an adjustable pickup range is preset, namely a speaker A (0 degree +/-10 degrees), a participant B (-30 degrees +/-10 degrees) and a participant C (35 degrees +/-10 degrees), a phase difference delta phi=0.5 rad between a channel 1 and a channel 2 is acquired by a spatial spectrum estimation algorithm, a channel distance d=0.15 m, and a voice main frequency is acquired(Wavelength λ=0.34 m), substituting the formula θ=arcsin (ΔΦLambda/(2pi d)), calculating azimuth angle 1.2 degrees of A (within 0 degree + -10 degrees), azimuth angle-28 degrees of B, azimuth angle 32 degrees of C, all falling into corresponding interval, improving gain of three channels to 1.2 times, and suppressing interference outside the interval;
Self-adaptive multi-mode noise reduction:
Time domain analysis, namely, a peak value of a door closing sound frame amplitude is 1.2V (average 0.3V of the previous 5 frames is more than 3 times), the peak value lasts 150ms (< 200 ms), the door closing sound frame amplitude is judged to be unsteady noise, a dynamic threshold is set to be 0.6V, the gain is reduced to 0.5 times when the door closing sound frame amplitude exceeds the threshold, and the noise attenuation is 32dB;
the frequency domain analysis comprises that the air conditioning noise is 20-200Hz frequency band, the continuous 10 frames of energy fluctuation is 8% (< 10%), the air conditioning noise is judged to be steady-state noise, the notch frequency is 20-200Hz, and the attenuation is 28dB;
acoustic signal adaptive processing:
reverberation time delay estimation, namely, identifying direct wave start time stamp =0.2 S, first reflected waveCalculating the reverberation delay τ=0.4s=0.6 s;
dynamic filtering, window length l=2×0.4=0.8 s, cut-off frequency The tail of the reverberation after filtering is shortened from 0.5 second to 0.1 second, and the reverberation suppression amount is 22dB;
Voice masking neural network and far field pickup:
The neural network is trained by adopting a data set of 'pure voice (10 ten thousand conference voices) +residual noise (5 ten thousand air conditioner/reverberation noise)', 64 units are arranged on a Bi-LSTM layer, an attention mechanism focuses on a 200-3400Hz frequency band, the accuracy rate of an output mask is 92% during reasoning, and the residual noise is distinguished from a target voice;
far field pickup, namely calculating the distance between A and a microphone by TDOA to obtain an energy attenuation coefficient alpha= (1/3) 2 approximately equal to 0.11, compensating coefficient G=9, increasing the energy from 0.2V to 1.8V, seriously attenuating the energy in a high frequency band (2000-3400 Hz), setting a gain coefficient to 1.8, increasing the energy proportion of the high frequency band from 15% to 30% after the energy attenuation, and increasing the voice definition by 35%.
S4, speech feature perception lossless coding and storage
Feature extraction:
pitch period: 120ms (vowel segment) for A, 110ms for B, 130ms for C, labeled as energy concentration region;
A mute section, namely, the energy of a non-voice time frame is 0.02V (1/15 of the average energy of 0.3V), and the mute section is judged;
Differential coding:
the energy concentration area is used for coding and sampling the 24 bits of precision, reserving the characteristics of fundamental tone and harmonic wave and grading 3.8 of sound quality PESQ;
the sampling precision is reduced to 16 bits, the compression ratio is 1:2.5, and the storage occupation is reduced by 60%;
Storing 3 paths of coded voice as MP3 format (lossless compression) according to 'timestamp + speaker identification', and generating a noise-reducing recording file by 120MB of total storage capacity of 10 minutes recording
Implementation effect verification
The signal quality is that the signal-to-noise ratio of the processed voice is improved from 12dB to 48dB, and the intelligibility is improved from 65% to 98%;
integrity, namely, no frame loss and sound breaking are carried out in the whole process, and the recording integrity is 100%;
and compared with general lossless coding, the storage efficiency is reduced by 45%, and the tone quality and efficiency are both considered.
The present invention is not limited in any way by the above-described preferred embodiments, but is not limited to the above-described preferred embodiments, and any person skilled in the art will appreciate that the present invention can be embodied in the form of a program for carrying out the method of the present invention, while the above disclosure is directed to equivalent embodiments capable of being altered or modified in a slight manner, any and all concise modifications, equivalent variations and alterations of the above embodiments are still within the scope of the present disclosure, all as may be made without departing from the scope of the present disclosure.
Claims (12)
1. The array microphone noise reduction recording method based on cascade noise reduction and blind source separation is characterized by comprising the following steps of:
S1, configuring a multichannel array microphone, wherein the multichannel array microphone maintains the amplitude and phase collaborative consistency of channel recording signals through a collaborative signal processing algorithm, monitors the channel recording process through an anti-frame-loss sound breaking mechanism, and collects original multichannel voice signals in the environment;
S2, preprocessing the original multi-channel voice signal based on a weighted kernel function blind source separation algorithm, firstly extracting signal-to-noise ratio distribution characteristics, giving kernel function weight, and then decoupling the original multi-channel voice signal through an independent component analysis model to obtain a target voice signal component and an interference signal component;
S3, executing a target-oriented adaptive cascade noise reduction model on the target voice signal component, namely firstly locking a voice signal sent by a speaker through a directional pickup technology based on a preset adjustable pickup range parameter, then adopting an adaptive multi-mode noise reduction algorithm to inhibit stable noise and unsteady noise in the voice signal, then adopting an acoustic signal adaptive processing algorithm to filter voice reverberation tail in the voice signal after noise reduction, finally combining a voice mask neural network and a far-field pickup algorithm, distinguishing residual noise and target voice in the voice signal output after reverberation filtering through the voice mask neural network, calculating energy expectation of the target voice, and enhancing the target voice in a far-field environment based on the far-field pickup algorithm;
And S4, adopting a voice characteristic perception lossless coding algorithm to the target voice signal, namely extracting the pitch period and harmonic distribution characteristics of the target voice signal, adopting a high-fidelity coding strategy to a voice energy concentration area, and storing the coded target voice signal to generate a noise reduction recording process file.
2. The method of claim 1, wherein in S1, the collaborative signal processing algorithm comprises a dynamic amplitude compensation sub-algorithm and a phase calibration sub-algorithm, wherein the dynamic amplitude compensation sub-algorithm collects amplitude differences of channel recording signals in real time, analyzes reasons generated by the differences, regenerates dynamic proportionality coefficients, adjusts signal gains of amplitude channels through the dynamic proportionality coefficients, presets reference channels, calculates phase delay amounts of other channels and the reference channels in real time based on phases of the reference channels, generates reverse compensation signals according to the phase delay amounts, and superimposes the reverse compensation signals on corresponding channels.
3. The method of claim 1, wherein in S1, the frame loss prevention and sound breaking mechanism comprises a signal buffer pool and a frame loss detection retransmission sub-algorithm, wherein the signal buffer pool is provided with a buffer space, the buffer space temporarily stores real-time recording signals in continuous time, the frame loss detection retransmission sub-algorithm monitors frame loss conditions in the signal transmission process in real time, judges the frame loss conditions by comparing signal integrity identifiers before and after transmission, and immediately triggers a retransmission request when detecting that the frame loss phenomenon affects recording continuity, and retrieves non-frame loss signals stored before a frame loss period from the signal buffer pool to supplement the transmission link.
4. The method of claim 1, wherein in S2, the weighting kernel function blind source separation algorithm is characterized in that the signal-to-noise ratio distribution characteristics of the original multi-channel voice signals are subjected to multi-dimensional analysis, fluctuation frequency and amplitude of the energy of the original multi-channel voice signals are monitored in a time domain, the energy ratio of the original multi-channel voice signals to noise signals in a frequency domain is calculated, a low signal-to-noise ratio area and a high signal-to-noise ratio area are divided through fusion judgment of the time domain and the frequency domain characteristics, a nonlinear weighting function is adopted to calculate kernel function weights, and the weight values of the functions are dynamically adjusted in combination with the noise energy duty ratio in the corresponding area, so that the weighting distribution is differentiated based on the actual characteristics of the original multi-channel voice signals.
5. The method according to claim 1, wherein in the step S2, the independent component analysis model integrates a gradient descent optimization sub-algorithm, the gradient descent optimization sub-algorithm uses mutual information of signals after separation reduction as an objective function, reduces the relevance between a target voice signal component and an interference signal component through optimization, adjusts the independent component analysis model parameters according to the currently calculated mutual information value during iteration, gradually optimizes the separation effect, and stops iteration when the mutual information difference obtained by two adjacent iteration calculation is reduced to a preset threshold.
6. The method of claim 1, wherein in step S3, the directional pickup technique is combined with a spatial spectrum estimation algorithm, the spatial spectrum estimation algorithm collects phase information of channel voice signals first, calculates an attitude of a source of the channel voice signals by comparing phase differences among channels, compares the attitude with an azimuth range corresponding to the preset adjustable pickup range, determines a target speaker voice signal when the attitude is in the azimuth range, and locks the target speaker voice signal by enhancing channel signal gain.
7. The method of claim 1, wherein in S3, the adaptive multi-modal noise reduction algorithm comprises a noise type classification sub-algorithm, wherein the noise type classification sub-algorithm firstly extracts time domain and frequency domain features of the voice signal, generates a feature analysis result, separates noise into stationary noise and non-stationary noise according to the feature analysis result, sets a filtering notch in a frequency band of a noise set by adopting a frequency domain notch filtering strategy for the stationary noise, weakens the stationary noise of the frequency band, sets a signal amplitude threshold for the non-stationary noise by adopting a time domain threshold suppression strategy, determines the non-stationary noise when the voice signal exceeds a threshold, and suppresses the voice signal by adopting a threshold filter.
8. The method of claim 1, wherein in S3, the acoustic signal adaptive processing algorithm comprises a reverberation time delay estimation sub-algorithm and a dynamic filtering sub-algorithm, wherein the reverberation time delay estimation sub-algorithm firstly identifies a direct arrival wave and a first reflected wave in the voice signal, the direct wave directly propagates from a sound source to a microphone, the first reflected wave propagates to the microphone after primary reflection and lags behind the direct wave, a reverberation time delay is obtained by calculating a time difference between the direct wave and the first reflected wave, and the dynamic filtering sub-algorithm adjusts a length of a filtering window according to the reverberation time delay, wherein the window length needs to cover a duration of a reverberation tail, and the reverberation tail is completely filtered.
9. The method of claim 1, wherein in S3, the voice masking neural network adopts a structure of combining a two-way long-short-term memory network layer and an attention mechanism, wherein the two-way long-term memory network layer can simultaneously extract historical context characteristics and future context characteristics of a voice signal, extract time sequence rules of the voice signal from the beginning to the current moment in a forward direction, extract time sequence rules of the voice signal from the current moment to the end in a backward direction, capture the time sequence characteristics of the voice signal through two-way feature fusion, and the attention mechanism focuses on a frequency band in which human voice is mainly distributed, so that the voice masking neural network prioritizes the frequency band when distinguishing residual noise from the human voice by strengthening feature weights of the frequency band.
10. The method of claim 1, wherein in S3, the far-field pickup algorithm comprises a voice energy compensation sub-algorithm and a frequency equalization sub-algorithm, wherein the voice energy compensation sub-algorithm is based on a decay law of far-field voice transmission, namely a law that the voice energy is gradually weakened along with the increase of a transmission distance, firstly analyzes a sound source distance in a current recording environment, then calculates an energy decay degree according to the sound source distance, generates a corresponding energy compensation coefficient, improves the overall energy of the far-field voice through the energy compensation coefficient, and identifies a high-frequency region in the far-field voice according to the characteristic that the decay of a high-frequency band signal in the far-field transmission is obvious, and adjusts the signal intensity of the high-frequency region through gain.
11. The method of claim 1, wherein in S4, the speech feature aware lossless coding algorithm comprises a feature driven coding redundancy optimization sub-algorithm, wherein the feature driven coding redundancy optimization sub-algorithm is used for analyzing the energy distribution feature of a target speech signal first, identifying a speech energy concentration region and a mute region, wherein the speech energy concentration region is a core bearing part of speech content, the mute region has no effective speech and only comprises weak background noise, and the detail feature of the signal is reserved by adopting a high-fidelity coding strategy for the speech energy concentration region, and the redundancy information in the signal is removed by adopting a coding redundancy compression strategy for the mute region.
12. The method according to claim 1, wherein the voice quality detection feedback step comprises detecting the target voice signal quality by using a voice quality perception evaluation algorithm, comprehensively judging the voice quality level by analyzing the target voice definition, naturalness and noise residual index, feeding the quality evaluation result back to the noise reduction link when the target voice signal quality does not reach the preset standard, repeatedly executing the noise reduction operation until the detected voice quality reaches the preset standard, and entering the encoding storage step.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202511338173.9A CN121237113A (en) | 2025-09-18 | 2025-09-18 | A noise reduction recording method for array microphones based on cascaded noise reduction and blind source separation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202511338173.9A CN121237113A (en) | 2025-09-18 | 2025-09-18 | A noise reduction recording method for array microphones based on cascaded noise reduction and blind source separation |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN121237113A true CN121237113A (en) | 2025-12-30 |
Family
ID=98155190
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202511338173.9A Pending CN121237113A (en) | 2025-09-18 | 2025-09-18 | A noise reduction recording method for array microphones based on cascaded noise reduction and blind source separation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN121237113A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN121725807A (en) * | 2026-02-26 | 2026-03-24 | 深圳市长丰影像器材有限公司 | Audio processing methods, devices, and systems based on multimodal noise reduction and sound source separation |
-
2025
- 2025-09-18 CN CN202511338173.9A patent/CN121237113A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN121725807A (en) * | 2026-02-26 | 2026-03-24 | 深圳市长丰影像器材有限公司 | Audio processing methods, devices, and systems based on multimodal noise reduction and sound source separation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Akbari et al. | Lip2audspec: Speech reconstruction from silent lip movements video | |
| CN110648678B (en) | Scene identification method and system for conference with multiple microphones | |
| EP3719798B1 (en) | Voiceprint recognition method and device based on memorability bottleneck feature | |
| CN111508498B (en) | Conversational speech recognition method, conversational speech recognition system, electronic device, and storage medium | |
| Kingsbury et al. | Robust speech recognition using the modulation spectrogram | |
| Stern et al. | Hearing is believing: Biologically inspired methods for robust automatic speech recognition | |
| Ibrahim | Preprocessing technique in automatic speech recognition for human computer interaction: an overview | |
| EP2363852B1 (en) | Computer-based method and system of assessing intelligibility of speech represented by a speech signal | |
| CN113192535B (en) | Voice keyword retrieval method, system and electronic device | |
| US20070118374A1 (en) | Method for generating closed captions | |
| Hu et al. | Segregation of unvoiced speech from nonspeech interference | |
| JPH06332492A (en) | VOICE DETECTION METHOD AND DETECTION DEVICE | |
| US20090271198A1 (en) | Producing phonitos based on feature vectors | |
| US20070118364A1 (en) | System for generating closed captions | |
| CN111724809A (en) | Vocoder implementation method and device based on variational self-encoder | |
| US20240249714A1 (en) | Multi-encoder end-to-end automatic speech recognition (asr) for joint modeling of multiple input devices | |
| Richardson et al. | Channel compensation for speaker recognition using MAP adapted PLDA and denoising DNNs | |
| CN121237113A (en) | A noise reduction recording method for array microphones based on cascaded noise reduction and blind source separation | |
| CN120690210A (en) | A real-time monitoring method and system for multi-channel output status based on voiceprint recognition | |
| US10229686B2 (en) | Methods and apparatus for speech segmentation using multiple metadata | |
| CN120913580B (en) | Language data preprocessing method for multilingual complex scene | |
| CN120766655B (en) | Cross-language AI voiceprint cloning method, system and storage medium | |
| CN119851671B (en) | Speech enhancement training method, device, equipment and medium based on speaker perception | |
| Tzudir et al. | Low-resource dialect identification in Ao using noise robust mean Hilbert envelope coefficients | |
| Kaur et al. | Optimizing feature extraction techniques constituting phone based modelling on connected words for Punjabi automatic speech recognition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |