CN104410761B - A kind of affine projection symbol subband convex combination adaptive echo cancellation method - Google Patents

Abstract

A kind of affine projection symbol subband convex combination adaptive echo cancellation method, its step mainly: A, remote signaling filtering, obtain the output vector Y of large step-length filter ₁(n), and the output vector Y of little step-length filter ₂(n); B, convex combination, by the output vector Y of large step-length filter and little step-length filter ₁(n) and Y ₂n () is carried out convex combination and is obtained combined filter value Y (n), Y (n)=λ (n) Y ₁(n)+(1-λ (n)) Y ₂n (), by the tap weights of large step-length filter vector W ₁the tap weights vector W of (n) and little step-length filter ₂(n) also carry out convex combination obtain always bank of filters tap weights vector W (n), W (n)=λ (n) W ₁(n)+(1-λ (n)) W ₂(n), C, echo cancelltion, near-end microphone pickup is obtained net signal E (n) to being with affine projection near end signal D (n) of echo and combined filter value Y (n) to subtract each other, E (n)=D (n)-Y (n), loopback is to far-end, D, filter tap weight coefficient upgrade, and the weight of E, filter upgrades, and hybrid parameter a (n) upgrades; F, make n=n+1, repeat above-mentioned steps, until end of conversation.It is strong that the acoustic echo of this method to this Sparse System of communication eliminates ability, fast convergence rate, and steady-state error is little; Echo cancellor is effective.

Description

A kind of affine projection symbol subband convex combination adaptive echo cancellation method

Technical field

The invention belongs to the adaptive echo cancellation techniques field of communication.

Background technology

In recent years, along with the development of various emerging communication, the care of communication user to speech quality grows with each passing day.In a communications system, the problem of noise and echo interference can often be encountered.In echo interference problem, acoustic echo (Echo) is the main factor affecting voice call quality.Acoustic echo in communication process refers to that user repeatedly hears oneself sound in communication process.This phenomenon is coupled by acoustic reflection and the acoustics between microphone and loud speaker to produce.Acoustic echo time delay is short, have multipath, time become feature.In the communications, people's ear is again for this echo and sensitivity thereof, and the echo postponing 10ms just can be caught by people's ear and perceive, and will cause great interference more than the echo of 32ms to communication quality.

At present, for the removing method of acoustic echo, remain a popular research topic.Be summed up, the method for echo cancellor roughly has 6 kinds both at home and abroad at present: (1) frequency shift technique; (2) subband center chopping techniques; (3) speech control switching technique: (4) comb filtering technology; (5) microphone array technology; (6) adaptive echo technology for eliminating.Front 5 kinds of Method means complexity, technical requirement is high, involves great expense.Adaptive echo cancellation method cost is low, is the promising echo cancellation technology of putative most, is also the mainstream technology of present echo cancellor.

So-called adaptive echo technology for eliminating is a kind of echo cancel method utilizing sef-adapting filter.Sef-adapting filter can independently adjust according to input signal the digital filter that performance carries out Digital Signal Processing, as a comparison, the parameter of traditional some non-self-adapting filters (as FIR filter, iir filter) is that presetting of static state is nonadjustable.For the environment of some the unknowns, sef-adapting filter can progressively learn out required statistical property in the course of the work, and automatically adjusts filter coefficient on this basis, to reach the effect of optimum filtering; Once the statistical property of input signal changes, it can follow the tracks of this change again, automatically adjusts filter parameter, makes filtering performance again reach best.In adaptive echo cancellation method, traditional LMS (lowest mean square) filter, affine projection APA filter, recurrence least square RLS filter etc., because structure is simple, be easy to realization, obtained the extensive concern of domestic and international researcher.But the echo channel in communication, great majority are all condition of sparse channel, and the impulse response of this condition of sparse channel is mostly very little, and exponent number is long.The echo cancellation convergence of tradition sef-adapting filter under this condition of sparse channel is slow, and steady-state error is large.And the echo in reality is coherent signal mostly, introduces sub-filter for this situation, the disposal ability of coherent signal can be improved, improve convergence rate.Further, the filter of fixed step size has the intrinsic contradictions of convergence rate and steady-state error, introduce convex combination strategy, by a fast filter and a slow filter convex combination, can ensure that filter has convergence rate and lower steady-state error faster simultaneously.

In reality call, often there will be the situation of dual end communication, namely the caller at communication system two ends talks simultaneously.This situation is equivalent to microphones and has arrived very large noise (here using all sound except echo all as noise), and existing adaptive filter algorithm is very responsive to the change of noise, often needs bilateral to talk about the auxiliary of detector DTD.Further, when dual end communication, the constringency performance of echo cancellor adaptive algorithm sharply declines.It is one of effective way improving algorithm antijamming capability by affine projection subband algorithm created symbol algorithm, this kind of algorithm is only relevant to the symbol of noise, have nothing to do with the size of noise, reduce the susceptibility that algorithm changes noise, and without the need to bilateral words detector DTD, while reducing cost, enhance the robustness of algorithm to dual end communication.

In the application of current Sparse System identification, more ripe method has following two kinds:

(1) a kind of symbol subband (VRP-SSAF) echo cancel method becoming regularization parameter

List of references 1 " Variableregularisationparametersignsubbandadaptivefilter " (J.Ni and F.Li, Electron.Lett., vol.46, no.24, pp.1605 – 1607, Nov.2010.) the method is added by symbol subband (SSAF) algorithm to become regularization parameter strategy, reduces the impact of fixed step size on convergence rate and steady-state error, be better than symbol subband (SSAF) echo cancel method under bilateral speech phase.But the steady-state error of this algorithm is still very large.

(2) affine projection symbol subband (APSSAF) echo cancel method

List of references 2 " Twovariantsofthesignsubbandadaptivefilterwithimprovedcon vergencerate " (J.Ni, X.Chen and J.Yang, SignalProcess., vol.96, pp.325 – 331, May.2014).This algorithm is in affine projection method created symbol subband (SSAF) method, and further increase the echo cancellor effect under the two call scenarios of condition of sparse channel, the more original algorithm of initial convergence speed is compared and improved.But still there are the intrinsic contradictions of sef-adapting filter convergence rate and steady-state error in this algorithm.

Summary of the invention

Goal of the invention of the present invention is just to provide a kind of affine projection symbol subband convex combination adaptive echo cancellation method, and the method is good to the eradicating efficacy of the acoustic echo of communication system, fast convergence rate, and steady-state error is little.

The present invention realizes the technical scheme that its goal of the invention adopts, a kind of affine projection symbol subband convex combination adaptive echo cancellation method, and its step is as follows:

A, remote signaling filtering

The remote signaling sampling transmitted by far-end obtains the centrifugal pump u (n) of the current time n of remote signaling, remote signaling is at current time n and the centrifugal pump u (n) in a front L-1 moment, u (n-1) ..., u (n-L+1), form sub-filter input vector U (n) of current time n, U (n)=[u (n), u (n-1) ..., u (n-L+1)] ^t; Wherein L=512, be filter tap number, subscript T represents transposition;

Again by current time n and sub-filter input vector U (n) in a front P-1 moment, U (n-1) ... U (n-P+1) combines, obtain sub-filter affine projection input vector G (n), G (n)=[U (n), U (n-1), ..., U (n-P+1)], wherein P represents affine projection exponent number, P=4,8,16;

Subsequently, the analysis filterbank that sub-filter affine projection input vector G (n) is eliminated in filter by subband convex combination adaptive echo is obtained i-th sub-filter affine projection input vector G _i(n)=[U _i(n), U _i(n-1) ..., U _i(n-P+1)], i is the sequence number that subband convex combination adaptive echo eliminates filter, i=1,2,3 ... N, N<32;

I-th sub-filter affine projection input vector G _in () obtains after eliminating filter filtering by subband convex combination adaptive echo respectively: the output vector Y of the large step-length filter of the i-th subband of current time _{i, 1}(n), Y _{i, 1}(n)=W ₁(n) ^tg _i(n)=[y _{i, 1}(n), y _{i, 1}(n-1) ..., y _{i, 1}(n-P+1)]; The output vector Y of the little step-length filter of the i-th subband _{i, 2}(n), Y _{i, 2}(n)=W ₂(n) ^tg _i(n)=[y _{i, 2}(n), y _{i, 2}(n-1) ..., y _{i, 2}(n-P+1)]; Wherein, W ₁(n) and W ₂n convex combination adaptive echo that () is respectively current time n eliminates the tap weights vector of large step-length filter in filter each subband and little step-length filter, and initial value is zero;

By the output vector Y of the large step-length filter of the i-th subband _{i, 1}the output vector Y of (n) and the little step-length filter of the i-th subband _{i, 2}n () obtains the output vector Y of large step-length filter after eliminating the synthesis filter banks in filter respectively by subband convex combination adaptive echo ₁(n), Y ₁(n)=W ₁(n) ^tthe output vector Y of the little step-length filter of G (n) and current time ₂(n), Y ₂(n)=W ₂(n) ^tg (n);

B, convex combination

By the output vector Y of the large step-length filter of current time ₁the output vector Y of the little step-length filter of (n) and current time ₂n () carries out the output vector Y (n) that convex combination obtains the junction filter of current time, Y (n)=λ (n) Y ₁(n)+[1-λ (n)] Y ₂(n); Wherein, λ (n) is the weight of the large step-length filter of current time, and its expression formula is a (n) is hybrid parameter, and its initial value is 0;

By the tap weights of the large step-length filter of current time vector W ₁the tap weights vector W of the little step-length filter of (n) and current time ₂n () also carries out tap weights vector W (n) that convex combination obtains the junction filter of current time, W (n)=λ (n) W ₁(n)+[1-λ (n)] W ₂(n);

C, echo cancelltion

Near end signal centrifugal pump d (n) of the band echo in the current time n that near-end microphone pickup is arrived and a front P-1 moment, d (n-1), ..., d (n-P+1) combines, obtain affine projection near-end vector D (n) of current time, D (n)=[d (n), d (n-1), ..., d (n-P+1)]; Net signal E (n) of the current time after echo cancellor is obtained after being subtracted each other with the output vector Y (n) of the junction filter of current time by the affine projection near-end of current time vector D (n), E (n)=D (n)-Y (n)=[e (n), e (n-1), ..., e (n-P+1)], by the net signal loopback of current time to far-end;

D, filter tap weight vector upgrade

By the affine projection near-end of current time vector D (n), respectively with the output vector Y of the large step-length filter of current time ₁the output vector Y of the little step-length filter of (n), current time ₂n () subtracts each other, obtain the large step-length net signal E of current time ₁the little step-length net signal E of (n) and current time ₂(n), that is:

E ₁(n)＝D(n)-Y ₁(n)，E ₂(n)＝D(n)-Y ₂(n)；

Exploitation right vector transition strategy calculates the tap weights vector W of large step-length filter in the adaptive echo elimination filter of subsequent time n+1 ₁(n+1), $W_1(n+1)=W_1\left(n\right)+{\mathrm{\&mu;}}_1\frac{G\left(n\right)\mathrm{sgn}\left[E_1\left(n\right)\right]}{{\left|\right|G\left(n\right)\mathrm{sgn}\left[E_1\left(n\right)\right]\left|\right|}_2+\mathrm{\&delta;}};$ Wherein, μ ₁for the step-length of large step-length filter, its value is 0.05 ~ 0.1; δ is regularization parameter, and its value is 0.001 ~ 0.01; || || ₂represent 2 norms; Sgn is symbolic operation function, and when being positive number by operation values, its value is 1; When being 0 by operation values, its value is 0, and when being negative by operation values, its value is-1;

Exploitation right vector transition strategy calculates the tap weights vector W of the medium and small step-length filter of adaptive echo elimination filter of subsequent time n+1 simultaneously ₂(n+1);

As hybrid parameter a (n)>=a ⁺time: $W_2(n+1)=W\left(n\right)+{\mathrm{\&mu;}}_2\frac{G\left(n\right)\mathrm{sgn}\left[E_2\left(n\right)\right]}{{\left|\right|G\left(n\right)\mathrm{sgn}\left[E_2\left(n\right)\right]\left|\right|}_2+\mathrm{\&delta;}}$

As hybrid parameter a (n) <a ⁺time, $W_2(n+1)=W_2\left(n\right)+{\mathrm{\&mu;}}_2\frac{G\left(n\right)\mathrm{sgn}\left[E_2\left(n\right)\right]}{{\left|\right|G\left(n\right)\mathrm{sgn}\left[E_2\left(n\right)\right]\left|\right|}_2+\mathrm{\&delta;}}$

Wherein: a ⁺be a normal number, its value is 4 ~ 5; μ ₂for the step-length of little step-length filter, its value is 0.005 ~ 0.01;

The weight of E, filter upgrades

Formula after hybrid parameter a (n) is simplified by sign function upgrades:

$a(n+1)=a\left(n\right)+{\mathrm{\&mu;}}_a\mathrm{\&lambda;}\left(n\right)[1-\mathrm{\&lambda;}\left(n\right)]\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}e\left(n\right)[y_{i,1}\left(n\right)-y_{i,2}\left(n\right)]$

Wherein N represents the number of subband, μ _aa constant, first element of value to be 0.1, e (n) be current time net signal E (n), y _{i, 1}n () is the output vector Y of the large step-length filter of current time i-th subband _{i, 1}first element in (n), y _{i, 2}n () is the output vector Y of the large step-length filter of current time i-th subband _{i, 2}first element in (n);

Hybrid parameter a (n+1) after renewal is substituted into the weight expression formula of step B in, obtain the filter weight λ (n+1) of subsequent time n+1,

F, iteration

Make n=n+1, repeat the step of A, B, C, D, E, until end of conversation.

Compared with prior art, the invention has the beneficial effects as follows:

The output vector Y (n) of junction filter is the estimated value of echo signal, itself and near-end microphone pickup, affine projection near end signal D (n) that processes the band echo obtained are subtracted each other, be proximally loopback to the signal of far-end, this signal is net signal E (n) after eliminating echo.The convergence rate of net signal is faster, and its echo cancellor effect is better.This patent adopts sub-filter, can multiple sub-band is divided into carry out filtering respectively the remote end input signal of Whole frequency band, then by filtered signal combination, its good wave filtering effect, fast convergence rate.Affine projection, by affine for input signal combination, can improve convergence rate further, resists stronger signal disturbing.And by the Fast Convergent of large step-length filter and the little steady-state error of little step-length filter, the fast convergence rate while of ensure that the steady-state error of the whole convex combination adaptive echo elimination filter after the two convex combination is little.

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail

Accompanying drawing explanation

Fig. 1 is the condition of sparse channel figure of emulation experiment of the present invention.

Fig. 2 is near end signal (voice) in emulation experiment of the present invention (bilateral words) and remote signaling (voice) figure.

Fig. 3 is document 1, the normalization steady output rate curve of document 2 and emulation experiment of the present invention.

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail.

Embodiment

Embodiment

A kind of embodiment of the present invention is, a kind of affine projection symbol subband convex combination adaptive echo cancellation method, and its step is as follows:

A, remote signaling filtering

The remote signaling sampling transmitted by far-end obtains the centrifugal pump u (n) of the current time n of remote signaling, remote signaling is at current time n and the centrifugal pump u (n) in a front L-1 moment, u (n-1) ..., u (n-L+1), form sub-filter input vector U (n) of current time n, U (n)=[u (n), u (n-1) ..., u (n-L+1)] ^t; Wherein L=512, be filter tap number, subscript T represents transposition;

Again by current time n and sub-filter input vector U (n) in a front P-1 moment, U (n-1) ... U (n-P+1) combines, obtain sub-filter affine projection input vector G (n), G (n)=[U (n), U (n-1), ..., U (n-P+1)], wherein P represents affine projection exponent number, P=4,8,16;

Subsequently, the analysis filterbank that sub-filter affine projection input vector G (n) is eliminated in filter by subband convex combination adaptive echo is obtained i-th sub-filter affine projection input vector G _i(n)=[U _i(n), U _i(n-1) ..., U _i(n-P+1)], i is the sequence number that subband convex combination adaptive echo eliminates filter, i=1,2,3 ... N, N<32;

I-th sub-filter affine projection input vector G _in () obtains after eliminating filter filtering by subband convex combination adaptive echo respectively: the output vector Y of the large step-length filter of the i-th subband of current time _{i, 1}(n), Y _{i, 1}(n)=W ₁(n) ^tg _i(n)=[y _{i, 1}(n), y _{i, 1}(n-1) ..., y _{i, 1}(n-P+1)]; The output vector Y of the little step-length filter of the i-th subband _{i, 2}(n), Y _{i, 2}(n)=W ₂(n) ^tg _i(n)=[y _{i, 2}(n), y _{i, 2}(n-1) ..., y _{i, 2}(n-P+1)]; Wherein, W ₁(n) and W ₂n convex combination adaptive echo that () is respectively current time n eliminates the tap weights vector of large step-length filter in filter each subband and little step-length filter, and initial value is zero;

By the output vector Y of the large step-length filter of the i-th subband _{i, 1}the output vector Y of (n) and the little step-length filter of the i-th subband _{i, 2}n () obtains the output vector Y of large step-length filter after eliminating the synthesis filter banks in filter respectively by subband convex combination adaptive echo ₁(n), Y ₁(n)=W ₁(n) ^tthe output vector Y of the little step-length filter of G (n) and current time ₂(n), Y ₂(n)=W ₂(n) ^tg (n);

B, convex combination

By the output vector Y of the large step-length filter of current time ₁the output vector Y of the little step-length filter of (n) and current time ₂n () carries out the output vector Y (n) that convex combination obtains the junction filter of current time, Y (n)=λ (n) Y ₁(n)+[1-λ (n)] Y ₂(n); Wherein, λ (n) is the weight of the large step-length filter of current time, and its expression formula is a (n) is hybrid parameter, and its initial value is 0;

By the tap weights of the large step-length filter of current time vector W ₁the tap weights vector W of the little step-length filter of (n) and current time ₂n () also carries out tap weights vector W (n) that convex combination obtains the junction filter of current time, W (n)=λ (n) W ₁(n)+[1-λ (n)] W ₂(n);

C, echo cancelltion

Near end signal centrifugal pump d (n) of the band echo in the current time n that near-end microphone pickup is arrived and a front P-1 moment, d (n-1), ..., d (n-P+1) combines, obtain affine projection near-end vector D (n) of current time, D (n)=[d (n), d (n-1), ..., d (n-P+1)]; Net signal E (n) of the current time after echo cancellor is obtained after being subtracted each other with the output vector Y (n) of the junction filter of current time by the affine projection near-end of current time vector D (n), E (n)=D (n)-Y (n)=[e (n), e (n-1), ..., e (n-P+1)], by the net signal loopback of current time to far-end;

D, filter tap weight vector upgrade

By the affine projection near-end of current time vector D (n), respectively with the output vector Y of the large step-length filter of current time ₁the output vector Y of the little step-length filter of (n), current time ₂n () subtracts each other, obtain the large step-length net signal E of current time ₁the little step-length net signal E of (n) and current time ₂(n), that is:

E ₁(n)＝D(n)-Y ₁(n)，E ₂(n)＝D(n)-Y ₂(n)；

Exploitation right vector transition strategy calculates the tap weights vector W of large step-length filter in the adaptive echo elimination filter of subsequent time n+1 ₁(n+1), $W_1(n+1)=W_1\left(n\right)+{\mathrm{\&mu;}}_1\frac{G\left(n\right)\mathrm{sgn}\left[E_1\left(n\right)\right]}{{\left|\right|G\left(n\right)\mathrm{sgn}\left[E_1\left(n\right)\right]\left|\right|}_2+\mathrm{\&delta;}};$ Wherein, μ ₁for the step-length of large step-length filter, its value is 0.05 ~ 0.1; δ is regularization parameter, and its value is 0.001 ~ 0.01; || || ₂represent 2 norms; Sgn is symbolic operation function, and when being positive number by operation values, its value is 1; When being 0 by operation values, its value is 0, and when being negative by operation values, its value is-1;

Exploitation right vector transition strategy calculates the tap weights vector W of the medium and small step-length filter of adaptive echo elimination filter of subsequent time n+1 simultaneously ₂(n+1);

As hybrid parameter a (n)>=a ⁺time: $W_2(n+1)=W\left(n\right)+{\mathrm{\&mu;}}_2\frac{G\left(n\right)\mathrm{sgn}\left[E_2\left(n\right)\right]}{{\left|\right|G\left(n\right)\mathrm{sgn}\left[E_2\left(n\right)\right]\left|\right|}_2+\mathrm{\&delta;}}$

As hybrid parameter a (n) <a ⁺time, $W_2(n+1)=W_2\left(n\right)+{\mathrm{\&mu;}}_2\frac{G\left(n\right)\mathrm{sgn}\left[E_2\left(n\right)\right]}{{\left|\right|G\left(n\right)\mathrm{sgn}\left[E_2\left(n\right)\right]\left|\right|}_2+\mathrm{\&delta;}}$

Wherein: a ⁺be a normal number, its value is 4 ~ 5; μ ₂for the step-length of little step-length filter, its value is 0.005 ~ 0.01;

The weight of E, filter upgrades

Formula after hybrid parameter a (n) is simplified by sign function upgrades:

$a(n+1)=a\left(n\right)+{\mathrm{\&mu;}}_a\mathrm{\&lambda;}\left(n\right)[1-\mathrm{\&lambda;}\left(n\right)]\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}e\left(n\right)[y_{i,1}\left(n\right)-y_{i,2}\left(n\right)]$

Wherein N represents the number of subband, μ _aa constant, first element of value to be 0.1, e (n) be current time net signal E (n), y _{i, 1}n () is the output vector Y of the large step-length filter of current time i-th subband _{i, 1}first element in (n), y _{i, 2}n () is the output vector Y of the large step-length filter of current time i-th subband _{i, 2}first element in (n);

Hybrid parameter a (n+1) after renewal is substituted into the weight expression formula of step B in, obtain the filter weight λ (n+1) of subsequent time n+1,

F, iteration

Make n=n+1, repeat the step of A, B, C, D, E, until end of conversation.

Emulation experiment:

In order to verify validity of the present invention, carry out emulation experiment, and contrasted with existing document 1,2 algorithm.

Remote signaling u (n) of emulation experiment is voice signal, and sample frequency is 8000Hz, sampled point number 40000.Echo channel impulse response is at high 2.5m, and wide 3.75m, long 6.25m, temperature 20 DEG C, obtains in the quiet closed room of humidity 50%, impulse response length and filter tap number L=512.The background noise of experiment is white Gaussian noise, and signal to noise ratio is 30dB.And microphones near end signal in, add from the 15000th sampled point the voice signal that length is 20000 sampled points.

According to above experiment condition, carry out echo cancellor experiment by the inventive method and existing two kinds of methods.The concrete value of parameter of various method is as table 1.

The optimized parameter that each algorithm tested by table 1 is similar to value

Fig. 1 is the condition of sparse channel figure of the communication system that the quiet closed room of experiment is formed.Fig. 2 is near end signal (voice) and remote signaling (voice) figure in the bilateral words of experiment.

Fig. 3 is document 1, the normalization steady output rate curve of the experiment of document 2 and the present invention's (being denoted as CAPSSAF in Fig. 3).As shown in Figure 3, under two call scenarios, the present invention is faster than document 1,2 convergence rate, and steady-state error is less.

Claims (1)

1. an affine projection symbol subband convex combination adaptive echo cancellation method, its step is as follows:

A, remote signaling filtering

The remote signaling sampling transmitted by far-end obtains the centrifugal pump u (n) of the current time n of remote signaling, remote signaling is at current time n and the centrifugal pump u (n) in a front L-1 moment, u (n-1) ..., u (n-L+1), form sub-filter input vector U (n) of current time n, U (n)=[u (n), u (n-1) ..., u (n-L+1)] ^t; Wherein L=512, be filter tap number, subscript T represents transposition;

Again by current time n and sub-filter input vector U (n) in a front P-1 moment, U (n-1) ... U (n-P+1) combines, obtain sub-filter affine projection input vector G (n), G (n)=[U (n), U (n-1), ..., U (n-P+1)], wherein P represents affine projection exponent number, P=4,8,16;

Subsequently, the analysis filterbank that sub-filter affine projection input vector G (n) is eliminated in filter by subband convex combination adaptive echo is obtained i-th sub-filter affine projection input vector G _i(n)=[U _i(n), U _i(n-1) ..., U _i(n-P+1)], i is the sequence number that subband convex combination adaptive echo eliminates filter, i=1,2,3 ... N, N<32;

I-th sub-filter affine projection input vector G _in () obtains after eliminating filter filtering by subband convex combination adaptive echo respectively: the output vector Y of the large step-length filter of the i-th subband of current time _{i, 1}(n), Y _{i, 1}(n)=W ₁(n) ^tg _i(n)=[y _{i, 1}(n), y _{i, 1}(n-1) ..., y _{i, 1}(n-P+1)]; The output vector Y of the little step-length filter of the i-th subband _{i, 2}(n), Y _{i, 2}(n)=W ₂(n) ^tg _i(n)=[y _{i, 2}(n), y _{i, 2}(n-1) ..., y _{i, 2}(n-P+1)]; Wherein, W ₁(n) and W ₂n convex combination adaptive echo that () is respectively current time n eliminates the tap weights vector of large step-length filter in filter each subband and little step-length filter, and initial value is zero;

By the output vector Y of the large step-length filter of the i-th subband _{i, 1}the output vector Y of (n) and the little step-length filter of the i-th subband _{i, 2}n () obtains the output vector Y of large step-length filter after eliminating the synthesis filter banks in filter respectively by subband convex combination adaptive echo ₁(n), Y ₁(n)=W ₁(n) ^tthe output vector Y of the little step-length filter of G (n) and current time ₂(n), Y ₂(n)=W ₂(n) ^tg (n);

B, convex combination

By the output vector Y of the large step-length filter of current time ₁the output vector Y of the little step-length filter of (n) and current time ₂n () carries out the output vector Y (n) that convex combination obtains the junction filter of current time, Y (n)=λ (n) Y ₁(n)+[1-λ (n)] Y ₂(n); Wherein, λ (n) is the weight of the large step-length filter of current time, and its expression formula is a (n) is hybrid parameter, and its initial value is 0;

By the tap weights of the large step-length filter of current time vector W ₁the tap weights vector W of the little step-length filter of (n) and current time ₂n () also carries out tap weights vector W (n) that convex combination obtains the junction filter of current time, W (n)=λ (n) W ₁(n)+[1-λ (n)] W ₂(n);

C, echo cancelltion

Near end signal centrifugal pump d (n) of the band echo in the current time n that near-end microphone pickup is arrived and a front P-1 moment, d (n-1), ..., d (n-P+1) combines, obtain affine projection near-end vector D (n) of current time, D (n)=[d (n), d (n-1), ..., d (n-P+1)]; Net signal E (n) of the current time after echo cancellor is obtained after being subtracted each other with the output vector Y (n) of the junction filter of current time by the affine projection near-end of current time vector D (n), E (n)=D (n)-Y (n)=[e (n), e (n-1), ..., e (n-P+1)], by the net signal loopback of current time to far-end;

D, filter tap weight vector upgrade

By the affine projection near-end of current time vector D (n), respectively with the output vector Y of the large step-length filter of current time ₁the output vector Y of the little step-length filter of (n), current time ₂n () subtracts each other, obtain the large step-length net signal E of current time ₁the little step-length net signal E of (n) and current time ₂(n), that is:

E ₁(n)＝D(n)-Y ₁(n)，E ₂(n)＝D(n)-Y ₂(n)；

Exploitation right vector transition strategy calculates the tap weights vector W of large step-length filter in the adaptive echo elimination filter of subsequent time n+1 ₁(n+1), $W_1(n+1)=W_1\left(n\right)+{\mathrm{\&mu;}}_1\frac{G\left(n\right)\mathrm{sgn}\&lsqb;E_1\left(n\right)\&rsqb;}{\left|\right|G\left(n\right)\mathrm{sgn}\&lsqb;E_1\left(n\right)\&rsqb;|{|}_2+\mathrm{\&delta;}};$ Wherein, μ ₁for the step-length of large step-length filter, its value is 0.05 ~ 0.1; δ is regularization parameter, and its value is 0.001 ~ 0.01; || || ₂represent 2 norms; Sgn is symbolic operation function, and when being positive number by operation values, its value is 1; When being 0 by operation values, its value is 0, and when being negative by operation values, its value is-1;

Exploitation right vector transition strategy calculates the tap weights vector W of the medium and small step-length filter of adaptive echo elimination filter of subsequent time n+1 simultaneously ₂(n+1);

As hybrid parameter a (n)>=a ⁺time: $W_2(n+1)=W\left(n\right)+{\mathrm{\&mu;}}_2\frac{G\left(n\right)\mathrm{sgn}\&lsqb;E_2\left(n\right)\&rsqb;}{\left|\right|G\left(n\right)\mathrm{sgn}\&lsqb;E_2\left(n\right)\&rsqb;|{|}_2+\mathrm{\&delta;}}$

As hybrid parameter a (n) < a ⁺time, $W_2(n+1)=W_2\left(n\right)+{\mathrm{\&mu;}}_2\frac{G\left(n\right)\mathrm{sgn}\&lsqb;E_2\left(n\right)\&rsqb;}{\left|\right|G\left(n\right)\mathrm{sgn}\&lsqb;E_2\left(n\right)\&rsqb;|{|}_2+\mathrm{\&delta;}}$

Wherein: a ⁺be a normal number, its value is 4 ~ 5; μ ₂for the step-length of little step-length filter, its value is 0.005 ~ 0.01;

The weight of E, filter upgrades

Formula after hybrid parameter a (n) is simplified by sign function upgrades:

$a(n+1)=a\left(n\right)+{\mathrm{\&mu;}}_a\mathrm{\&lambda;}\left(n\right)\&lsqb;1-\mathrm{\&lambda;}\left(n\right)\&rsqb;\underset{i=1}{\overset{N}{\&Sigma;}}e\left(n\right)\&lsqb;y_{i,1}\left(n\right)-y_{i,2}\left(n\right)\&rsqb;$

Wherein N represents the number of subband, μ _aa constant, first element of value to be 0.1, e (n) be current time net signal E (n), y _{i, 1}n () is the output vector Y of the large step-length filter of current time i-th subband _{i, 1}first element in (n), y _{i, 2}n () is the output vector Y of the little step-length filter of current time i-th subband _{i, 2}first element in (n);

Hybrid parameter a (n+1) after renewal is substituted into the weight expression formula of step B in, obtain the large step-length filter weight λ (n+1) of subsequent time n+1,

F, iteration

Make n=n+1, repeat the step of A, B, C, D, E, until end of conversation.