CN106060295B - A kind of proportional affine projection echo cancel method of convex combination coefficient difference - Google Patents
A kind of proportional affine projection echo cancel method of convex combination coefficient difference Download PDFInfo
- Publication number
- CN106060295B CN106060295B CN201610326009.0A CN201610326009A CN106060295B CN 106060295 B CN106060295 B CN 106060295B CN 201610326009 A CN201610326009 A CN 201610326009A CN 106060295 B CN106060295 B CN 106060295B
- Authority
- CN
- China
- Prior art keywords
- length
- current time
- filter
- vector
- convex combination
- 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.)
- Active
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04M—TELEPHONIC COMMUNICATION
- H04M9/00—Arrangements for interconnection not involving centralised switching
- H04M9/08—Two-way loud-speaking telephone systems with means for conditioning the signal, e.g. for suppressing echoes for one or both directions of traffic
- H04M9/082—Two-way loud-speaking telephone systems with means for conditioning the signal, e.g. for suppressing echoes for one or both directions of traffic using echo cancellers
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Processing of the speech or voice signal 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
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Processing of the speech or voice signal 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
- G10L2021/02082—Noise filtering the noise being echo, reverberation of the speech
Abstract
A kind of proportional affine projection echo cancel method of convex combination coefficient difference the steps include: A, the filtering of distal end sampled signal, and A1, distal end discrete input signal constitute the input vector X (n) of convex combination adaptive-filtering echo cancellation filter；A2, filter input vector X (n) is obtained to big step-length, small step-length filter value y1(n) and y2(n)；B, near end signal d (n) and junction filter output valve y (n) are subtracted each other to obtain total residual signals e (n) after echo cancellor, and send back to distal end by echo cancelltion；C, convex combination, by large and small step-length output valve y1(n) and y2(n) convex combination is carried out by weight λ (n) and obtains combination output valve y (n) and its tap weights coefficient W1(n) and W2(n) convex combination is carried out by weight λ (n) and obtains tap filter tap weight coefficient W (n)；D, filter tap weight vector updates；E, filter weight updates；F, filter weight limits G, repeats the step of A, B, C, D, E, F, and echo cancellor can be realized.The identification capability of this method is strong, and fast convergence rate, tracking ability are strong and steady-state error is low, and echo cancellor effect is good.
Description
Technical field
The invention belongs to communicate adaptive echo technology for eliminating field, proportional affine projection echo cancellor skill is especially belonged to
Art field.
Background technique
With the development of communication technology, more and more communication equipments provide hand-free function, this function makes communication more
It is convenient to add, while also producing echo problem, and echo can seriously affect the stability of voice call quality and system.For present
For communication system (including telephone, hands-free phone, mobile phone and TeleConference Bridge etc.), echo cancellor is that these are
One of the key factor considered emphatically is needed when system design.Echo cancellor is by the remote signaling input adaptive filtering with echo
Device realizes and carries out System Discrimination to unknown echo channel, i.e., by the adjustment of adaptive filter algorithm, analog echo path,
Approach its shock response mutually with actual echo path, to obtain echo prediction signal, then proximally by echo prediction signal
It is subtracted in the voice signal that microphone receives, echo cancellor can be realized.
The most commonly used is least mean square algorithm (LMS) in adaptive filter algorithm, since it does not need to calculate related phase
Function is closed, matrix inversion is not needed yet, so that it becomes the reference standard of other linear adaption filtering algorithms.Since echo is believed
Road is largely Sparse System, i.e., there are many in system (channel) close to zero or null element, only a small number of elements are
Amplitude is greater than zero vector.Therefore, this Sparse System is recognized using least mean square algorithm, convergence rate is very slow.
The preferable method of echo cancellor effect is affine projection algorithm (APA) at present, and affine projection algorithm (APA) repeats benefit
There is better constringency performance in the case where the temporal correlation of input signal is strong with the signal of past period.
" the Double-talk robust fast converging algorithms for network of document 1
echocancellation”(T.Gansler,S.L.Gay,M.M.Sondhi,and J.Benesty,
IEEE.Trans.Speech, Audio process, vol.8, no.6, pp.656-663, Nov.2000.) it is proportional affine
Echo cancel method is projected, then on the basis of affine projection algorithm, the signal (vector) of past different periods is assigned different
Affine projection weight, input signal is big and predicts that the weight of echo value big period is big.Make again the step-length of filter with it is previous
The weight of step-length period is proportional.Namely when voice signal is strong, the step-length of filter is big, accelerates filtering speed, thus
Accelerate algorithm global convergence speed.But the algorithm is lost more data when increasing step-length, leads to its steady-state error
Greatly.
Summary of the invention
It is an object of the invention to provide a kind of adaptive-filtering echoes of the proportional affine projection of convex combination coefficient difference
Removing method.This method accommodative ability of environment is strong, fast convergence rate and low steady-state error.
The technical scheme adopted by the invention for realizing the object of the invention is a kind of proportional radiation throwing of convex combination coefficient difference
Shadow echo cancel method, its step are as follows:
A, remote signaling sampling filter
A1, the remote end input signal for obtaining the remote signaling distally transmitted in current time n and preceding L-1 instance sample
Discrete value x (n), x (n-1) ..., x (n-L+1) constitute the input vector X (n), i.e. X of the current time n of convex combination filter
(n)=[x (n), x (n-1) ..., x (n-L+1)]T；Wherein, L=64 is filter tap number, and subscript T represents transposition operation；
A2, by the input vector X (n) of current time n by respectively obtaining the big step of current time n after convex combination filter
Long filtering output value y1(n), y1(n)=W1 T(n) the small step-length filtering output value y of X (n) and current time n2(n): y2(n)=W2 T
(n)X(n)；
Wherein W1(n)=[w1,1(n),w1,2(n),...,w1,l(n),...,w1,L(n)]TIt is current in convex combination filter
The tap weights vector of the big step-length filter of moment n, initial value are null vector, w1,l(n) it is filtered for the big step-length of current time n
First of tap weight coefficient in the tap weights vector of wave device； W2(n)=[w2,1(n),w2,2(n),...,w2,l(n),...,w2,L
It (n)] is the tap weights vector of the small step-length filter of current time n in convex combination filter, initial value is null vector, w2,l
It (n) is first of tap weight coefficient in the tap weights vector of the small step-length filter of current time n；
B, convex combination
By the big step-length output valve y of current time n1(n) and the small step-length output valve y of current time n2(n) convex combination is carried out
Obtain the output valve y (n) of convex combination filter, y (n)=λ (n) y1(n)+(1-λ(n))y2(n)；By the big step-length of current time n
The tap weights vector W of filter1(n) and the small step-length filter tap weight vector W of current time n2(n) convex combination is carried out to obtain
Filter tap weight vector W (n): W (n)=λ (n) W1(n)+(1-λ(n))W2(n)；Wherein λ (n) is the big step-length of current time n
The weight coefficient of filter, its calculation formula isA (n) is the hybrid parameter of current time n, initial value
It is 0；
C, echo cancelltion
The convex combination of the near end signal d (n) and current time n of the current time n with echo that proximal end is picked up filter
Device output valve y (n) subtracts each other, total residual signals e (n) after the echo that is eliminated, e (n)=d (n)-y (n)；And total residual error is believed
Number send back to distal end.
Meanwhile by the big step-length output valve y of the near end signal d (n) of current time n and current time n1(n) subtract each other, obtain
The big step-length residual signals e of current time n1(n), e1(n)=d (n)-y1(n)；By the near end signal d (n) of current time n with work as
The small step-length output valve y of preceding moment n2(n) subtract each other, obtain the small step-length residual signals e of current time n2(n), e2(n)=d (n)-
y2(n)；
D, filter tap weight vector updates
The composition of D1, affine projection input vector
By convex combination filter in the current time n and input vector X (n) at preceding P-1 moment, X (n-1) ..., X (n-P
+ 1) affine projection input vector U (n), U (n)=[X (n), X (n-1) ... the X (n-P+1)] of current time n, are constituted, wherein P
Represent affine projection order, P=4,8,16；
The composition of D2, residual signals vector
By current time n and total residual signals e (n) at preceding P-1 moment, e (n-1) ..., e (n-P+1) are constituted current
Total residual signals vector E (n) of moment n, E (n)=[e (n), e (n-1) ..., e (n-P+1)]T；
By the current time n and big step-length residual signals e at preceding P-1 moment1(n),e1(n-1),...,e1(n-P+1), structure
At the big step-length residual signals vector E of current time n1(n), E1(n)=[e1(n),e1(n-1),...,e1(n-P+1)]T；
By the current time n and small step-length residual signals e at preceding P-1 moment2(n),e2(n-1),...,e2(n-P+1), structure
At the small step-length residual signals vector E of current time n2(n), E2(n)=[e2(n),e2(n-1),...,e2(n-P+1)]T；
The calculating of D3, the proportional matrix of big step-length
Calculate the proportional coefficient g of first big step-length of current time n1,l(n):
g1,l(n)=max { ρ1,|w1,l(n)-w1,l(kM)|}
Wherein max indicates maximum operation, | | indicate the operation that takes absolute value, ρ1For the tap weights system of big step-length filter
Several threshold values, value range be 0.001~0.01, M be the length of time window, its value range is that 100~200, k is the time
The serial number of window, Indicate downward round numbers operation；
The proportional coefficient g of the big step-length of l-th is arrived by first of current time n1,1(n),g1,2(n)...g1,l(n)...,
g1,L(n), the proportional matrix G of big step-length is constituted1(n), G1(n)=diag { g1,1(n),g1,2(n)...g1,l(n)...g1,L(n) },
Wherein, diag indicates construction diagonal matrix；
D4, small step grow up to the calculating of scaling matrices
First of small step for calculating current time n grows up to proportionality coefficient g2,l(n):
g2,l(n)=max { ρ2,|w2,l(n)-w2,l(kM)|}
Wherein, ρ2For the threshold value of the tap weight coefficient of small step-length filter, value range is 0.001~0.01；
Grow up to proportionality coefficient g to l-th small step for first of current time n2,1(n),g2,2(n) ... g2,l(n),...,
g2,L(n), it constitutes small step and grows up to scaling matrices G2(n):
G2(n)=diag { g2,1(n),g2,2(n)...g2,l(n)...g2,L(n)}
D5, tap weights vector update
The tap weights vector W of the big step-length filter of subsequent time n+11(n+1) it is obtained by following formula, W1(n+1)=W1(n)+
μ1G1(n)U(n)[UT(n)G1(n)U(n)+δIP]-1E1(n):
If threshold value σ of the hybrid parameter a (n) of current time n more than or equal to hybrid parameter a (n), σ=4, then down for the moment
Carve the tap weights vector W of the small step-length filter of n+12(n+1) are as follows: W2(n+1)=W (n)+μ2G2(n)U(n)[UT(n)G2(n)U
(n)+δIP]-1E2(n)；
Otherwise, the tap weights vector W of the small step-length filter of subsequent time n+12(n+1) are as follows:
W2(n+1)=W2(n)+μ2G2(n)U(n)[UT(n)G2(n)U(n)+δIP]-1E2(n)；
Wherein, μ1For the step-length of big step-length filter, value range is 0.5~2.0；μ2The step-length of small step-length filter, takes
Being worth range is 0.005~0.20；δ is the normal number for preventing matrix inversion dyscalculia, and value is usually 0.001~0.01, IP
For the unit matrix of P × P；
E, filter weight updates
The hybrid parameter a (n+1) of subsequent time n+1 is obtained by following formula:
A (n+1)=a (n)+μαλ(n)(1-λ(n))e(n)(y1(n)-y2(n))
Wherein, μαIt is the change step value of hybrid parameter, value 1000；
And then update and obtain the weight coefficient λ (n+1) of the big step-length filter of subsequent time n+1,
F, the weight of filter limits
If the hybrid parameter a (n+1) at n+1 moment is less than or equal to the negative of threshold value σ=4, i.e. a (n+1) <=- 4 then enables a
(n+1)=- 4, λ (n+1)=0；If the hybrid parameter a (n+1) at n+1 moment is more than or equal to threshold value σ, i.e. >=4 a (n+1) then enable
A (n+1)=4, λ (n+1)=1；Otherwise, directly turn G step；
G, the step of enabling n=n+1, repeating A, B, C, D, E, F, until end of conversation.
Compared with prior art, the beneficial effects of the present invention are:
The tap weight coefficient w of two sef-adapting filters at current time of the invention1,l(n) and w2,l(n) step-length
Proportional coefficient is respectively g1,i(n) and g2,i(n), g1,i(n)=max { ρ1,|w1,i(n)-w1,i(kM) | }, g2,i(n)=max
{ρ2,|w2,i(n)-w2,i(kM) | }, i.e., within the period (time window) that the time is M, the sef-adapting filter at current time is taken out
The tap weights of the proportional coefficient of step-length the tap weight coefficient equal to current time n and the period initial time of head weight coefficient
The difference of coefficient.It is also larger to the difference of big coefficient calculating when initial, so as to obtain faster initial convergence speed；And it connects
The difference becomes smaller when nearly stable state, so that the step-length of tap weight coefficient also accordingly becomes smaller when close to stable state, namely in target impulse
Faster convergence rate can be still obtained when responding non-sparse.Pass through the filtering of the fast convergence and small step-length of big step-length filter
The low steady-state error of device, obtains for echo cancellor that not only fast convergence is simultaneously but also steady-state error is low provides guarantor after the two convex combination
Card.During convex combination, weight is big in the early stage for big step-length filter, and junction filter y (n) is mainly big step-length filter value,
And be mainly that small step-length filter weight is big in the later period, junction filter is mainly small step-length filter value, by error signal e (n)
It influences relatively small, under the conditions of weight variation tendency is metastable, big step-length filter will be given full play to by convex combination
Fast convergence and small step-length low steady-state error the advantages of.
Result after convex combination operation is updated into the pumping of big adaptive step filter as direction vector is updated respectively
Head weight coefficient w1(n) and the tap weight coefficient w of small step-length sef-adapting filter2(n), the adaptive filter then by the two combined
The tap weight coefficient w (n) of wave device, value can be more preferable closer to true value, echo cancellor effect.
In short, method accommodative ability of environment of the invention is strong, fast convergence rate and low steady-state error.
Present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.
Detailed description of the invention
Near end signal d (n) figure used when Fig. 1 is emulation experiment of the present invention.
Fig. 2 is the shock response figure of emulation experiment of the present invention.
Fig. 3 be document 1 method and normalization steady output rate curve of the invention.
Specific embodiment
Embodiment
A kind of specific embodiment of the invention is a kind of proportional radial projection echo cancellor side of convex combination coefficient difference
Method, its step are as follows:
A, remote signaling sampling filter
A1, the remote end input signal for obtaining the remote signaling distally transmitted in current time n and preceding L-1 instance sample
Discrete value x (n), x (n-1) ..., x (n-L+1) constitute the input vector X (n), i.e. X of the current time n of convex combination filter
(n)=[x (n), x (n-1) ..., x (n-L+1)]T；Wherein, L=64 is filter tap number, and subscript T represents transposition operation；
A2, by the input vector X (n) of current time n by respectively obtaining the big step of current time n after convex combination filter
Long filtering output value y1(n), y1(n)=W1 T(n) the small step-length filtering output value y of X (n) and current time n2(n): y2(n)=W2 T
(n)X(n)；
Wherein W1(n)=[w1,1(n),w1,2(n),...,w1,l(n),...,w1,L(n)]TIt is current in convex combination filter
The tap weights vector of the big step-length filter of moment n, initial value are null vector, w1,l(n) it is filtered for the big step-length of current time n
First of tap weight coefficient in the tap weights vector of wave device；W2(n)=[w2,1(n),w2,2(n),...,w2,l(n),...,w2,L
(n)]TFor the tap weights vector of the small step-length filter of current time n in convex combination filter, initial value is null vector, w2,l
It (n) is first of tap weight coefficient in the tap weights vector of the small step-length filter of current time n；
B, convex combination
By the big step-length output valve y of current time n1(n) and the small step-length output valve y of current time n2(n) convex combination is carried out
Obtain the output valve y (n) of convex combination filter, y (n)=λ (n) y1(n)+(1-λ(n))y2(n)；By the big step-length of current time n
The tap weights vector W of filter1(n) and the small step-length filter tap weight vector W of current time n2(n) convex combination is carried out to obtain
Filter tap weight vector W (n): W (n)=λ (n) W1(n)+(1-λ(n))W2(n)；Wherein λ (n) is the big step-length of current time n
The weight coefficient of filter, its calculation formula isA (n) is the hybrid parameter of current time n, initial value
It is 0；
C, echo cancelltion
The convex combination of the near end signal d (n) and current time n of the current time n with echo that proximal end is picked up filter
Device output valve y (n) subtracts each other, total residual signals e (n) after the echo that is eliminated, e (n)=d (n)-y (n)；And total residual error is believed
Number send back to distal end.
Meanwhile by the big step-length output valve y of the near end signal d (n) of current time n and current time n1(n) subtract each other, obtain
The big step-length residual signals e of current time n1(n), e1(n)=d (n)-y1(n)；By the near end signal d (n) of current time n with work as
The small step-length output valve y of preceding moment n2(n) subtract each other, obtain the small step-length residual signals e of current time n2(n), e2(n)=d (n)-
y2(n)；
D, filter tap weight vector updates
The composition of D1, affine projection input vector
By convex combination filter in the current time n and input vector X (n) at preceding P-1 moment, X (n-1) ..., X (n-P
+ 1) affine projection input vector U (n), U (n)=[X (n), X (n-1) ... the X (n-P+1)] of current time n, are constituted, wherein P
Represent affine projection order, P=4,8,16；
The composition of D2, residual signals vector
By current time n and total residual signals e (n) at preceding P-1 moment, e (n-1) ..., e (n-P+1), composition is worked as
Total residual signals vector E (n) of preceding moment n, E (n)=[e (n), e (n-1) ..., e (n-P+1)]T；
By the current time n and big step-length residual signals e at preceding P-1 moment1(n),e1(n-1),...,e1(n-P+1), structure
At the big step-length residual signals vector E of current time n1(n), E1(n)=[e1(n),e1(n-1),...,e1(n-P+1)]T；
By the current time n and small step-length residual signals e at preceding P-1 moment2(n),e2(n-1),...,e2(n-P+1), structure
At the small step-length residual signals vector E of current time n2(n), E2(n)=[e2(n),e2(n-1),...,e2(n-P+1)]T；
The calculating of D3, the proportional matrix of big step-length
Calculate the proportional coefficient g of first big step-length of current time n1,l(n):
g1,l(n)=max { ρ1,|w1,l(n)-w1,l(kM)|}
Wherein max indicates maximum operation, | | indicate the operation that takes absolute value, ρ1For the tap weights system of big step-length filter
Several threshold values, value range be 0.001~0.01, M be the length of time window, its value range is that 100~200, k is the time
The serial number of window, Indicate downward round numbers operation；
The proportional coefficient g of the big step-length of l-th is arrived by first of current time n1,1(n),g1,2(n)...g1,l(n)...,
g1,L(n), the proportional matrix G of big step-length is constituted1(n), G1(n)=diag { g1,1(n),g1,2(n)...g1,l(n)...g1,L(n) },
Wherein, diag indicates construction diagonal matrix；
D4, small step grow up to the calculating of scaling matrices
First of small step for calculating current time n grows up to proportionality coefficient g2,l(n):
g2,l(n)=max { ρ2,|w2,l(n)-w2,l(kM)|}
Wherein, ρ2For the threshold value of the tap weight coefficient of small step-length filter, value range is 0.001~0.01；
Grow up to proportionality coefficient g to l-th small step for first of current time n2,1(n),g2,2(n) ... g2,l(n),...,
g2,L(n), it constitutes small step and grows up to scaling matrices G2(n):
G2(n)=diag { g2,1(n),g2,2(n)...g2,l(n)...g2,L(n)}
D5, tap weights vector update
The tap weights vector W of the big step-length filter of subsequent time n+11(n+1) it is obtained by following formula, W1(n+1)=W1(n)+
μ1G1(n)U(n)[UT(n)G1(n)U(n)+δIP]-1E1(n):
If threshold value σ of the hybrid parameter a (n) of current time n more than or equal to hybrid parameter a (n), σ=4, then down for the moment
Carve the tap weights vector W of the small step-length filter of n+12(n+1) are as follows: W2(n+1)=W (n)+μ2G2(n)U(n)[UT(n)G2(n)U
(n)+δIP]-1E2(n)；
Otherwise, the tap weights vector W of the small step-length filter of subsequent time n+12(n+1) are as follows:
W2(n+1)=W2(n)+μ2G2(n)U(n)[UT(n)G2(n)U(n)+δIP]-1E2(n)；
Wherein, μ1For the step-length of big step-length filter, value range is 0.5~2.0；μ2The step-length of small step-length filter, takes
Being worth range is 0.005~0.20；δ is the normal number for preventing matrix inversion dyscalculia, and value is usually 0.001~0.01, IP
For the unit matrix of P × P；
E, filter weight updates
The hybrid parameter a (n+1) of subsequent time n+1 is obtained by following formula:
A (n+1)=a (n)+μαλ(n)(1-λ(n))e(n)(y1(n)-y2(n))
Wherein, μαIt is the change step value of hybrid parameter, value 1000；
And then update and obtain the weight coefficient λ (n+1) of the big step-length filter of subsequent time n+1,
F, the weight of filter limits
If the hybrid parameter a (n+1) at n+1 moment is less than or equal to the negative of threshold value σ=4, i.e. a (n+1) <=- 4 then enables a
(n+1)=- 4, λ (n+1)=0；If the hybrid parameter a (n+1) at n+1 moment is more than or equal to threshold value σ, i.e. >=4 a (n+1) then enable
A (n+1)=4, λ (n+1)=1；Otherwise, directly turn G step；
G, the step of enabling n=n+1, repeating A, B, C, D, E, F, until end of conversation.
Emulation experiment
In order to verify effectiveness of the invention, emulation experiment is carried out, and carry out emulation pair with the method for existing document 1
Than.
In emulation experiment sef-adapting filter tap length L be 64, remote end input signal use pole for 0.95 single order
Autoregression (AR (1)) signal, in a length of 6.25m in room, wide 3.75m, high 2.5m, temperature is 20 DEG C, and the peace and quiet of temperature 50% are close
It closing in room, received remote signaling is in a room 8kHz by sample frequency with microphone after loudspeaker plays by proximal end,
Sampling order is the 64 near end signal d (n) for sampling 20000 moment points altogether.
The above near end signal d (n) is used to method (the big step size mu of the method for the present invention and document 11=0.6 and small step size mu2=
0.005 two kinds of situations) carry out echo cancellor.The optimal value of the parameter of various methods in an experiment such as the following table 1.
The optimized parameter approximation value of each method method emulation experiment
| Method (the big step size mu of document 11=0.6) | μ1=0.6；δ=0.01；ρ1=0.01 |
| Method (the small step size mu of document 12=0.005) | μ2=0.005；δ=0.01；ρ2=0.005 |
| The present invention | μα=1000；δ=0.01；λ0=0.5；α0=0 |
Table 1
Simulation result is averagely obtained by independent operating 30 times.Fig. 1 is near end signal d (n) figure, and Fig. 2 is experiment of the present invention
Convex combination sef-adapting filter shock response figure, Fig. 3 is the long method of small step, large-step method and the method for the present invention of document 1
The normalization steady output rate curve of emulation experiment.
From figure 3, it can be seen that the present invention solves the contradiction of fast convergence and low steady-state error, it may be assumed that the present invention is kept
The fast convergence of the big step-length DPAPA method of document 1, and maintain the low steady-state error of the small step-length DPAPA method of document 1
Performance, and tracking ability of the present invention is strong.Under identical convergence rate, steady-state error of the invention is significantly more steady than PAPA algorithm
State error is small and tracking ability is strong.
Claims (1)
1. a kind of proportional affine projection echo cancel method of convex combination coefficient difference, its step are as follows:
A, remote signaling sampling filter
A1, the remote end input signal for obtaining the remote signaling distally transmitted in current time n and preceding L-1 instance sample are discrete
Value x (n), x (n-1) ..., x (n-L+1), the input vector X (n) of the current time n of composition convex combination filter, i.e. X (n)=
[x (n), x (n-1) ..., x (n-L+1)]T；Wherein, L=64 is filter tap number, and subscript T represents transposition operation；
A2, the input vector X (n) of current time n is filtered by respectively obtaining the big step-length of current time n after convex combination filter
Wave output valve y1(n), y1(n)=W1 T(n) the small step-length filtering output value y of X (n) and current time n2(n): y2(n)=W2 T(n)X
(n)；
Wherein W1(n)=[w1,1(n), w1,2(n) ..., w1, l(n) ..., w1, L(n)]TFor current time n in convex combination filter
Big step-length filter tap weights vector, initial value be null vector, w1, lIt (n) is the big step-length filter of current time n
First of tap weight coefficient in tap weights vector；W2(n)=[w2,1(n), w2,2(n) ..., w2, l(n) ..., w2, L(n)]TFor
The tap weights vector of the small step-length filter of current time n in convex combination filter, initial value are null vector, w2, lIt (n) is to work as
First of tap weight coefficient in the tap weights vector of the small step-length filter of preceding moment n；
B, convex combination
By the big step-length output valve y of current time n1(n) and the small step-length output valve y of current time n2(n) convex combination is carried out to obtain
The output valve y (n) of convex combination filter, y (n)=λ (n) y1(n)+(1-λ(n))y2(n)；The big step-length of current time n is filtered
The tap weights vector W of device1(n) and the small step-length filter tap weight vector W of current time n2(n) convex combination is carried out to be filtered
Device tap weights vector W (n): W (n)=λ (n) W1(n)+(1-λ(n))W2(n)；The big step-length that wherein λ (n) is current time n filters
The weight coefficient of device, its calculation formula isA (n) is the hybrid parameter of current time n, initial value 0；
C, echo cancelltion
The convex combination filter of the near end signal d (n) and current time n of the current time n with echo that proximal end is picked up are defeated
Value y (n) subtracts each other out, total residual signals e (n) after the echo that is eliminated, e (n)=d (n)-y (n)；And total residual signals are sent
Back to distal end；
Meanwhile by the big step-length output valve y of the near end signal d (n) of current time n and current time n1(n) subtract each other, obtain current
The big step-length residual signals e of moment n1(n), c1(n)=d (n)-y1(n)；By the near end signal d (n) of current time n and it is current when
Carve the small step-length output valve y of n2(n) subtract each other, obtain the small step-length residual signals e of current time n2(n), e2(n)=d (n)-y2
(n)；
D, filter tap weight vector updates
The composition of D1, affine projection input vector
By convex combination filter in the current time n and input vector X (n) at preceding P-1 moment, X (n-1) ..., X (n-P+1),
Affine projection input vector U (n), U (n)=[X (n), X (n-1) ... the X (n-P+1)] of current time n are constituted, wherein P is represented
Affine projection order, P=4,8,16；
The composition of D2, residual signals vector
By current time n and total residual signals e (n) at preceding P-1 moment, e (n-1) ..., e (n-P+1) constitute current time
Total residual signals vector E (n) of n, E (n)=[e (n), e (n-1) ..., e (n-P+1)]T；
By the current time n and big step-length residual signals e at preceding P-1 moment1(n), e1(n-1) ..., e1(n-P+1), composition is worked as
The big step-length residual signals vector E of preceding moment n1(n), E1(n)=[e1(n), e1(n-1) ..., e1(n-P+1)]T；
By the current time n and small step-length residual signals e at preceding P-1 moment2(n), e2(n-1) ..., e2(n-P+1), composition is worked as
The small step-length residual signals vector E of preceding moment n2(n), E2(n)=[e2(n), e2(n-1) ..., e2(n-P+1)]T；
The calculating of D3, the proportional matrix of big step-length
Calculate the proportional coefficient g of first big step-length of current time n1, l(n):
g1, l(n)=max { ρ1, | w1, l(n)-w1, l(kM)|}
Wherein max indicates maximum operation, | | indicate the operation that takes absolute value, ρ1For the tap weight coefficient of big step-length filter
Threshold value, value range be 0.001~0.01, M be the length of time window, its value range is that 100~200, k is time window
Serial number, Indicate downward round numbers operation；
The proportional coefficient g of the big step-length of l-th is arrived by the 1st of current time n1,1(n), g1,2(n)...g1, l(n) ..., g1, L
(n), the proportional matrix G of big step-length is constituted1(n), G1(n)=diag { g1,1(n), g1,2(n)...g1, l(n)...g1, L(n) },
In, diag indicates construction diagonal matrix；
D4, small step grow up to the calculating of scaling matrices
First of small step for calculating current time n grows up to proportionality coefficient g2, l(n):
g2, l(n)=max { ρ2, | w2, l(n)-w2, l(kM)|}
Wherein, ρ2For the threshold value of the tap weight coefficient of small step-length filter, value range is 0.001~0.01:
Grow up to proportionality coefficient g to l-th small step for the 1st of current time n2,1(n), g2,2(n) ... g2, l(n) ..., g2, L
(n), it constitutes small step and grows up to scaling matrices G2(n):
G2(n)=diag { g2,1(n), g2,2(n)...g2, l(n)...g2, L(n)}
D5, tap weights vector update
The tap weights vector W of the big step-length filter of subsequent time n+11(n+1) it is obtained by following formula, W1(n+1)=W1(n)+μ1G1
(n)U(n)[UT(n)G1(n)U(n)+δIP]-1E1(n):
If the hybrid parameter a (n) of current time n is more than or equal to the threshold value σ of hybrid parameter a (n), σ=4, then subsequent time n+1
Small step-length filter tap weights vector W2(n+1) are as follows: W2(n+1)=W (n)+μ2G2(n)U(n)[UT(n)G2(n)U(n)+δ
IP]-1E2(n)；
Otherwise, the tap weights vector W of the small step-length filter of subsequent time n+12(n+1) are as follows:
W2(n+1)=W2(n)+μ2G2(n)U(n)[UT(n)G2(n)U(n)+δIP]-1E2(n)；
Wherein, μ1For the step-length of big step-length filter, value range is 0.5~2.0；μ2The step-length of small step-length filter, value model
Enclose is 0.005~0.20；δ is the normal number for preventing matrix inversion dyscalculia, and value is usually 0.001~0.01, IPFor P ×
The unit matrix of P；
E, filter weight updates
The hybrid parameter a (n+1) of subsequent time n+1 is obtained by following formula:
A (n+1)=a (n)+μαλ(n)(1-λ(n))e(n)(y1(n)-y2(n))
Wherein, μαIt is the change step value of hybrid parameter, value 1000；
And then update and obtain the weight coefficient λ (n+1) of the big step-length filter of subsequent time n+1,
F, the weight of filter limits
If the hybrid parameter a (n+1) at n+1 moment is less than or equal to the negative of threshold value σ=4, i.e. a (n+1) <=- 4 then enables a (n+1)
=-4, λ (n+1)=0；If the hybrid parameter a (n+1) at n+1 moment is more than or equal to threshold value σ, i.e. >=4 a (n+1) then enable a (n+
1)=4, λ (n+1)=1；Otherwise, directly turn G step；
G, the step of enabling n=n+1, repeating A, B, C, D, E, F, until end of conversation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610326009.0A CN106060295B (en) | 2016-05-17 | 2016-05-17 | A kind of proportional affine projection echo cancel method of convex combination coefficient difference |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610326009.0A CN106060295B (en) | 2016-05-17 | 2016-05-17 | A kind of proportional affine projection echo cancel method of convex combination coefficient difference |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106060295A CN106060295A (en) | 2016-10-26 |
| CN106060295B true CN106060295B (en) | 2019-05-24 |
Family
ID=57177694
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610326009.0A Active CN106060295B (en) | 2016-05-17 | 2016-05-17 | A kind of proportional affine projection echo cancel method of convex combination coefficient difference |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106060295B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107819963B (en) * | 2017-09-13 | 2019-11-26 | 西南交通大学 | A kind of equal cube absolute value echo cancel method of the minimum of convex combination |
| CN109102794A (en) * | 2018-07-24 | 2018-12-28 | 西南交通大学 | M based on convex combination estimates the echo cancel method of proportional class affine projection |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103077726A (en) * | 2012-01-09 | 2013-05-01 | 美商威睿电通公司 | Pretreatment and after-treatment for linear acoustic echo elimination system |
| CN104392727A (en) * | 2014-11-12 | 2015-03-04 | 华为技术有限公司 | Audio signal processing method and related device |
| CN104410761A (en) * | 2014-09-13 | 2015-03-11 | 西南交通大学 | Convex combination adaptive echo cancellation method for affine projection sign subband adaptive filter |
-
2016
- 2016-05-17 CN CN201610326009.0A patent/CN106060295B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103077726A (en) * | 2012-01-09 | 2013-05-01 | 美商威睿电通公司 | Pretreatment and after-treatment for linear acoustic echo elimination system |
| CN104410761A (en) * | 2014-09-13 | 2015-03-11 | 西南交通大学 | Convex combination adaptive echo cancellation method for affine projection sign subband adaptive filter |
| CN104392727A (en) * | 2014-11-12 | 2015-03-04 | 华为技术有限公司 | Audio signal processing method and related device |
Non-Patent Citations (1)
| Title |
|---|
| Affine projection M-estimate subband adaptive filters for robust adaptive filtering in impulsive noise;Zongsheng Zheng et al.;《SIGNAL PROCESSING》;20160331;64-70 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106060295A (en) | 2016-10-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109040499B (en) | Adaptive echo cancellation method for resisting impact interference | |
| CN104683614B (en) | Based on the proportional illumination-imitation projection self-adoptive echo cancel method of memory that M estimates | |
| CN105391879B (en) | A kind of acoustic echo removing method of echoless residual dual end communication robust | |
| CN103369162B (en) | A kind of listener's echo self adaptive elimination method of low complex degree | |
| CN104506746B (en) | A kind of proportional adaptive echo cancellation method of convex combination decorrelation of improvement | |
| CN110138990A (en) | A method of eliminating mobile device voip phone echo | |
| CN109102794A (en) | M based on convex combination estimates the echo cancel method of proportional class affine projection | |
| CN110572525B (en) | Self-adaptive communication echo cancellation method for voice communication | |
| CN105721729B (en) | Based on the sparse proportional echo cancel method for reusing weight coefficient affine projection of block | |
| CN106210370A (en) | A kind of affine projection maximum entropy sub-band adaptive echo cancel method | |
| CN104270539A (en) | Proportional affine projection echo elimination method based on coefficient difference | |
| CN104410761A (en) | Convex combination adaptive echo cancellation method for affine projection sign subband adaptive filter | |
| CN109040497A (en) | A kind of proportional class illumination-imitation projection self-adoptive echo cancel method based on M estimation | |
| CN110191245B (en) | Self-adaptive echo cancellation method based on time-varying parameters | |
| CN104144269B (en) | A kind of proportional self adaptation listener's echo removing method based on decorrelation | |
| CN107819963B (en) | A kind of equal cube absolute value echo cancel method of the minimum of convex combination | |
| CN104601837A (en) | Robust convex combination type adaptive phone echo canceling method | |
| CN106060295B (en) | A kind of proportional affine projection echo cancel method of convex combination coefficient difference | |
| CN108877824B (en) | Combined step echo cancellation method with high tracking performance | |
| CN106412352B (en) | A kind of time-frequency memory subband ratio adaptive echo cancellation method | |
| CN105491256A (en) | Robust step size adjustment method in initialization phase of acoustic echo cancellation | |
| CN107610714A (en) | The echo cancel method of the minimum cube absolute value attracted based on a norm zero | |
| CN106170049B (en) | A kind of normalization sub-band adaptive echo cancel method having offset compensation | |
| CN109151237B (en) | Affine projection self-adaptive echo cancellation method based on zero attraction | |
| CN101568058A (en) | Digital hearing aid echo path estimation method based on weighted subgradient projection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |