CN103559888B - Based on non-negative low-rank and the sound enhancement method of sparse matrix decomposition principle - Google Patents
Based on non-negative low-rank and the sound enhancement method of sparse matrix decomposition principle Download PDFInfo
- Publication number
- CN103559888B CN103559888B CN201310548773.9A CN201310548773A CN103559888B CN 103559888 B CN103559888 B CN 103559888B CN 201310548773 A CN201310548773 A CN 201310548773A CN 103559888 B CN103559888 B CN 103559888B
- Authority
- CN
- China
- Prior art keywords
- matrix
- rank
- noisy speech
- low
- sparse matrix
- 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.)
- Expired - Fee Related
Links
Landscapes
- Circuit For Audible Band Transducer (AREA)
- Noise Elimination (AREA)
Abstract
The invention discloses a kind of based on non-negative low-rank with the sound enhancement method of sparse matrix decomposition principle.First noisy speech signal is smoothed by the method, framing and discrete Fourier transform, obtains noisy speech frequency spectrum；Then the noisy speech amplitude spectrum of every frame is sequentially arranged one noisy speech time-frequency matrix of composition as column vector, by noisy speech time-frequency matrix is carried out non-negative low-rank and sparse matrix decomposition, it is thus achieved that the low-rank of non-negative and sparse matrix；Utilize sparse matrix and noisy speech phase reconstruction to strengthen voice spectrum, obtain the enhancing voice of forms of time and space finally by inverse Fourier transform.The present invention is strong to noise adaptation, be made without end-point detection and model training, parameter the most easily regulate, and strong noise environment performance is good, has good application prospect.
Description
Technical field
The present invention relates to signal processing field, it is adaptable to the noise suppressed of noisy speech, be based particularly on non-negative low-rank and sparse square
The sound enhancement method of battle array decomposition principle.
Background technology
Voice signal is Human communication's information means the most natural, maximally effective.Along with the mankind enter the information age, in the urgent need to
Use the voice processing technology of advanced person to promote that human society is intelligent.As far back as 2000, Bill Gates was just it is proposed that " following 10
Year is the epoch of voice ".Recent years, successively release intelligent sound service, intelligence along with companies such as Fructus Mali pumilae, Google, Microsofts
Voice industry has become as the new industry in areas of information technology, and user cognition degree and market scale the most gradually expand.Especially
It is that the smart mobile phone that Fructus Mali pumilae is released recently has voice assistant's function, and the voice " cloud " that University of Science and Technology news fly is released, and makes intelligence
Voice technology can face more wide application.But, inevitably by from ring around in voice communication and application process
Border, communication media and the interference of inside communication equipment noise, had a strong impact on the actual application of intelligent sound technology.
Speech enhan-cement is the effective technology solving sound pollution.Speech enhan-cement is by the suppression noise interference to voice so that strengthen
The voice signal processed is minimum with the distortion between original clean voice signal.Come in the past few decades, emerged in large numbers many speech enhan-cement
Algorithm, typical algorithm includes spectrum-subtraction, based on spectral amplitude least mean-square error, subspace method, wavelet de-noising method.In noise
In the environment of the highest, speech enhan-cement has been obtained for effectively solving.But, due to the multiformity of noise and voice in natural environment
The complexity of signal itself, voice enhancement algorithm is different according to the difference of applied environment, and this makes its research work difficulty very big,
The Speech Enhancement problem of very noisy and multiple noise circumstance is solved the most very well.
In existing voice enhancement algorithm, many methods attempt to use the pdf model of voice signal and noise signal to come
Remove noise signal in big degree, but research in recent years shows that certain single distribution can not be applicable to all of voice or make an uproar
Sound, needs mathematical model and model algorithm for estimating the most flexibly are to adapt to the feature of signal self.Additionally, increase at existing voice
In strong algorithms, Noise Estimation is the early stage indispensability work of voice enhancement algorithm.By Noise Estimation can obtain noise power spectrum and
The prior weight of voice signal, the improvement to speech enhan-cement effect is most important.Existing sound enhancement method passes through end-speech
The voice signal collected is divided into noise segment and noisy speech section by some detection, utilizes noise segment to estimate and updates Noise Estimation amount,
But this is a kind of suboptimal estimation mode, the instantaneous noise of noise segment and noisy speech section being not exclusively consistent in reality, therefore,
This noise estimation method always brings error, moreover existing voice end-point detection technology is at low signal-to-noise ratio and nonstationary noise environment
Under the most immature, easily cause erroneous judgement, can cause voice exists the biggest residual noise.
In recent years compressive sensing theory research shows, many actual observed quantities can be attributed to a low-rank component and sparse point
The pattern that amount is added, by low-rank and the sparse matrix decomposition of matrix, can recover former from big noise or exceptional value contamination data
Beginning data message.The low-rank of matrix and sparse matrix decomposition have been used to image enhaucament, video object detection, data mining etc. and are permitted
Many sciemtifec and technical spheres.
Stationary random noise and periodic noise are modal two kinds of noise class.Stationary random noise's single order and second-order statistic
Describing its stochastic process, its average and auto-correlation function are unrelated with the time, owing to the Fourier of stochastic signal auto-correlation function becomes
Changing is power spectrum, therefore the time-frequency matrix of stationary random noise be an order number be the low-rank matrix of 1.Equally, if noise is week
Phase property noise, owing to its time-frequency matrix only has value at some fixed frequency, its rectangular array vector has stronger dependency, must
Also it is so a low-rank matrix.
In sum, the time-frequency rectangular array vector of background noise has the strongest dependency, and therefore the time-frequency matrix of noise has low
Order.Relatively for background noise, speech source signals value in major part time frequency point is zero or close to zero, only minority
Sample point value is relatively big, so speech source signals has certain openness, applicable sparse matrix describes.Therefore, it can
Consider to use for reference the low-rank of matrix and sparse resolution theory solves Speech Enhancement problem.Chinese patent disclose a kind of based on low-rank with
The single channel of sparse matrix decomposition is made an uproar separation method (publication number: CN102915742A) without supervision language.The method first by
Noisy speech time domain waveform is transformed to time-frequency domain thus obtains the amplitude spectrum of noisy speech by Short Time Fourier Transform；Utilize low-rank with
The amplitude spectrum of noisy speech is decomposed into noise amplitude spectrum, voice amplitudes spectrum and residual noise amplitude spectrum three by sparse matrix decomposition algorithm
Person's sum；Finally, the voice time domain waveform that inverse Fourier transform reconstructs from the amplitude spectrum of voice in short-term is utilized.The method
Deficiency is low-rank and sparse matrix decomposition not to be added nonnegativity restriction, is easily caused isolated from noisy speech amplitude spectrum
Voice amplitudes spectrum containing negative result.And the amplitude spectrum of reality is non-negative physical quantity, negative value phenomenon should not occur.Negative value width
Degree spectrum not only causes resolution error, and can produce the music noise that human ear is felt to feel bad, thus affects phonetic hearing quality.
The present invention devises a kind of based on non-negative low-rank with the sound enhancement method of sparse matrix decomposition principle, and the method uses non-negative
Low-rank and sparse matrix decomposition principle decompose noisy speech amplitude spectrum, and the voice amplitudes spectrum that decomposition can be made to obtain meets nonnegativity, has
Effect improves low-rank and sparse matrix decomposition effect.The method has strong robustness, be made without end-point detection and parameter is few
The advantages such as easy regulation, are suitable for the speech enhan-cement task under strong noise environment.
Summary of the invention
The technical problem to be solved is to provide a kind of based on non-negative low-rank and the speech enhan-cement side of sparse matrix decomposition principle
Method, by introducing noise and the low-rank of voice and sparse constraint and nonnegativity restriction carries out low-rank and sparse in time-frequency domain
Matrix decomposition, it is achieved the language of noisy speech is made an uproar separation.
The present invention takes techniques below scheme, based on non-negative low-rank and the sound enhancement method of sparse matrix decomposition principle, uses non-negative
Low-rank and sparse matrix decomposition method isolate voice signal from noisy speech, and implementation step is as follows:
(1) discrete noisy speech signal being carried out pretreatment, pretreatment includes signal smoothing and framing；
(2) noisy speech signal after framing is carried out discrete Fourier transform, obtain noisy speech frequency spectrum；
(3) in a frequency domain, using the spectrum amplitude of every frame voice as column vector, it is sequentially arranged, by several speech frame structures
Become noisy speech time-frequency matrix；
(4) utilize non-negative low-rank and sparse matrix decomposition algorithm that noisy speech time-frequency matrix is decomposed, it is thus achieved that the low-rank square of non-negative
Battle array and sparse matrix；Decomposing expression formula is:
Y=L+S+E meets rank (L)≤r, | | S | |0≤h,L≥0,S≥0；
Wherein: Y is noisy speech time-frequency matrix；L is low-rank matrix, the amplitude spectrum of corresponding noise；S is sparse matrix, corresponding language
The amplitude spectrum of sound, | | S | |0Representing the non-zero element number that sparse matrix S contains, the order of rank (L) representing matrix L, E is residual matrix,
R and h represents low-rank and sparse constraint upper limit parameter；
(5) the phase spectrum reconstruct utilizing sparse matrix S and noisy speech strengthens voice spectrum, then by inverse Fourier transform,
Enhancing voice to forms of time and space.
The processing procedure that discrete noisy speech signal carries out in described step (1) pretreatment is:
(1) P point arest neighbors signal average is used to carry out signal smoothing, in order to the amplitude wave-shape of smooth noisy speech；
(2) to noisy speech signal framing, the window function that framing uses is Hamming window, and a length of 200 points of window, each interframe moves
Overlap to count be 80 points.
The step calculating low-rank matrix L and sparse matrix S is as follows:
(1) initialize: Y0=Y；L0=S0=[0]N×K；
Iterations initial value i=1；Maximum iteration time imax=103；Relative error threshold value δ=10-3；
(2) use NMF renewal low-rank matrix: (W, H)=NMF (Yi-1), Li=WH；W∈RN×r, H ∈ Rr×K；
NMF represent Non-negative Matrix Factorization, NMF represent Non-negative Matrix Factorization, W and H be order be the NMF decomposition result of r, the survey of NMF
Degree function selects Itakura-Saito to estimate；
(3) Soft-thresholding operator is used to update sparse matrix: Si=(Yi-1-Li+Si-1＞ λ) (Yi-1-Li+Si-1-λ)；
Wherein: symbolRepresenting matrix correspondence position element product, λ is thresholding constant；λ is relevant with noise level, it is recommended that value
λ=σ, wherein σ is the mean square deviation of noise；
(4) superposition matrix: Y is updatedi=Li+Si；
(5) if i reach maximum iteration time i=imax orStop iteration, the estimated value of output L and S
LiAnd Si；Otherwise jump to step (2), i=i+1；Continue executing with iterative process.
The phase spectrum utilizing sparse matrix and noisy speech in described step (5) reconstructs and strengthens voice spectrum:
Wherein: ∠ Y (n, k) is the spectral phase of noisy speech, and S is sparse matrix, S (n, k) is sparse matrix spectrum amplitude angle value,For the enhancing voice spectrum of reconstruct, n is time frame index, and k is frequency indices.
The sound enhancement method that the present invention provides, by non-negative low-rank and sparse matrix decomposition, can make the low-rank matrix that decomposition obtains
It it is all nonnegative value with the element in sparse matrix.The method is made without end-point detection and model training, have strong robustness,
The advantages such as parameter the most easily regulation, the speech enhan-cement task being particularly suitable under strong noise environment.
Accompanying drawing explanation
Fig. 1 is the speech-enhancement system block diagram of the present invention.
Detailed description of the invention
In conjunction with accompanying drawing, the invention will be further described, sees Fig. 1, based on non-negative low-rank and the voice of sparse matrix decomposition principle
Enhancement Method, including step in detail below:
1) noisy speech signal y (t) is carried out pretreatment 101；Pretreatment 101 stage includes signal smoothing and framing so that it is be easy to
Subsequent processes.Signal smoothing refer to use y (t) P point arest neighbors signal average to calculate noisy speech signal currency, in order to smooth
The amplitude wave-shape of noisy speech signal.In the present invention, the value of P is 3, i.e.Framing uses
Window function be Hamming window, a length of 200 points of window, it is 80 points that the overlap that each interframe moves is counted；
2) noisy speech signal after framing is carried out DFT (discrete Fourier transform) 102, obtain signal spectrum, signal spectrum bag
Include the amplitude spectrum 104 of signal | Y (n, k) | and phase spectrum 103 ∠ Y (n, k).Wherein n represents frame index, n=1,2 ..., N；K represents frequency
Rate indexes, k=1, and 2 ..., k；N is total time frame number；K is that Fourier transformation is counted；
3) in a frequency domain, being arranged as column vector order by the amplitude spectrum 104 of every frame voice, so several speech frames just constitute one
The noisy time-frequency matrix Y of individual N × K.
4) noisy time-frequency matrix Y is carried out NLSMD (non-negative low-rank and sparse matrix decomposition) 105, calculate non-negative low-rank matrix L and non-
Negative sparse matrix S.
Y=L+S+E meets rank (L)≤r, | | S | |0≤h,L≥0,S≥0
Here the amplitude spectrum of L correspondence noise；The amplitude spectrum of S correspondence voice；||S||0The non-zero element number that representing matrix S contains；E
For residual matrix；The order of rank (L) representing matrix L；R and h represents low-rank and sparse constraint upper limit parameter；Through contrast test, r
Preferable noise reduction is obtained during value 1～3.
The calculating process of NLSMD (non-negative low-rank and sparse matrix decomposition) 105 is as follows:
1. initialize: Y0=Y；L0=S0=[0]N×K；Iterations i=1；Maximum iteration time imax=103；Relative error threshold
Value δ=10-3；
2. use Non-negative Matrix Factorization renewal low-rank matrix: (W, H)=NMF (Yi-1), Li=WH；W∈RN×r, H ∈ Rr×K；
Wherein: LiThe estimated value through ith iteration L, NMF represent Non-negative Matrix Factorization, W and H be order be r NMF decompose
As a result, owing to W and H is nonnegative value, so LiInevitable is also nonnegative matrix.The measure function of NMF algorithm can select Euclidean
Distance, Kullback-Leibler estimate and estimate with Itakura-Saito.Through contrast test, use Itakura-Saito to estimate and obtain
Good effect.Therefore, the present invention uses the NMF method estimated based on Itakura-Saito to calculate L.
3. Soft-thresholding operator is used to update sparse matrix Si: Si=(Yi-1-Li+Si-1＞ λ) (Yi-1-Li+ Si-1-λ)；
Wherein: symbolRepresenting matrix correspondence position element product, λ is threshold value, and the value of λ is relevant with noise intensity, pushes away
Recommending value λ=σ, wherein σ is noise mean square deviation.
4. superposition matrix: Y is updatedi=Li+Si；
If 5. i reach maximum iteration time i=imax orStop iteration, estimated value Li of output L and S
And Si；Otherwise jump to step 2., i=i+1, continue executing with iterative process；
5) sparse matrix S and the reconstruct of noisy speech phase spectrum is utilized to strengthen voice spectrum, owing to human ear is unwise to the phase information of sound
Sense, can with the phase place ∠ Y of noisy speech frequency spectrum (n, k) replaces the phase place strengthening voice, obtains strengthening the complex number spectrum of voice:
6) the complex number spectrum matrix strengthening voice is expanded into vector, it is carried out IDFT (inverse discrete Fourier transform) 106, obtain
The discrete time strengthening voice represents:
Wherein:Rectangular array vector temporally frame sequential is concatenated into the operation of one-dimensional vector by vec function representation.
Claims (1)
1. based on non-negative low-rank and the sound enhancement method of sparse matrix decomposition principle, by non-negative low-rank and sparse matrix decomposition method
Isolating voice signal from noisy speech, discrete noisy speech signal is carried out pretreatment, pretreatment includes signal smoothing and divides
Frame；Noisy speech signal after framing is carried out discrete Fourier transform, obtains noisy speech frequency spectrum；In a frequency domain, by every frame
The spectrum amplitude of voice, as column vector, is sequentially arranged, and is made up of noisy speech time-frequency matrix several speech frames；Profit
With non-negative low-rank and sparse matrix decomposition algorithm, noisy speech time-frequency matrix is decomposed, it is thus achieved that the low-rank matrix of non-negative and sparse
Matrix；Decomposing expression formula is:
Y=L+S+E meets rank (L)≤r, | | S | |0≤h,L≥0,S≥0；
Wherein: Y is noisy speech time-frequency matrix；L is low-rank matrix, the amplitude spectrum of corresponding noise；S is sparse matrix, corresponding language
The amplitude spectrum of sound, | | S | |0Representing the non-zero element number that sparse matrix S contains, the order of rank (L) representing matrix L, E is residual matrix,
R and h represents low-rank and sparse constraint upper limit parameter；The phase spectrum reconstruct utilizing sparse matrix S and noisy speech strengthens voice spectrum,
Then by inverse Fourier transform, the enhancing voice of forms of time and space is obtained；It is characterized in that, calculate low-rank matrix L and sparse matrix
The step of S is as follows:
(1) initialize: Y0=Y；L0=S0=[0]N×K；
Iterations initial value i=1；Maximum iteration time imax=103；Relative error threshold value δ=10-3；
(2) use NMF renewal low-rank matrix: (W, H)=NMF (Yi-1), Li=WH；W∈RN×r, H ∈ Rr×K；
NMF represent Non-negative Matrix Factorization, W and H be order be the NMF decomposition result of r, the measure function of NMF selects Itakura-Saito
Estimate；
(3) Soft-thresholding operator is used to update sparse matrix: Si=(Yi-1-Li+Si-1＞ λ) (Yi-1-Li+Si-1-λ)；
Wherein: symbol representing matrix correspondence position element product, λ is thresholding constant；λ is relevant with noise level, it is recommended that value
λ=σ, wherein σ is the mean square deviation of noise；
(4) superposition matrix: Y is updatedi=Li+Si；
(5) if i reach maximum iteration time i=imax orStop iteration, the estimated value of output L and S
LiAnd Si；Otherwise jump to step (2), i=i+1；Continue executing with iterative process.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310548773.9A CN103559888B (en) | 2013-11-07 | 2013-11-07 | Based on non-negative low-rank and the sound enhancement method of sparse matrix decomposition principle |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310548773.9A CN103559888B (en) | 2013-11-07 | 2013-11-07 | Based on non-negative low-rank and the sound enhancement method of sparse matrix decomposition principle |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103559888A CN103559888A (en) | 2014-02-05 |
| CN103559888B true CN103559888B (en) | 2016-10-05 |
Family
ID=50014118
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310548773.9A Expired - Fee Related CN103559888B (en) | 2013-11-07 | 2013-11-07 | Based on non-negative low-rank and the sound enhancement method of sparse matrix decomposition principle |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103559888B (en) |
Families Citing this family (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104021797A (en) * | 2014-06-19 | 2014-09-03 | 南昌大学 | Voice signal enhancement method based on frequency domain sparse constraint |
| EP3010017A1 (en) * | 2014-10-14 | 2016-04-20 | Thomson Licensing | Method and apparatus for separating speech data from background data in audio communication |
| CN104751855A (en) * | 2014-11-25 | 2015-07-01 | 北京理工大学 | Speech enhancement method in music background based on non-negative matrix factorization |
| CN104505100B (en) * | 2015-01-06 | 2017-12-12 | 中国人民解放军理工大学 | A kind of unsupervised sound enhancement method based on robust Non-negative Matrix Factorization and data fusion |
| US10528147B2 (en) | 2017-03-06 | 2020-01-07 | Microsoft Technology Licensing, Llc | Ultrasonic based gesture recognition |
| US10276179B2 (en) | 2017-03-06 | 2019-04-30 | Microsoft Technology Licensing, Llc | Speech enhancement with low-order non-negative matrix factorization |
| CN108573711A (en) * | 2017-03-09 | 2018-09-25 | 中国科学院声学研究所 | A kind of single microphone speech separating method based on NMF algorithms |
| US10984315B2 (en) | 2017-04-28 | 2021-04-20 | Microsoft Technology Licensing, Llc | Learning-based noise reduction in data produced by a network of sensors, such as one incorporated into loose-fitting clothing worn by a person |
| CN107528648A (en) * | 2017-10-23 | 2017-12-29 | 北京邮电大学 | A kind of blind frequency spectrum sensing method and device based on low-rank sparse matrix decomposition |
| CN108899045A (en) * | 2018-06-29 | 2018-11-27 | 中国航空无线电电子研究所 | Subspace sound enhancement method based on constraint low-rank and sparse decomposition |
| CN108986834B (en) * | 2018-08-22 | 2023-04-07 | 中国人民解放军陆军工程大学 | Bone conduction voice blind enhancement method based on codec framework and recurrent neural network |
| CN109036452A (en) * | 2018-09-05 | 2018-12-18 | 北京邮电大学 | A kind of voice information processing method, device, electronic equipment and storage medium |
| CN109215671B (en) * | 2018-11-08 | 2022-12-02 | 西安电子科技大学 | Voice enhancement system and method based on MFrSRRPCA algorithm |
| WO2020113575A1 (en) * | 2018-12-07 | 2020-06-11 | 广东省智能制造研究所 | Sound classification method, device and medium based on semi-nonnegative materix factorization with constraint |
| CN109658944B (en) * | 2018-12-14 | 2020-08-07 | 中国电子科技集团公司第三研究所 | Helicopter acoustic signal enhancement method and device |
| CN111863014B (en) * | 2019-04-26 | 2024-09-17 | 北京嘀嘀无限科技发展有限公司 | Audio processing method, device, electronic equipment and readable storage medium |
| CN111312270B (en) * | 2020-02-10 | 2022-11-22 | 腾讯科技（深圳）有限公司 | Voice enhancement method and device, electronic equipment and computer readable storage medium |
| CN111402909B (en) * | 2020-03-02 | 2023-07-07 | 东华大学 | Speech enhancement method based on constant frequency domain transformation |
| CN114913874B (en) * | 2021-02-08 | 2024-06-18 | 北京小米移动软件有限公司 | Voice signal processing method and device, electronic equipment and storage medium |
| CN113129872B (en) * | 2021-04-06 | 2023-03-14 | 新疆大学 | Voice enhancement method based on deep compressed sensing |
| CN117935834B (en) * | 2024-03-12 | 2024-05-28 | 深圳市声优创科技有限公司 | Intelligent audio noise reduction method and equipment |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101441872A (en) * | 2007-11-19 | 2009-05-27 | 三菱电机株式会社 | Denoising acoustic signals using constrained non-negative matrix factorization |
| CN102855884A (en) * | 2012-09-11 | 2013-01-02 | 中国人民解放军理工大学 | Speech time scale modification method based on short-term continuous nonnegative matrix decomposition |
| CN102915742A (en) * | 2012-10-30 | 2013-02-06 | 中国人民解放军理工大学 | Single-channel monitor-free voice and noise separating method based on low-rank and sparse matrix decomposition |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20100111499A (en) * | 2009-04-07 | 2010-10-15 | 삼성전자주식회사 | Apparatus and method for extracting target sound from mixture sound |
| US8080724B2 (en) * | 2009-09-14 | 2011-12-20 | Electronics And Telecommunications Research Institute | Method and system for separating musical sound source without using sound source database |
-
2013
- 2013-11-07 CN CN201310548773.9A patent/CN103559888B/en not_active Expired - Fee Related
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101441872A (en) * | 2007-11-19 | 2009-05-27 | 三菱电机株式会社 | Denoising acoustic signals using constrained non-negative matrix factorization |
| CN102855884A (en) * | 2012-09-11 | 2013-01-02 | 中国人民解放军理工大学 | Speech time scale modification method based on short-term continuous nonnegative matrix decomposition |
| CN102915742A (en) * | 2012-10-30 | 2013-02-06 | 中国人民解放军理工大学 | Single-channel monitor-free voice and noise separating method based on low-rank and sparse matrix decomposition |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103559888A (en) | 2014-02-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103559888B (en) | Based on non-negative low-rank and the sound enhancement method of sparse matrix decomposition principle | |
| Sigg et al. | Speech enhancement using generative dictionary learning | |
| CN111445905B (en) | Mixed voice recognition network training method, mixed voice recognition method, device and storage medium | |
| CN107564513B (en) | Voice recognition method and device | |
| CN109524020B (en) | Speech enhancement processing method | |
| CN102915742B (en) | Single-channel monitor-free voice and noise separating method based on low-rank and sparse matrix decomposition | |
| CN104685562B (en) | Method and apparatus for reconstructing echo signal from noisy input signal | |
| Talmon et al. | Single-channel transient interference suppression with diffusion maps | |
| CN104505100B (en) | A kind of unsupervised sound enhancement method based on robust Non-negative Matrix Factorization and data fusion | |
| CN111508518B (en) | Single-channel speech enhancement method based on joint dictionary learning and sparse representation | |
| Mavaddaty et al. | A novel speech enhancement method by learnable sparse and low-rank decomposition and domain adaptation | |
| Wang | Multi-band multi-centroid clustering based permutation alignment for frequency-domain blind speech separation | |
| Li et al. | A multi-objective learning speech enhancement algorithm based on IRM post-processing with joint estimation of SCNN and TCNN | |
| JP2013068938A (en) | Signal processing apparatus, signal processing method, and computer program | |
| Nie et al. | Exploiting spectro-temporal structures using NMF for DNN-based supervised speech separation | |
| Kang et al. | NMF-based speech enhancement incorporating deep neural network. | |
| Xu et al. | Methods for learning adaptive dictionary in underdetermined speech separation | |
| JP5807914B2 (en) | Acoustic signal analyzing apparatus, method, and program | |
| Casebeer et al. | Deep tensor factorization for spatially-aware scene decomposition | |
| Narayanaswamy et al. | Audio source separation via multi-scale learning with dilated dense u-nets | |
| Fang et al. | Integrating statistical uncertainty into neural network-based speech enhancement | |
| CN102509268B (en) | Immune-clonal-selection-based nonsubsampled contourlet domain image denoising method | |
| Ben Messaoud et al. | Sparse representations for single channel speech enhancement based on voiced/unvoiced classification | |
| Zdunek | Improved convolutive and under-determined blind audio source separation with MRF smoothing | |
| Ben et al. | Chirp signal denoising based on convolution neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20161005 Termination date: 20171107 |