CN108962276A - A kind of speech separating method and device - Google Patents
A kind of speech separating method and device Download PDFInfo
- Publication number
- CN108962276A CN108962276A CN201810820474.9A CN201810820474A CN108962276A CN 108962276 A CN108962276 A CN 108962276A CN 201810820474 A CN201810820474 A CN 201810820474A CN 108962276 A CN108962276 A CN 108962276A
- Authority
- CN
- China
- Prior art keywords
- signal
- separation
- iteration
- separated
- residual coefficients
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 50
- 238000000926 separation method Methods 0.000 claims abstract description 263
- 230000008569 process Effects 0.000 claims abstract description 21
- 238000005070 sampling Methods 0.000 claims abstract description 13
- 239000011159 matrix material Substances 0.000 claims description 44
- 238000012545 processing Methods 0.000 claims description 8
- 238000007476 Maximum Likelihood Methods 0.000 claims description 7
- 238000013528 artificial neural network Methods 0.000 claims description 7
- 238000012880 independent component analysis Methods 0.000 claims description 7
- 230000011664 signaling Effects 0.000 claims description 7
- 238000000513 principal component analysis Methods 0.000 claims description 6
- 230000017105 transposition Effects 0.000 claims description 6
- WDEFWOPTRFWZLM-LEQGEALCSA-N (2,5-dioxopyrrolidin-1-yl) (2r)-2,5,7,8-tetramethyl-6-(oxan-2-yloxy)-3,4-dihydrochromene-2-carboxylate Chemical compound C([C@@](C)(OC=1C(C)=C2C)C(=O)ON3C(CCC3=O)=O)CC=1C(C)=C2OC1CCCCO1 WDEFWOPTRFWZLM-LEQGEALCSA-N 0.000 claims 2
- 230000006870 function Effects 0.000 description 26
- 230000000694 effects Effects 0.000 description 6
- 238000010586 diagram Methods 0.000 description 4
- 238000001914 filtration Methods 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000002068 genetic effect Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000013518 transcription Methods 0.000 description 1
- 230000035897 transcription Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/0208—Noise filtering
- G10L21/0216—Noise filtering characterised by the method used for estimating noise
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/0272—Voice signal separating
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
- G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/0208—Noise filtering
- G10L2021/02087—Noise filtering the noise being separate speech, e.g. cocktail party
Landscapes
- Engineering & Computer Science (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Signal Processing (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Human Computer Interaction (AREA)
- Acoustics & Sound (AREA)
- Quality & Reliability (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Cable Transmission Systems, Equalization Of Radio And Reduction Of Echo (AREA)
- Circuit For Audible Band Transducer (AREA)
- Telephone Function (AREA)
Abstract
The invention discloses a kind of speech separating methods, comprising: obtains the voice data to be separated of each signal path；For each preset sampling instant, separating treatment is carried out to voice data to be separated using blind source separation algorithm, obtains P separation signal；It calculates current separation signal and is separated with other that P separate in signal in addition to currently separating signal and intersect residual coefficients between signal；And whether judge to intersect residual coefficients less than the first preset threshold；If not, using echo cancellation algorithm all intersection residual coefficients are not less than with the separation signal of the first preset threshold, echo cancellation process is carried out, and separates signal for treated and separates signal as target with all set for intersecting the separation signal that residual coefficients are not less than the first preset threshold；If so, will separate signal as target separates signal.The embodiment of the invention also provides a kind of speech Separation devices.Using the embodiment of the present invention, the crossbar signal residual in voice signal can be reduced.
Description
Technical field
The present invention relates to a kind of method of speech processing and device, it is more particularly to a kind of speech separating method and device.
Background technique
In speech processes field, it will usually it encounters to more people while the voice signal spoken carries out speech Separation processing, into
And everyone voice signal is obtained, the effect of better Sound seperation how is obtained, that is, other speakers' after separating is residual
Stay voice less.This problem is academicly being known as " cocktail party problem ", is to perplex man machine language's interactive application for a long time
Problem, so far still without stablizing available product or scheme in the actual environment.
Currently, the algorithm of common speech Separation specifically includes that neural network algorithm, maximum entropy algorithm, Minimum mutual information
Algorithm, maximum likelihood algorithm, Independent Component Analysis Algorithm, genetic algorithm, machine learning, the Wave beam forming based on microphone array
Algorithm etc..
But since the basic theory of existing algorithm limits, cause the universal separating effect of existing algorithm undesirable, intersect letter
Number residual is larger.
Summary of the invention
Technical problem to be solved by the present invention lies in a kind of speech separating method and device is provided, to solve existing skill
Crossbar signal remains larger technical problem in art.
The present invention is to solve above-mentioned technical problem by the following technical programs:
The embodiment of the invention provides a kind of speech separating methods, which comprises
Obtain the voice data to be separated of each signal path, wherein the voice data to be separated contains at least two
The voice data that people generates when speaking simultaneously；
For each preset sampling instant, the voice data to be separated is carried out at separation using blind source separation algorithm
Reason, obtains P separation signal；
Signal is separated for each, current separation signal is calculated and is currently separated with a separate in signal of the P except described
The intersection residual coefficients between other separation signals except signal；And judge whether the intersection residual coefficients are pre- less than first
If threshold value；
If it is not, using echo cancellation algorithm all intersection residual coefficients are not less than with the separation signal of the first preset threshold,
Echo cancellation process is carried out, and separation signal with all intersects separation of the residual coefficients less than the first preset threshold by treated
The set of signal separates signal as target；
If so, separating signal for the separation signal as target.
Optionally, the blind source separation algorithm includes: non-linear principal component analysis, independent component analysis, neural network calculation
One of method, maximum entropy algorithm, Minimum mutual information algorithm, maximum likelihood algorithm or multiple combinations.
It is optionally, described that separating treatment is carried out to the voice data to be separated using blind source separation algorithm, comprising:
For each voice data to be separated, the generation for being directed to the voice data to be separated is established using NPCA criterion
Valence functionWherein,
J (W) is the cost of the separation matrix of t moment；E { } is desired operation function；X (t) is corresponding for each microphone
The observation signal that is observed of signal path；W is separation matrix；(.)TFor transposition operation；For nonlinear function；T is to work as
The preceding moment；
Minimum processing is carried out to the cost function, obtains the iterative estimate of separation matrix are as follows:
W (t+1)=W (t)+θ * z (t) [xT(t)-zT(t) W (t)], wherein
W (t+1) is the separation matrix at t+1 moment；W (t) is the separation matrix of t moment；θ is iteration step length, andθ (t) is the iteration step length of t moment, and θ (t-1) is the iteration at t-1 moment
Step-length, ρ are constant,For gradient function, J (t) is the cost of t moment；Z (t) is nonlinear function, and
Utilize formula, W (t+1)=W (t)+θ * z (t) [xT(t)-zT(t) W (t)], iterate to calculate the separation square of subsequent time
Battle array obtains the target separation matrix of each voice data to be separated until the separation matrix is restrained；
Utilize formula, y (t)=Wx (t), the signal after obtaining the separation of the voice data to be separated, wherein y (t) is
Signal after the separation of Current observation signal.
Optionally, the current separation signal of the calculating with the P separate in signal except it is described it is current separate signal in addition to
Other separation signals between intersection residual coefficients, comprising:
Using formula,Current separation signal is calculated to separate in signal with the P
The intersection residual coefficients between other separation signals in addition to the current separation signal, wherein
For current separation signal and the P of i-th of channel separate in signal except it is described currently separate signal in addition to
Other separation signals between intersection residual coefficients；I is the number in the channel of current separation signal；J is P separation letter
The number in the channel of other separation signals in number in addition to the current separation signal；ai,kFor the separation signal in i-th of channel
The mixed coefficint between signal is separated with k-th；aj,kMixing between signal is separated with k-th for the separation signal in j-th of channel
Collaboration number；ykFor the sound-source signal in k-th of channel；∑ is summing function.
Optionally, the separation that using echo cancellation algorithm all intersection residual coefficients are not less than with the first preset threshold
Signal carries out echo cancellation process, comprising:
For all intersection residual coefficients not less than each separation signal in the separation signal of the first preset threshold, will work as
Preceding separation signal makees near end signal；Residual coefficients will be intersected currently to divide not less than in the separation signal of the first preset threshold except described
From other signals except signal as remote signaling；
Using formula,Obtain error signal；Wherein, e (n) is error letter
Number；D (n) is desired output signal；N is the corresponding duration of each audio frame, and value is filter length；K is in audio frame
The serial number of sampled point；K-th of sampled point corresponding filter coefficient when for nth iteration；N is the number of iterations；x(n-k)
Observation signal when iteration secondary for the n-th-k；
Using formula,Update iteration step length, wherein
Iteration step length when μ (n) is nth iteration；For the variance of near end signal；When N is that each audio frame is corresponding
Long, value is filter length, and k ∈ (0, N)；Observation signal when x (n-i) is the n-th-i iteration；Λ (n) is the l times
Imbalance when iteration；
Using formula,Update filtering
The estimated value of the coefficient of device, wherein
The estimated value of filter coefficient when for (n+1)th iteration；μ (n) is iteration step length；For l
The estimated value of filter coefficient when secondary iteration；N is the corresponding duration of each audio frame, and value is filter length；x(n-i)
Observation signal when iteration secondary for the n-th-i；x*(n-k) observation signal conjugate when iteration secondary for the n-th-k；| | it is modulus letter
Number；
Utilize formula, d (n)=v (n)+∑kwk(n) x (n-k) calculates desired signal when nth iteration, wherein v (n)
For near end signal；wk(n) be nth iteration when the corresponding filter coefficient of k-th of sampled point theoretical value；X (n-k) is n-th-
Observation signal when k iteration；
Judge whether desired signal when nth iteration restrains, makees closely if so, returning and executing the signal that will currently separate
The step of end signal；If it is not, desired signal when using the nth iteration is as the signal after eliminating echo.
The embodiment of the invention provides a kind of speech Separation device, described device includes:
First obtains module, for obtaining the voice data to be separated of each signal path, wherein the voice to be separated
The voice data that data contain at least two people while generating when speaking；
Second obtains module, for being directed to each preset sampling instant, using blind source separation algorithm to described to be separated
Voice data carries out separating treatment, obtains P separation signal；
Computing module, for for each separation signal, the current separation signal of calculating separates in signal with the P to be removed
The intersection residual coefficients between other separation signals except the current separation signal；And judge that the intersection residual coefficients are
It is no less than the first preset threshold；
Cancellation module, for utilizing echo cancellation algorithm pair in the case where the calculated result of the computing module is no
All separation signals for intersecting residual coefficients and being not less than the first preset threshold, carry out echo cancellation process, and will treated point
Intersect residual coefficients with all from signal and separate signal as target less than the set of the separation signal of the first preset threshold；
Setup module, for the calculated result of the computing module be in the case where, using the separation signal as
Target separates signal.
Optionally, the blind source separation algorithm includes: non-linear principal component analysis, independent component analysis, neural network calculation
One of method, maximum entropy algorithm, Minimum mutual information algorithm, maximum likelihood algorithm or multiple combinations.
Optionally, described second module is obtained, is also used to:
For each voice data to be separated, the generation for being directed to the voice data to be separated is established using NPCA criterion
Valence functionWherein,
J (W) is the cost of the separation matrix of t moment；E { } is desired operation function；X (t) is corresponding for each microphone
The observation signal that is observed of signal path；W is separation matrix；(.)TFor transposition operation；For nonlinear function；T is to work as
The preceding moment；
Minimum processing is carried out to the cost function, obtains the iterative estimate of separation matrix are as follows:
W (t+1)=W (t)+θ * z (t) [xT(t)-zT(t) W (t)], wherein
W (t+1) is the separation matrix at t+1 moment；W (t) is the separation matrix of t moment；θ is iteration step length, andθ (t) is the iteration step length of t moment, and θ (t-1) is the iteration at t-1 moment
Step-length, ρ are constant,For gradient function, J (t) is the cost of t moment；Z (t) is nonlinear function, and
Utilize formula, W (t+1)=W (t)+θ * z (t) [xT(t)-zT(t) W (t)], iterate to calculate the separation square of subsequent time
Battle array obtains the target separation matrix of each voice data to be separated until the separation matrix is restrained；
Utilize formula, y (t)=Wx (t), the signal after obtaining the separation of the voice data to be separated, wherein y (t) is
Signal after the separation of Current observation signal.
Optionally, the computing module, is also used to:
Using formula,Current separation signal is calculated to separate in signal with the P
The intersection residual coefficients between other separation signals in addition to the current separation signal, wherein
For current separation signal and the P of i-th of channel separate in signal except it is described currently separate signal in addition to
Other separation signals between intersection residual coefficients；I is the number in the channel of current separation signal；J is P separation letter
The number in the channel of other separation signals in number in addition to the current separation signal；ai,kFor the separation signal in i-th of channel
The mixed coefficint between signal is separated with k-th；aj,kMixing between signal is separated with k-th for the separation signal in j-th of channel
Collaboration number；ykFor the sound-source signal in k-th of channel；∑ is summing function.
Optionally, the cancellation module, is also used to:
For all intersection residual coefficients not less than each separation signal in the separation signal of the first preset threshold, will work as
Preceding separation signal makees near end signal；Residual coefficients will be intersected currently to divide not less than in the separation signal of the first preset threshold except described
From other signals except signal as remote signaling；
Using formula,Obtain error signal；Wherein, e (n) is error letter
Number；D (n) is desired output signal；N is the corresponding duration of each audio frame, and value is filter length；K is in audio frame
The serial number of sampled point；K-th of sampled point corresponding filter coefficient when for nth iteration；N is the number of iterations；x(n-k)
Observation signal when iteration secondary for the n-th-k；
Using formula,Update iteration step length, wherein
Iteration step length when μ (n) is nth iteration；For the variance of near end signal；When N is that each audio frame is corresponding
Long, value is filter length, and k ∈ (0, N)；Observation signal when x (n-i) is the n-th-i iteration；Λ (n) is the l times
Imbalance when iteration；
Using formula,Update filtering
The estimated value of the coefficient of device, wherein
The estimated value of filter coefficient when for (n+1)th iteration；μ (n) is iteration step length；For l
The estimated value of filter coefficient when secondary iteration；N is the corresponding duration of each audio frame, and value is filter length；x(n-i)
Observation signal when iteration secondary for the n-th-i；x*(n-k) observation signal conjugate when iteration secondary for the n-th-k；| | it is modulus letter
Number；
Utilize formula, d (n)=v (n)+∑kwk(n) x (n-k) calculates desired signal when nth iteration, wherein v (n)
For near end signal；wk(n) be nth iteration when the corresponding filter coefficient of k-th of sampled point theoretical value；X (n-k) is n-th-
Observation signal when k iteration；
Judge whether desired signal when nth iteration restrains, makees closely if so, returning and executing the signal that will currently separate
The step of end signal；If it is not, desired signal when using the nth iteration is as the signal after eliminating echo.
The present invention has the advantage that compared with prior art
Using the embodiment of the present invention, crossbar signal remaining in the signal after separation can be can be regarded as into other sound sources
Echo, reuse echo cancellation algorithm to each separation signal carry out echo cancellation process, so as to reach improve separation
Effect, and then reduce the crossbar signal residual in echo signal.
Detailed description of the invention
Fig. 1 is a kind of flow diagram of speech separating method provided in an embodiment of the present invention；
Fig. 2 is a kind of structural schematic diagram of speech Separation device provided in an embodiment of the present invention.
Specific embodiment
It elaborates below to the embodiment of the present invention, the present embodiment carries out under the premise of the technical scheme of the present invention
Implement, the detailed implementation method and specific operation process are given, but protection scope of the present invention is not limited to following implementation
Example.
The embodiment of the invention provides a kind of speech separating method and devices, just provided in an embodiment of the present invention first below
A kind of speech separating method is introduced.
Firstly the need of explanation, the embodiment of the present invention has a wide range of applications scene, such as: (1) traditionally for public affairs
The monitoring of occasion only has video monitoring altogether, can not accomplish sound monitoring, because there may be multiple speakers are same in public arena
When speak, in addition can also have various ambient noises, background music etc..It can be in protection and monitor field using the embodiment of the present invention
It is monitored while realizing voice and video.(2) occur the meeting transcription system for completing meeting summary in real time in the industry at present,
To efficiently complete the function of meeting summary, but this system is made a speech (if session discussing mistake simultaneously for there are more people
When debating actively in journey) the case where will fail, existing speech recognition system can not cope with more speaker's voices completely
The scene of identification.Using the embodiment of the present invention, intelligent meeting system can be applied to.(3) it can be applied to general voice drop
Scene of making an uproar remains with the channel of user's normal speech by Sound seperation, removes the channel of not normal speech, can be realized
Voice de-noising.
Fig. 1 is a kind of flow diagram of speech separating method provided in an embodiment of the present invention, as shown in Figure 1, the side
Method includes:
S101: the voice data to be separated of each signal path is obtained, wherein the voice data to be separated contains at least
The voice data that two people generate when speaking simultaneously.
Specifically, the microphone that different location is arranged in using at least two, it is same to obtain two or two people or more
When voice data when speaking, one of microphone obtains voice data to be separated all the way, for example, microphone -1 obtain to
The voice data -4 to be separated of voice data -2 to be separated, the acquisition of microphone -3 that separation voice data -1, microphone -2 obtain,
The voice data -5 to be separated that microphone -1 obtains.
It is understood that voice data to be separated corresponds to a signal path all the way.
S102: being directed to each preset sampling instant, is carried out using blind source separation algorithm to the voice data to be separated
Separating treatment obtains P separation signal.
Specifically, the blind source separation algorithm includes: non-linear principal component analysis, independent component analysis, neural network calculation
One of method, maximum entropy algorithm, Minimum mutual information algorithm, maximum likelihood algorithm or multiple combinations
Specifically, with the voice data position input to be separated of the road S101 step Zhong Ge, benefit is with the following method, available every
The corresponding road P of one preset sampling instant separates signal.It can be directed to each voice data to be separated, utilize NPCA
(nonlinear principle component analysis, non-linear principal component analysis) criterion is established for described wait divide
Cost function from voice data:
Wherein, J (W) is the cost of the separation matrix of t moment；E { } is
Expectation computing function；The observation signal that x (t) is observed by the corresponding signal path of each microphone；W is separation matrix；(.
)TFor transposition operation；For nonlinear function；T is current time；
Minimum processing is carried out to the cost function, obtains the iterative estimate of separation matrix are as follows:
W (t+1)=W (t)+θ * z (t) [xT(t)-zT(t) W (t)], wherein
W (t+1) is the separation matrix at t+1 moment；W (t) is the separation matrix of t moment；θ is iteration step length, andθ (t) is the iteration step length of t moment, and θ (t-1) is the iteration at t-1 moment
Step-length, ρ are constant,For gradient function, J (t) is the cost of t moment；Z (t) is nonlinear function, and
Utilize formula, W (t+1)=W (t)+θ * z (t) [xT(t)-zT(t) W (t)], iterate to calculate the separation square of subsequent time
Battle array obtains the target separation matrix of each voice data to be separated until the separation matrix is restrained；
Utilize formula, y (t)=Wx (t), the signal after obtaining the separation of the voice data to be separated, wherein y (t) is
Signal after the separation of Current observation signal.
For example, the available road the P separation signal of the 1st sampling instant, the available road the P separation letter of the 2nd sampling instant
Number, the available road P of n-th of sampling instant separate signal.
It is emphasized that aforementioned observed signal refers to, per voice data to be separated all the way.
S103: it for each separation signal, calculates current separation signal and separates in signal with the P except described current
Separate the intersection residual coefficients between other separation signals except signal；And judge the intersection residual coefficients whether less than
One preset threshold；If it is not, executing S104 step；If so, executing S105 step.
Specifically, s can be usedi(n) each obtained separation signal is indicated, wherein i is the corresponding letter of separation signal
Number channel position, the i.e. serial number of microphone；N is the serial number of the sampling instant in each signal path.Inventors have found that in reality
In the application of border, any one separation signal can be regarded as other roads P-1 in the same time and intersect what residual signal mixed, because
This, can use formula,Current separation signal is calculated to remove with a separate in signal of the P
The intersection residual coefficients between other separation signals except the current separation signal, wherein
For current separation signal and the P of i-th of channel separate in signal except it is described currently separate signal in addition to
Other separation signals between intersection residual coefficients；I is the number in the channel of current separation signal；J is P separation letter
The number in the channel of other separation signals in number in addition to the current separation signal；ai,kFor the separation signal in i-th of channel
The mixed coefficint between signal is separated with k-th, it is to be understood that when k is 1, ai,kValue be 1；aj,kFor j-th of channel
Separation signal the mixed coefficint between signal is separated with k-th；ykFor the sound-source signal in k-th of channel；∑ is summing function.
It is not less than the first preset threshold when intersecting residual coefficients, when such as 0.0125, executes S104 step；If intersecting residual system
Number executes S105 step less than the first preset threshold.
In practical applications, t can be usedi(n) it indicates to need to carry out the separation signal of echo cancellation process, wherein i is
The corresponding signal path serial number of the separation signal, the i.e. serial number of microphone；N is the sequence of the sampling instant in each signal path
Number, and i ∈ (1,2 ..., Q), Q≤P.
S104: being not less than the separation signal of the first preset threshold using echo cancellation algorithm to all intersection residual coefficients,
Echo cancellation process is carried out, and separation signal with all intersects separation of the residual coefficients less than the first preset threshold by treated
The set of signal separates signal as target.
Specifically, the echo cancellation algorithm, comprising: frequency domain MDF algorithm.In practical applications, it will successively need to carry out
The separation signal t of echo cancellation processi(n) it is successively used as near end signal, others need to carry out the separation of echo cancellation process
Signal is handled using echo cancellation algorithm respectively as remote signaling.
Specifically, the separation that using echo cancellation algorithm all intersection residual coefficients are not less than with the first preset threshold
Signal carries out echo cancellation process, comprising:
For all intersection residual coefficients not less than each separation signal in the separation signal of the first preset threshold, will work as
Preceding separation signal makees near end signal；Residual coefficients will be intersected currently to divide not less than in the separation signal of the first preset threshold except described
From other signals except signal as remote signaling；
Using formula,Obtain error signal；Wherein, e (n) is error letter
Number；D (n) is desired output signal；N is the corresponding duration of each audio frame, and value is filter length；K is in audio frame
The serial number of sampled point；K-th of sampled point corresponding filter coefficient when for nth iteration；N is the number of iterations；x(n-k)
Observation signal when iteration secondary for the n-th-k；
Using formula,Update iteration step length, wherein
Iteration step length when μ (n) is nth iteration；For the variance of near end signal；When N is that each audio frame is corresponding
Long, value is filter length, and k ∈ (0, N)；Observation signal when x (n-i) is the n-th-i iteration；Λ (n) is the l times
Imbalance when iteration；
Using formula,Update filtering
The estimated value of the coefficient of device, wherein
The estimated value of filter coefficient when for (n+1)th iteration；μ (n) is iteration step length；For l
The estimated value of filter coefficient when secondary iteration；N is the corresponding duration of each audio frame, and value is filter length；x(n-i)
Observation signal when iteration secondary for the n-th-i；Observation signal conjugate when x* (n-k) is the n-th-k iteration；| | it is modulus letter
Number；
Utilize formula, d (n)=v (n)+∑kwk(n) x (n-k) calculates desired signal when nth iteration, wherein v (n)
For near end signal；wk(n) be nth iteration when the corresponding filter coefficient of k-th of sampled point theoretical value；X (n-k) is n-th-
Observation signal when k iteration；
Judge whether desired signal when nth iteration restrains, makees closely if so, returning and executing the signal that will currently separate
The step of end signal；If it is not, desired signal when using the nth iteration is as the signal after eliminating echo.
By the separation signal after progress echo cancellation process with all residual coefficients of intersecting not less than the first preset threshold
Signal is separated, that is, the set for not needing the separation signal for carrying out echo cancellation process separates signal as target.
S105: signal is separated using the separation signal as target.
Separation signal of the residual coefficients not less than the first preset threshold will be intersected, that is, do not need to carry out echo cancellation process
The set for separating signal separates signal as target.
It should be noted that the P separation signal for the (n+1)th moment is also handled according to the method described above.Finally
The target separation signal at each moment is arrived.
Using embodiment illustrated in fig. 1 of the present invention, crossbar signal remaining in the signal after separation can be can be regarded as
The echo of other sound sources reuses echo cancellation algorithm and carries out echo cancellation process to each separation signal, so as to reach
Improve separating effect, and then reduces the crossbar signal residual in echo signal.
In addition, if all echo cancellor is used to post-process each separation signal after blind source separating, and can bring
Great extra computation amount can effectively judge which signal after blind source separating is suitble to use using the embodiment of the present invention
Echo cancellation process improves separating effect, and then effectively promotes the working efficiency of whole system.
Corresponding with embodiment illustrated in fig. 1 of the present invention, the embodiment of the invention also provides a kind of speech Separation devices.
Fig. 2 is a kind of structural schematic diagram of speech Separation device provided in an embodiment of the present invention, as shown in Fig. 2, the dress
It sets and includes:
First obtains module 201, for obtaining the voice data to be separated of each signal path, wherein described to be separated
The voice data that voice data contains at least two people while generating when speaking；
Second obtains module 202, for being directed to each preset sampling instant, using blind source separation algorithm to described wait divide
Separating treatment is carried out from voice data, obtains P separation signal；
Computing module 203, for calculating current separation signal with the P and separating signal for each separation signal
In except it is described it is current separation signal in addition to other separation signals between intersection residual coefficients；And judge intersection residual system
Whether number is less than the first preset threshold；
Cancellation module 204, for utilizing echo cancellation algorithm in the case where the calculated result of the computing module is no
All intersection residual coefficients are not less than with the separation signal of the first preset threshold, carries out echo cancellation process, and by treated
Separation signal intersects residual coefficients and separates signal as target less than the set of the separation signal of the first preset threshold with all；
Setup module 205, in the case where the calculated result of the computing module, which is, is, the separation signal to be made
Signal is separated for target.
Using embodiment illustrated in fig. 2 of the present invention, crossbar signal remaining in the signal after separation can be can be regarded as
The echo of other sound sources reuses echo cancellation algorithm and carries out echo cancellation process to each separation signal, so as to reach
Improve separating effect, and then reduces the crossbar signal residual in echo signal.
In a kind of specific embodiment of the embodiment of the present invention, the blind source separation algorithm includes: non-linear principal component
One of analysis, independent component analysis, neural network algorithm, maximum entropy algorithm, Minimum mutual information algorithm, maximum likelihood algorithm
Or multiple combination.
In a kind of specific embodiment of the embodiment of the present invention, described second obtains module 202, is also used to:
For each voice data to be separated, the generation for being directed to the voice data to be separated is established using NPCA criterion
Valence functionWherein,
J (w) is the cost of the separation matrix of t moment；E { } is desired operation function；X (t) is corresponding for each microphone
The observation signal that is observed of signal path；W is separation matrix；(.)TFor transposition operation；For nonlinear function；T is to work as
The preceding moment；
Minimum processing is carried out to the cost function, obtains the iterative estimate of separation matrix are as follows:
W (t+1)=W (t)+θ * z (t) [xT(t)-zT(t) W (t)], wherein
W (t+1) is the separation matrix at t+1 moment；W (t) is the separation matrix of t moment；θ is iteration step length, andθ (t) is the iteration step length of t moment, and θ (t-1) is the iteration step length at t-1 moment,
ρ is constant,For gradient function, J (t) is the cost of t moment；Z (t) is nonlinear function, and
Utilize formula, W (t+1)=W (t)+θ * z (t) [xT(t)-zT(t) W (t)], iterate to calculate the separation square of subsequent time
Battle array obtains the target separation matrix of each voice data to be separated until the separation matrix is restrained；
Utilize formula, y (t)=Wx (t), the signal after obtaining the separation of the voice data to be separated, wherein y (t) is
Signal after the separation of Current observation signal.
In a kind of specific embodiment of the embodiment of the present invention, the computing module 203 is also used to:
Using formula,Current separation signal is calculated to separate in signal with the P
The intersection residual coefficients between other separation signals in addition to the current separation signal, wherein
For current separation signal and the P of i-th of channel separate in signal except it is described currently separate signal in addition to
Other separation signals between intersection residual coefficients；I is the number in the channel of current separation signal；J is P separation letter
The number in the channel of other separation signals in number in addition to the current separation signal；ai,kFor the separation signal in i-th of channel
The mixed coefficint between signal is separated with k-th；aj,kMixing between signal is separated with k-th for the separation signal in j-th of channel
Collaboration number；ykFor the sound-source signal in k-th of channel；∑ is summing function.
In a kind of specific embodiment of the embodiment of the present invention, the cancellation module 204 is also used to:
For all intersection residual coefficients not less than each separation signal in the separation signal of the first preset threshold, will work as
Preceding separation signal makees near end signal；Residual coefficients will be intersected currently to divide not less than in the separation signal of the first preset threshold except described
From other signals except signal as remote signaling；
Using formula,Obtain error signal；Wherein, e (n) is error letter
Number；D (n) is desired output signal；N is the corresponding duration of each audio frame, and value is filter length；K is in audio frame
The serial number of sampled point；K-th of sampled point corresponding filter coefficient when for nth iteration；N is the number of iterations；x(n-k)
Observation signal when iteration secondary for the n-th-k；
Using formula,Update iteration step length, wherein
Iteration step length when μ (n) is nth iteration；For the variance of near end signal；When N is that each audio frame is corresponding
Long, value is filter length, and k ∈ (0, N)；Observation signal when x (n-i) is the n-th-i iteration；Λ (n) is the l times
Imbalance when iteration；
Using formula,Update filtering
The estimated value of the coefficient of device, wherein
The estimated value of filter coefficient when for (n+1)th iteration；μ (n) is iteration step length；For l
The estimated value of filter coefficient when secondary iteration；N is the corresponding duration of each audio frame, and value is filter length；x(n-i)
Observation signal when iteration secondary for the n-th-i；x*(n-k) observation signal conjugate when iteration secondary for the n-th-k；| | it is modulus letter
Number；
Utilize formula, d (n)=v (n)+∑kwk(n) x (n-k) calculates desired signal when nth iteration, wherein v (n)
For near end signal；wk(n) be nth iteration when the corresponding filter coefficient of k-th of sampled point theoretical value；X (n-k) is n-th-
Observation signal when k iteration；
Judge whether desired signal when nth iteration restrains, makees closely if so, returning and executing the signal that will currently separate
The step of end signal；If it is not, desired signal when using the nth iteration is as the signal after eliminating echo.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (10)
1. a kind of speech separating method, which is characterized in that the described method includes:
Obtain the voice data to be separated of each signal path, wherein it is same that the voice data to be separated contains at least two people
When the voice data that generates when speaking；
For each preset sampling instant, separating treatment is carried out to the voice data to be separated using blind source separation algorithm,
Obtain P separation signal；
Signal is separated for each, current separation signal is calculated with a separate in signal of the P and currently separates signal except described
Except other separation signals between intersection residual coefficients；And judge the intersection residual coefficients whether less than the first default threshold
Value；
If it is not, using echo cancellation algorithm all intersection residual coefficients are not less than with the separation signal of the first preset threshold, carry out
Echo cancellation process, and separation signal with all intersects separation signal of the residual coefficients less than the first preset threshold by treated
Set as target separate signal；
If so, separating signal for the separation signal as target.
2. a kind of speech separating method according to claim 1, which is characterized in that the blind source separation algorithm includes: non-
Linear principal component analysis, independent component analysis, neural network algorithm, maximum entropy algorithm, Minimum mutual information algorithm, maximum likelihood are calculated
One of method or multiple combinations.
3. a kind of speech separating method according to claim 1, which is characterized in that described to utilize blind source separation algorithm to institute
It states voice data to be separated and carries out separating treatment, comprising:
For each voice data to be separated, the cost letter for being directed to the voice data to be separated is established using NPCA criterion
NumberWherein,
J (W) is the cost of the separation matrix of t moment；E { } is desired operation function；X (t) is the corresponding letter of each microphone
The observation signal that number channel is observed；W is separation matrix；(.)TFor transposition operation；For nonlinear function；When t is current
It carves；
Minimum processing is carried out to the cost function, obtains the iterative estimate of separation matrix are as follows:
W (t+1)=W (t)+θ * z (t) [xT(t)-zT(t) W (t)], wherein
W (t+1) is the separation matrix at t+1 moment；W (t) is the separation matrix of t moment；θ is iteration step length, and θ (t)=θ
(t-1)-ρ*▽θ*(J(t)|θ=θ (t-1)), θ (t) is the iteration step length of t moment, and θ (t-1) is the iteration step length at t-1 moment, ρ
For constant, ▽θFor gradient function, J (t) is the cost of t moment；Z (t) is nonlinear function, and
Utilize formula, W (t+1)=W (t)+θ * z (t) [xT(t)-zT(t) W (t)], the separation matrix of subsequent time is iterated to calculate,
Until the separation matrix is restrained, the target separation matrix of each voice data to be separated is obtained；
Utilize formula, y (t)=Wx (t), the signal after obtaining the separation of the voice data to be separated, wherein y (t) is current
Signal after the separation of observation signal.
4. a kind of speech separating method according to claim 1, which is characterized in that the calculating is current to separate signal and institute
State the intersection residual coefficients between other separation signals in P separation signal in addition to the current separation signal, comprising:
Using formula,Current separation signal is calculated to separate in signal with the P except described
The intersection residual coefficients between other separation signals except current separation signal, wherein
For current separation signal and the P of i-th of channel separate in signal except it is described currently separate signal in addition to its
He separates the intersection residual coefficients between signal；I is the number in the channel of current separation signal；J is in the P separation signal
The number in the channel of other separation signals in addition to the current separation signal；ai,kFor the separation signal and in i-th of channel
Mixed coefficint between k separation signal；aj,kThe mixed stocker between signal is separated with k-th for the separation signal in j-th of channel
Number；ykFor the sound-source signal in k-th of channel；∑ is summing function.
5. a kind of speech separating method according to claim 1, which is characterized in that described to utilize echo cancellation algorithm to institute
Have and intersect the separation signal that residual coefficients are not less than the first preset threshold, carries out echo cancellation process, comprising:
For all intersection residual coefficients not less than each separation signal in the separation signal of the first preset threshold, will currently divide
Make near end signal from signal；Residual coefficients will be intersected and believed not less than the current separation is removed in the separation signal of the first preset threshold
Other signals except number are as remote signaling；
Using formula,Obtain error signal；Wherein, e (n) is error signal；d
It (n) is desired output signal；N is the corresponding duration of each audio frame, and value is filter length；K is to sample in audio frame
The serial number of point；K-th of sampled point corresponding filter coefficient when for nth iteration；N is the number of iterations；X (n-k) is the
Observation signal when n-k iteration；
Using formula,Update iteration step length, wherein
Iteration step length when μ (n) is nth iteration；For the variance of near end signal；N is the corresponding duration of each audio frame,
Value is filter length, and k ∈ (0, N)；Observation signal when x (n-i) is the n-th-i iteration；Λ (n) is the l times iteration
When imbalance；
Using formula,Update filter
The estimated value of coefficient, wherein
The estimated value of filter coefficient when for (n+1)th iteration；μ (n) is iteration step length；Repeatedly for the l times
For when filter coefficient estimated value；N is the corresponding duration of each audio frame, and value is filter length；X (n-i) is the
Observation signal when n-i iteration；x*(n-k) observation signal conjugate when iteration secondary for the n-th-k；| | it is mod function；
Utilize formula, d (n)=v (n)+∑kwk(n) x (n-k) calculates desired signal when nth iteration, wherein v (n) is close
End signal；wk(n) be nth iteration when the corresponding filter coefficient of k-th of sampled point theoretical value；X (n-k) is the n-th-k times
Observation signal when iteration；
Judge whether desired signal when nth iteration restrains, execution is described to make proximal end letter for current separation signal if so, returning
Number the step of；If it is not, desired signal when using the nth iteration is as the signal after eliminating echo.
6. a kind of speech Separation device, which is characterized in that described device includes:
First obtains module, for obtaining the voice data to be separated of each signal path, wherein the voice data to be separated
The voice data generated when speaking simultaneously containing at least two people；
Second obtains module, for being directed to each preset sampling instant, using blind source separation algorithm to the voice to be separated
Data carry out separating treatment, obtain P separation signal；
Computing module, for calculating current separation signal and separating in signal with the P except described for each separation signal
The intersection residual coefficients between other separation signals except current separation signal；And judge whether the intersection residual coefficients are small
In the first preset threshold；
Cancellation module, in the case where the calculated result of the computing module is no, using echo cancellation algorithm to all
Intersection residual coefficients are not less than the separation signal of the first preset threshold, carry out echo cancellation process, and separation is believed by treated
Number intersects residual coefficients with all and separate signal as target less than the set of the separation signal of the first preset threshold；
Setup module, for the calculated result of the computing module be in the case where, using the separation signal as target
Separate signal.
7. a kind of speech Separation device according to claim 6, which is characterized in that the blind source separation algorithm includes: non-
Linear principal component analysis, independent component analysis, neural network algorithm, maximum entropy algorithm, Minimum mutual information algorithm, maximum likelihood are calculated
One of method or multiple combinations.
8. a kind of speech Separation device according to claim 6, which is characterized in that described second obtains module, is also used to:
For each voice data to be separated, the cost letter for being directed to the voice data to be separated is established using NPCA criterion
NumberWherein,
J (W) is the cost of the separation matrix of t moment；E { } is desired operation function；X (t) is the corresponding letter of each microphone
The observation signal that number channel is observed；W is separation matrix；(.)TFor transposition operation；For nonlinear function；When t is current
It carves；
Minimum processing is carried out to the cost function, obtains the iterative estimate of separation matrix are as follows:
W (t+1)=W (t)+θ * z (t) [xT(t)-zT(t) W (t)], wherein
W (t+1) is the separation matrix at t+1 moment；W (t) is the separation matrix of t moment；θ is iteration step length, and θ (t)=θ
(t-1)-ρ*▽θ*(J(t)|θ=θ (t-1)), θ (t) is the iteration step length of t moment, and θ (t-1) is the iteration step length at t-1 moment, ρ
For constant, ▽θFor gradient function, J (t) is the cost of t moment；Z (t) is nonlinear function, and
Utilize formula, W (t+1)=W (t)+θ * z (t) [xT(t)-zT(t) W (t)], the separation matrix of subsequent time is iterated to calculate,
Until the separation matrix is restrained, the target separation matrix of each voice data to be separated is obtained；
Utilize formula, y (t)=Wx (t), the signal after obtaining the separation of the voice data to be separated, wherein y (t) is current
Signal after the separation of observation signal.
9. a kind of speech Separation device according to claim 6, which is characterized in that the computing module is also used to:
Using formula,Current separation signal is calculated to separate in signal with the P except described
The intersection residual coefficients between other separation signals except current separation signal, wherein
For current separation signal and the P of i-th of channel separate in signal except it is described currently separate signal in addition to its
He separates the intersection residual coefficients between signal；I is the number in the channel of current separation signal；J is in the P separation signal
The number in the channel of other separation signals in addition to the current separation signal；ai,kFor the separation signal and in i-th of channel
Mixed coefficint between k separation signal；aj,kThe mixed stocker between signal is separated with k-th for the separation signal in j-th of channel
Number；ykFor the sound-source signal in k-th of channel；∑ is summing function.
10. a kind of speech Separation device according to claim 6, which is characterized in that the cancellation module is also used to:
For all intersection residual coefficients not less than each separation signal in the separation signal of the first preset threshold, will currently divide
Make near end signal from signal；Residual coefficients will be intersected and believed not less than the current separation is removed in the separation signal of the first preset threshold
Other signals except number are as remote signaling；
Using formula,Obtain error signal；Wherein, e (n) is error signal；d
It (n) is desired output signal；N is the corresponding duration of each audio frame, and value is filter length；K is to sample in audio frame
The serial number of point；K-th of sampled point corresponding filter coefficient when for nth iteration；N is the number of iterations；X (n-k) is the
Observation signal when n-k iteration；
Using formula,Update iteration step length, wherein
Iteration step length when μ (n) is nth iteration；For the variance of near end signal；N is the corresponding duration of each audio frame,
Value is filter length, and k ∈ (0, N)；Observation signal when x (n-i) is the n-th-i iteration；Λ (n) is the l times iteration
When imbalance；
Using formula,Update filter
The estimated value of coefficient, wherein
The estimated value of filter coefficient when for (n+1)th iteration；μ (n) is iteration step length；Repeatedly for the l times
For when filter coefficient estimated value；N is the corresponding duration of each audio frame, and value is filter length；X (n-i) is the
Observation signal when n-i iteration；x*(n-k) observation signal conjugate when iteration secondary for the n-th-k；| | it is mod function；
Utilize formula, d (n)=v (n)+∑kwk(n) x (n-k) calculates desired signal when nth iteration, wherein v (n) is close
End signal；wk(n) be nth iteration when the corresponding filter coefficient of k-th of sampled point theoretical value；X (n-k) is the n-th-k times
Observation signal when iteration；
Judge whether desired signal when nth iteration restrains, execution is described to make proximal end letter for current separation signal if so, returning
Number the step of；If it is not, desired signal when using the nth iteration is as the signal after eliminating echo.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810820474.9A CN108962276B (en) | 2018-07-24 | 2018-07-24 | Voice separation method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810820474.9A CN108962276B (en) | 2018-07-24 | 2018-07-24 | Voice separation method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108962276A true CN108962276A (en) | 2018-12-07 |
| CN108962276B CN108962276B (en) | 2020-11-17 |
Family
ID=64464704
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810820474.9A Active CN108962276B (en) | 2018-07-24 | 2018-07-24 | Voice separation method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108962276B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020223952A1 (en) * | 2019-05-09 | 2020-11-12 | 广东省智能制造研究所 | Semi-non-negative matrix factorization-based sound signal separation method |
| CN113362847A (en) * | 2021-05-26 | 2021-09-07 | 北京小米移动软件有限公司 | Audio signal processing method and device and storage medium |
| CN113470689A (en) * | 2021-08-23 | 2021-10-01 | 杭州国芯科技股份有限公司 | Voice separation method |
Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101894558A (en) * | 2010-08-04 | 2010-11-24 | 华为技术有限公司 | Lost frame recovering method and equipment as well as speech enhancing method, equipment and system |
| CN101917527A (en) * | 2010-09-02 | 2010-12-15 | 杭州华三通信技术有限公司 | Method and device of echo elimination |
| CN101964670A (en) * | 2009-07-21 | 2011-02-02 | 雅马哈株式会社 | Echo suppression method and apparatus thereof |
| CN102142259A (en) * | 2010-01-28 | 2011-08-03 | 三星电子株式会社 | Signal separation system and method for automatically selecting threshold to separate sound source |
| CN103188184A (en) * | 2012-12-17 | 2013-07-03 | 中国人民解放军理工大学 | NPCA (Nonlinear Principal Component Analysis)-based self-adaptive variable step size blind source separation method |
| US20140105410A1 (en) * | 2012-10-12 | 2014-04-17 | Huawei Technologies Co., Ltd. | Echo cancellation method and device |
| CN103780522A (en) * | 2014-01-08 | 2014-05-07 | 西安电子科技大学 | Non-orthogonal joint diagonalization instantaneous blind source separation method based on double iteration |
| CN105845148A (en) * | 2016-03-16 | 2016-08-10 | 重庆邮电大学 | Convolution blind source separation method based on frequency point correction |
| CN106057210A (en) * | 2016-07-01 | 2016-10-26 | 山东大学 | Quick speech blind source separation method based on frequency point selection under binaural distance |
| CN106898361A (en) * | 2017-03-16 | 2017-06-27 | 杭州电子科技大学 | Single channel blind source separation method based on feedback variation Mode Decomposition |
| CN107316650A (en) * | 2016-04-26 | 2017-11-03 | 诺基亚技术有限公司 | Method, device and the computer program of the modification of the feature associated on the audio signal with separating |
| US20180182412A1 (en) * | 2016-12-28 | 2018-06-28 | Google Inc. | Blind source separation using similarity measure |
| CN108231087A (en) * | 2017-12-14 | 2018-06-29 | 宁波升维信息技术有限公司 | A kind of single channel blind source separating method |
-
2018
- 2018-07-24 CN CN201810820474.9A patent/CN108962276B/en active Active
Patent Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101964670A (en) * | 2009-07-21 | 2011-02-02 | 雅马哈株式会社 | Echo suppression method and apparatus thereof |
| CN102142259A (en) * | 2010-01-28 | 2011-08-03 | 三星电子株式会社 | Signal separation system and method for automatically selecting threshold to separate sound source |
| CN101894558A (en) * | 2010-08-04 | 2010-11-24 | 华为技术有限公司 | Lost frame recovering method and equipment as well as speech enhancing method, equipment and system |
| CN101917527A (en) * | 2010-09-02 | 2010-12-15 | 杭州华三通信技术有限公司 | Method and device of echo elimination |
| US20140105410A1 (en) * | 2012-10-12 | 2014-04-17 | Huawei Technologies Co., Ltd. | Echo cancellation method and device |
| CN103188184A (en) * | 2012-12-17 | 2013-07-03 | 中国人民解放军理工大学 | NPCA (Nonlinear Principal Component Analysis)-based self-adaptive variable step size blind source separation method |
| CN103780522A (en) * | 2014-01-08 | 2014-05-07 | 西安电子科技大学 | Non-orthogonal joint diagonalization instantaneous blind source separation method based on double iteration |
| CN105845148A (en) * | 2016-03-16 | 2016-08-10 | 重庆邮电大学 | Convolution blind source separation method based on frequency point correction |
| CN107316650A (en) * | 2016-04-26 | 2017-11-03 | 诺基亚技术有限公司 | Method, device and the computer program of the modification of the feature associated on the audio signal with separating |
| CN106057210A (en) * | 2016-07-01 | 2016-10-26 | 山东大学 | Quick speech blind source separation method based on frequency point selection under binaural distance |
| US20180182412A1 (en) * | 2016-12-28 | 2018-06-28 | Google Inc. | Blind source separation using similarity measure |
| CN106898361A (en) * | 2017-03-16 | 2017-06-27 | 杭州电子科技大学 | Single channel blind source separation method based on feedback variation Mode Decomposition |
| CN108231087A (en) * | 2017-12-14 | 2018-06-29 | 宁波升维信息技术有限公司 | A kind of single channel blind source separating method |
Non-Patent Citations (2)
| Title |
|---|
| LI HONG ET AL.: "An adaptive echo cancellation method based on a blind signal separation", 《2010 INTERNATIONAL CONFERENCE ON ELECTRICAL AND CONTROL ENGINEERING》 * |
| MUHAMMAD Z. IKRAM ET AL.: "BLIND SOURCE SEPARATION AND ACOUSTIC ECHO CANCELLATION: A UNIFIED FRAMEWORK", 《2012 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS,SPEECH AND SIGNAL PROCESSING(ICASSP)》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020223952A1 (en) * | 2019-05-09 | 2020-11-12 | 广东省智能制造研究所 | Semi-non-negative matrix factorization-based sound signal separation method |
| CN113362847A (en) * | 2021-05-26 | 2021-09-07 | 北京小米移动软件有限公司 | Audio signal processing method and device and storage medium |
| CN113470689A (en) * | 2021-08-23 | 2021-10-01 | 杭州国芯科技股份有限公司 | Voice separation method |
| CN113470689B (en) * | 2021-08-23 | 2024-01-30 | 杭州国芯科技股份有限公司 | Voice separation method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108962276B (en) | 2020-11-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Buchner et al. | TRINICON: A versatile framework for multichannel blind signal processing | |
| Ortega-García et al. | Overview of speech enhancement techniques for automatic speaker recognition | |
| CN108962276A (en) | A kind of speech separating method and device | |
| CN1914683A (en) | Methods and apparatus for blind separation of multichannel convolutive mixtures in the frequency domain | |
| WO2020256257A3 (en) | Combined learning method and device using transformed loss function and feature enhancement based on deep neural network for speaker recognition that is robust to noisy environment | |
| CN1261759A (en) | Adding blind source separate technology to hearing aid | |
| US20090043588A1 (en) | Sound-source separation system | |
| Braun et al. | Task splitting for dnn-based acoustic echo and noise removal | |
| Wang et al. | NN3A: Neural network supported acoustic echo cancellation, noise suppression and automatic gain control for real-time communications | |
| CN106782592B (en) | System and method for eliminating echo and howling of network sound transmission | |
| KR100446626B1 (en) | Noise suppression method and apparatus | |
| CN110992966B (en) | Human voice separation method and system | |
| Takeda et al. | Exploiting known sound source signals to improve ICA-based robot audition in speech separation and recognition | |
| CN108564961A (en) | A kind of voice de-noising method of mobile communication equipment | |
| KR101936242B1 (en) | Apparatus and method for noise removal, and recording medium thereof | |
| Krueger et al. | Bayesian Feature Enhancement for ASR of Noisy Reverberant Real-World Data. | |
| Rincón-Trujillo et al. | Analysis of Speech Separation Methods based on Deep Learning. | |
| Lakra et al. | Selective noise filtering of speech signals using an adaptive neuro-fuzzy inference system as a frequency pre-classifier | |
| Kokkinakis et al. | Multichannel speech separation using adaptive parameterization of source PDFs | |
| CN116648747A (en) | Apparatus for providing a processed audio signal, method for providing a processed audio signal, apparatus for providing a neural network parameter and method for providing a neural network parameter | |
| Adasme et al. | Proposed Integration Algorithm to Optimize the Separation of Audio Signals Using the ICA and Wavelet Transform | |
| Jadhav et al. | Blind source separation: trends of new age-a review | |
| Nezamdoust et al. | Frequency-Domain Functional Links For Nonlinear Feedback Cancellation In Hearing Aids | |
| Aoulass et al. | Noise Reduction using DUET algorithm for dual-microphone mobile station | |
| Harding et al. | Mask estimation based on sound localisation for missing data 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 | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20201027 Address after: 310000 station 07, room 704, building 8, No. 20, Keji Garden Road, Baiyang street, Qiantang New District, Hangzhou City, Zhejiang Province Applicant after: Hangzhou audiometry Technology Co.,Ltd. Address before: 100176 FC-3, 6th floor, No. 5 Building, 2 Ronghua South Road, Daxing Economic and Technological Development Zone, Beijing Applicant before: BEIJING SINWT SCIENCE & TECHNOLOGY Co.,Ltd. |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CP02 | Change in the address of a patent holder | ||
| CP02 | Change in the address of a patent holder |
Address after: Room 703-1, building 8, No.20, kekeyuan Road, Qiantang New District, Hangzhou City, Zhejiang Province Patentee after: Hangzhou audiometry Technology Co.,Ltd. Address before: 310000 station 07, room 704, building 8, No. 20, Keji Garden Road, Baiyang street, Qiantang New District, Hangzhou City, Zhejiang Province Patentee before: Hangzhou audiometry Technology Co.,Ltd. |
|
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20231213 Address after: 100176 1201-09, 12 / F, building 2, yard 1, No. 29, Kechuang 13th Street, Beijing Economic and Technological Development Zone, Daxing District, Beijing Patentee after: BEIJING SINWT SCIENCE & TECHNOLOGY Co.,Ltd. Address before: Room 703-1, building 8, No.20, kekeyuan Road, Qiantang New District, Hangzhou City, Zhejiang Province Patentee before: Hangzhou audiometry Technology Co.,Ltd. |