CN105491256A - Robust step size adjustment method in initialization phase of acoustic echo cancellation - Google Patents

Abstract

The invention discloses a steady step size adjustment method in an initialization phase of an acoustic echo cancellation. The method comprises: prior filtering: removing the data farthest away from n moment in a far-end signal vector, updating a current far-end signal to the far-end signal vector; evaluating a current echo signal according to the former state of a filter, filtering the evaluated current echo signal from a current near-end signal to obtain a prior error signal; determining the initialization step size of the filter; cancelling echo: cancelling the acoustic echo by DOUBLE-TALKROBUST VariableStepSizeNormalizedLeast-mean-squareForUnder-modeling according to the selected step size. According to the invention, the defect that the system is liable to maladjusted in the initialization phase of the VSS-NLMS-UMDT algorithm is solved, the trade-off between the convergence rate of the filter and the performance of the system is realized; and the stability and practicability of the system are enhanced.

Description

Step length regulating method sane startup stage of a kind of acoustic echo canceller

Technical field

The present invention relates to a kind of acoustic echo canceller.Step length regulating method sane startup stage of particularly relating to a kind of acoustic echo canceller.

Background technology

Along with the development of the communication technology, the requirement of people to convenient exchange way is more and more higher, and various hand-held telecommunication apparatus, video conference and the application of VoIP soft phone are more and more wider.In this kind of communication terminal, near-end speech passes to far-end by loud speaker, due to the coupling between far-end microphone and loud speaker, makes near-end speech pass this locality back, forms acoustic echo.Acoustic echo has a strong impact on quality of voice transmission, and therefore acoustic echo canceler system is indispensable.Acoustic echo canceller (AcousticEchoCancellation, AEC) be one of best solution solving acoustic echo elimination, general acoustic echo canceller at least comprises double talk detection module (DoubleTalkDetector, DTD), linear echo eliminates two parts.Fig. 1 is that a typical acoustic echo eliminates system.

The signal that microphone receives at moment n is:

d(n)＝y(n)+v(n)+w(n)1.1

In formula: d (n) represents near end signal, y (n) represents the echo that remote signaling x (n) produces after loud speaker is play, and v (n) represents near-end voice signals, and w (n) represents near-end noise signal.Remote signaling x (n) forms y (n) after ssystem transfer function h filtering

y(n)＝h ^T*x1.2

Wherein,

h ^T＝[h ₀(n)h ₁(n),...,h _N(n)] ^T

x＝[x(n)x(n-1),...,x(n-N+1)] ^T

X is remote signaling vector, and N is room impulse response length, and T represents transpose of a matrix.

The target of echo cancellor is exactly the echo path that design adaptive finite impulse response filter estimates between microphone and loud speaker then echo estimated signal is obtained according to this estimated path it is eliminated from d (n), and retains v (n).

$\hat{y}\left(n\right)=\hat{h}*x---1.3$

$e\left(n\right)=d\left(n\right)-\hat{y}\left(n\right)---1.4$

E (n) represents the error signal obtained after linear echo filter is eliminated, wherein,

$\hat{h}\left(n\right)={\[{\hat{h}}_0\left(n\right){\hat{h}}_1\left(n\right),...,{\hat{h}}_{L-1}\left(n\right)\]}^T---1.5$

L is sef-adapting filter length, general L<N in reality.It is generally acknowledged, the scene of acoustic echo canceller process is divided into three kinds of situations: far-end situation, only exists echo signal and there is not near-end voice signals; , there is not echo in near-end situation, only there is near-end voice signals; Dual end communication situation, echo signal and near-end voice signals exist simultaneously.

Owe variable step Normalized least mean squares (the DOUBLE-TALKROBUSTVariableStepSizeNormalizedLeast-mean-sq uareForUnder-modeling of cover half type dual end communication robust, VSS-NLMS-UMDT) be a kind of echo cancellation algorithm of novel more practical dual end communication robust, double talk detection device (DTD) is not needed compared with other Normalized least mean squares (VSS-NLMS) algorithm, just steady operation in fixed and dual end communication situation can owed, insensitive near end signal interference, still keep less and stable steady output rate, and it is easy to implement with control in actual applications, do not need any parameter of acoustic enviroment, robustness is very strong.Its step size proposed and filter update equation are

$\mathrm{\&mu;}\left(n\right)=\frac{1}{\mathrm{\&delta;}+x_L^T\left(n\right)*x_L\left(n\right)}\left|1-\frac{\sqrt{{\widehat{\mathrm{\&delta;}}}_d^2-{\widehat{\mathrm{\&delta;}}}_{\hat{y}}^2}}{\mathrm{\&xi;}+{\widehat{\mathrm{\&delta;}}}_e^2}\right|\frac{\mathrm{\&eta;}\left(n\right)}{\mathrm{\&eta;}\left(n\right)+{\mathrm{\&gamma;}}_{ed}\left(n\right)}---1.6$

$\hat{h}\left(n\right)=\hat{h}(n-1)+\mathrm{\&mu;}\left(n\right)x\left(n\right)e\left(n\right)---1.7$

In formula, the step-length that μ (n) is sef-adapting filter, γ _edbe the crosscorrelation estimation between e (n) and d (n), η (n) is the convergence statistical parameter of filter, with represent respectively d (n), with the energy expectation estimation of e (n), δ, ξ are a constant.Above parameter can be obtained by formula 1.8 and 1.9

γ _ed(n)＝E[e(n)*d(n)]＝λγ _ed(n-1)+(1-λ)e(n)d(n)1.8

$\mathrm{\&eta;}\left(n\right)=\left|\frac{{\mathrm{\&gamma;}}_{ed}\left(n\right)-{\widehat{\mathrm{\&delta;}}}_e^2\left(n\right)}{{\widehat{\mathrm{\&delta;}}}_d^2\left(n\right)-{\mathrm{\&gamma;}}_{ed}\left(n\right)}\right|---1.9$

E{} represents mathematic expectaion, and λ is a minimum normal number, mark represent that the energy of sequence p (n) is estimated, it can be calculated by index recurrence formula, and computational methods are

${\widehat{\mathrm{\&delta;}}}_p^2\left(n\right)=\mathrm{\&lambda;}{\widehat{\mathrm{\&delta;}}}_p^2(n-1)+(1-\mathrm{\&lambda;})p^2\left(n\right)---1.10$

Although this algorithm is with the obvious advantage, but still there are some drawbacks.This algorithm, at system start-up phase, needs by larger step-length upgrade filter, make it rapidly converge to a certain stable state.If system start-up phase is in dual end communication situation, now near-end voice filtering can fall by filter, and this is not allowed by system, greatly reduces systematic function; If system start-up phase adopts little step size $\mathrm{\&mu;}\left(n\right)=\frac{1}{\mathrm{\&delta;}+x_L^T\left(n\right)*x_L\left(n\right)}\left|1-\frac{\sqrt{{\widehat{\mathrm{\&delta;}}}_d^2-{\widehat{\mathrm{\&delta;}}}_{\hat{y}}^2}}{\mathrm{\&xi;}+{\widehat{\mathrm{\&delta;}}}_e^2}\right|\frac{\mathrm{\&eta;}\left(n\right)}{\mathrm{\&eta;}\left(n\right)+{\mathrm{\&gamma;}}_{ed}\left(n\right)}$ Adjustment, filter converges speed will be slack-off.Because this algorithm does not need DTD module, so system self cannot judge to be in near-end or to be far-end situation, so suitable step-length initialization filter can not be selected.

Summary of the invention

Technical problem to be solved by this invention is, there is provided a kind of and can not only ensure that filter rapidly converges to a certain stable state in dual end communication situation, can also startup stage complete reservation near-end speech, improve systematic function, step length regulating method sane startup stage of practicality stronger acoustic echo canceller.

The technical solution adopted in the present invention is: step length regulating method sane startup stage of a kind of acoustic echo canceller, comprises the steps:

1) algorithm parameter is determined;

2) priori filtering, described priori filtering, comprising: remove remote signaling vector x middle distance n moment data x farthest (n-L), current remote signaling x (n) is updated to remote signaling vector x; Utilize the previous state of filter estimate current echo signal, and the current echo signal filtering from current near end signal d (n) that will estimate, obtain prior uncertainty signal epsilon (n);

3) filter initialization step-length is determined;

4) carry out echo cancellor, according to step 3) selected by step-length, utilize and owe the variable step Normalized least mean squares of cover half type dual end communication robust and carry out acoustic echo elimination.

Step 1) described in algorithm parameter include: comprise speech sample frequency f s, filter length L, filter status filter step size μ (n), filter step size maximum μm ax, system start-up time initTime, remote signaling vector x, prior uncertainty signal epsilon (n), posteriori error signal e (n), the energy expectation estimation of near end signal the energy expectation estimation of remote signaling the energy expectation estimation of estimated echo signal with the energy expectation estimation of error signal crosscorrelation estimation γ between prior uncertainty signal and near end signal _ed, convergence statistical parameter η (n), convergence statistical parameter desired value exp η (n), convergence statistical parameter expects threshold value thres.

Step 3) described in determination filter initialization step-length comprise the steps:

(1) the crosscorrelation estimation γ between prior uncertainty signal and near end signal is calculated _ed, the energy expectation estimation of near end signal the energy expectation estimation of estimated echo signal with the energy expectation estimation of error signal

(2) by the parameter in step (1), substitute into convergence statistical parameter η (n) computing formula, obtain convergence statistical parameter η (n),

$\mathrm{\&eta;}\left(n\right)=\left|\frac{{\mathrm{\&gamma;}}_{ed}\left(n\right)-{\widehat{\mathrm{\&delta;}}}_e^2\left(n\right)}{{\widehat{\mathrm{\&delta;}}}_d^2\left(n\right)-{\mathrm{\&gamma;}}_{ed}\left(n\right)}\right|---1.9$

(3) define convergence statistical parameter desired value exp η (n), computing formula is

expη(n)＝λ*expη(n-1)+(1-λ)*η(n)1.11

In system start-up time initTime, if exp η (n) is less than convergence statistical parameter expect threshold value thres, then thinks that system is current and be in far-end situation, take filter step size maximum μm ax to upgrade filter; Otherwise step-length is upgraded.

Described in step (3) to step-length carry out renewal be adopt following formula:

$\mathrm{\&mu;}\left(n\right)=\frac{1}{\mathrm{\&delta;}+x_L^T\left(n\right)*x_L\left(n\right)}\left|1-\frac{\sqrt{{\widehat{\mathrm{\&delta;}}}_d^2-{\widehat{\mathrm{\&delta;}}}_{\hat{y}}^2}}{\mathrm{\&xi;}+{\widehat{\mathrm{\&delta;}}}_e^2}\right|\frac{\mathrm{\&eta;}\left(n\right)}{\mathrm{\&eta;}\left(n\right)+{\mathrm{\&gamma;}}_{ed}\left(n\right)}---1.6$

In formula, δ, ξ are a constant.

Step length regulating method sane startup stage of a kind of acoustic echo canceller of the present invention, does not need DTD module, between filter converges speed and systematic function, obtain compromise.Suitable step-length adjustment mode is selected the acoustic echo canceller startup stage by convergence statistical parameter desired value, realize in far-end situation, in the most of the time, select large step-length to upgrade filter, make its Fast Convergent, and keep little step-length to upgrade in other cases, can near-end speech be retained, also can maintain filter simultaneously and upgrade fast.In dual end communication situation, method of the present invention can not only ensure that filter rapidly converges to a certain stable state, can also startup stage complete reservation near-end speech, improve systematic function, practicality is stronger.The invention solves VSS-NLMS-UMDT algorithm startup stage the system drawback of easily lacking of proper care, between filter converges speed and systematic function, obtain compromise, enhance stability and the practicality of system.

Accompanying drawing explanation

Fig. 1 is that a typical acoustic echo eliminates system configuration;

Fig. 2 a is far-end speech signal time-domain diagram;

Fig. 2 b is near end signal time-domain diagram;

Fig. 2 c is near-end voice signals time-domain diagram;

Fig. 3 a is VSS-NLMS result;

Fig. 3 b is VSS-NLMS-UMDT result;

Fig. 3 c is VSS-NLMS-UMDT-CSE result;

Fig. 4 is the algorithm flow chart of the embodiment of the present invention.

Embodiment

Below in conjunction with embodiment and accompanying drawing, step length regulating method sane startup stage of a kind of acoustic echo canceller of the present invention is described in detail.

How step length regulating method sane startup stage of a kind of acoustic echo canceller of the present invention, select suitable step length regulating method startup stage that key to be solved being acoustic echo canceller.If startup stage in far-end situation, choose large step-length upgrade filter, make filter Fast Convergent; Otherwise choose little step-length, the renewal that near-end voice signals also can maintain filter simultaneously can be retained.

Step length regulating method sane startup stage of a kind of acoustic echo canceller of the present invention, comprises the steps:

1) algorithm parameter is determined;

Described algorithm parameter includes: comprise speech sample frequency f s, filter length L, filter status filter step size μ (n), filter step size maximum μm ax, system start-up time initTime, remote signaling vector x, prior uncertainty signal epsilon (n), posteriori error signal e (n), the energy expectation estimation of near end signal the energy expectation estimation of remote signaling the energy expectation estimation of estimated echo signal with the energy expectation estimation of error signal crosscorrelation estimation γ between prior uncertainty signal and near end signal _ed, convergence statistical parameter η (n), convergence statistical parameter desired value exp η (n), convergence statistical parameter expects threshold value thres.

2) priori filtering, described priori filtering, comprising: remove remote signaling vector x middle distance n moment data x farthest (n-L), current remote signaling x (n) is updated to remote signaling vector x; Utilize the previous state of filter estimate current echo signal, and the current echo signal filtering from current near end signal d (n) that will estimate, obtain prior uncertainty signal epsilon (n);

3) filter initialization step-length is determined;

Described determination filter initialization step-length comprises the steps:

(1) the crosscorrelation estimation γ between prior uncertainty signal and near end signal is calculated _ed, the energy expectation estimation of near end signal the energy expectation estimation of estimated echo signal with the energy expectation estimation of error signal

(2) by the parameter in step (1), substitute into convergence statistical parameter η (n) computing formula, obtain convergence statistical parameter η (n),

$\mathrm{\&eta;}\left(n\right)=\left|\frac{{\mathrm{\&gamma;}}_{ed}\left(n\right)-{\widehat{\mathrm{\&delta;}}}_e^2\left(n\right)}{{\widehat{\mathrm{\&delta;}}}_d^2\left(n\right)-{\mathrm{\&gamma;}}_{ed}\left(n\right)}\right|---1.9$

(3) convergence statistical parameter η (n) only can represent the current state of filter, and jumping characteristic is comparatively large, can not follow the tracks of description filter well.In order to the convergence of statistical zero-knowledge better, therefore define convergence statistical parameter desired value exp η (n), computing formula is

expη(n)＝λ*expη(n-1)+(1-λ)*η(n)1.11

Experimental studies have found that, upgrade if filter is stable in far-end situation, then convergence statistical parameter expects that exp η (n) most of the time will be in a certain below smaller value thres.Therefore in system start-up time initTime, if exp η (n) is less than convergence statistical parameter expect threshold value thres, then thinks that system is current and be in far-end situation, take filter step size maximum μm ax to upgrade filter; Otherwise step-length is upgraded.It is described that to carry out renewal to step-length be adopt following formula:

$\mathrm{\&mu;}\left(n\right)=\frac{1}{\mathrm{\&delta;}+x_L^T\left(n\right)*x_L\left(n\right)}\left|1-\frac{\sqrt{{\widehat{\mathrm{\&delta;}}}_d^2-{\widehat{\mathrm{\&delta;}}}_{\hat{y}}^2}}{\mathrm{\&xi;}+{\widehat{\mathrm{\&delta;}}}_e^2}\right|\frac{\mathrm{\&eta;}\left(n\right)}{\mathrm{\&eta;}\left(n\right)+{\mathrm{\&gamma;}}_{ed}\left(n\right)}---1.6$

In formula, δ, ξ are a constant.

4) carry out echo cancellor, according to step 3) selected by step-length, utilize and owe the variable step Normalized least mean squares of cover half type dual end communication robust and carry out acoustic echo elimination.

Below, for speech sample rate be 16K, the filter length system that is 1000 rank implements this patent and suggested plans, treatment step is with reference to figure 4 algorithm flow chart.

Algorithm parameter is arranged: speech sample frequency f s=16K, filter length L=1000, filter initial condition filter initial step length μ (n)=0, constant filter step size maximum with filter initial time initTime=6s, remote signaling vector x=0, prior uncertainty signal epsilon (n)=0, posteriori error signal e (n)=0, the energy expectation estimation of near end signal, estimated echo signal and error signal crosscorrelation estimation γ between prior uncertainty signal and near end signal _ed=0, convergence statistical parameter η (n)=0, convergence statistical parameter desired value exp η (n)=0, convergence statistical parameter expects threshold value thres=0.05.

The concrete implementation step of this algorithm is as follows:

1, read current remote signaling x (n) and current near end signal d (n), x (n) is updated to remote signaling vector x _lin (n);

2, by formula 1.3 and 1.4, prior uncertainty signal epsilon (n) is calculated;

3, by formula 1.8-1.10, characteristic value γ is calculated _ed, η (n) and exp η (n);

If 4 current time n<initTime*L, the startup stage that then filter being in, if now convergence statistical parameter desired value exp η (n) is less than convergence statistical parameter expectation threshold value thres, then make μ (n)=μm ax; Otherwise, $\mathrm{\&mu;}\left(n\right)=\frac{1}{\mathrm{\&delta;}+x_L^T\left(n\right)*x_L\left(n\right)}\left|1-\frac{\sqrt{{\widehat{\mathrm{\&delta;}}}_d^2-{\widehat{\mathrm{\&delta;}}}_{\hat{y}}^2}}{\mathrm{\&xi;}+{\widehat{\mathrm{\&delta;}}}_e^2}\right|\frac{\mathrm{\&eta;}\left(n\right)}{\mathrm{\&eta;}\left(n\right)+{\mathrm{\&gamma;}}_{ed}\left(n\right)};$

If 1 n>=initTime*L, then system starts end, order

$\mathrm{\&mu;}\left(n\right)=\frac{1}{\mathrm{\&delta;}+x_L^T\left(n\right)*x_L\left(n\right)}\left|1-\frac{\sqrt{{\widehat{\mathrm{\&delta;}}}_d^2-{\widehat{\mathrm{\&delta;}}}_{\hat{y}}^2}}{\mathrm{\&xi;}+{\widehat{\mathrm{\&delta;}}}_e^2}\right|\frac{\mathrm{\&eta;}\left(n\right)}{\mathrm{\&eta;}\left(n\right)+{\mathrm{\&gamma;}}_{ed}\left(n\right)};$

5, the step-length obtained is substituted into formula 1.6 and obtain filter status posteriori error signal e (n) is obtained after substitution formula 1.4 and 1.3.If read voice to be all disposed, algorithm stops; Otherwise the 1st step is returned in redirect.

Posteriori error signal e (n) is the final output of system.

Claims (4)

1. a step length regulating method sane startup stage of acoustic echo canceller, is characterized in that, comprises the steps:

1) algorithm parameter is determined;

2) priori filtering, described priori filtering, comprising: remove remote signaling vector x middle distance n moment data x farthest (n-L), current remote signaling x (n) is updated to remote signaling vector x; Utilize the previous state of filter estimate current echo signal, and the current echo signal filtering from current near end signal d (n) that will estimate, obtain prior uncertainty signal epsilon (n);

3) filter initialization step-length is determined;

4) carry out echo cancellor, according to step 3) selected by step-length, utilize and owe the variable step Normalized least mean squares of cover half type dual end communication robust and carry out acoustic echo elimination.

2. step length regulating method sane startup stage of a kind of acoustic echo canceller according to claim 1, is characterized in that, step 1) described in algorithm parameter include: comprise speech sample frequency f s, filter length L, filter status filter step size μ (n), filter step size maximum μm ax, system start-up time initTime, remote signaling vector x, prior uncertainty signal epsilon (n), posteriori error signal e (n), the energy expectation estimation of near end signal the energy expectation estimation of remote signaling the energy expectation estimation of estimated echo signal with the energy expectation estimation of error signal crosscorrelation estimation γ between prior uncertainty signal and near end signal _ed, convergence statistical parameter η (n), convergence statistical parameter desired value exp η (n), convergence statistical parameter expects threshold value thres.

3. step length regulating method sane startup stage of a kind of acoustic echo canceller according to claim 1, is characterized in that, step 3) described in determination filter initialization step-length comprise the steps:

(1) the crosscorrelation estimation γ between prior uncertainty signal and near end signal is calculated _ed, the energy expectation estimation of near end signal the energy expectation estimation of estimated echo signal with the energy expectation estimation of error signal

(2) by the parameter in step (1), substitute into convergence statistical parameter η (n) computing formula, obtain convergence statistical parameter η (n),

$\mathrm{\&eta;}\left(n\right)=\left|\frac{{\mathrm{\&gamma;}}_{ed}\left(n\right)-{\widehat{\mathrm{\&delta;}}}_e^2\left(n\right)}{{\widehat{\mathrm{\&delta;}}}_d^2\left(n\right)-{\mathrm{\&gamma;}}_{ed}\left(n\right)}\right|---1.9$

(3) define convergence statistical parameter desired value exp η (n), computing formula is

expη(n)＝λ*expη(n-1)+(1-λ)*η(n)1.11

In system start-up time initTime, if exp η (n) is less than convergence statistical parameter expect threshold value thres, then thinks that system is current and be in far-end situation, take filter step size maximum μm ax to upgrade filter; Otherwise step-length is upgraded.

4. step length regulating method sane startup stage of a kind of acoustic echo canceller according to claim 3, is characterized in that, described in step (3) to step-length carry out renewal be adopt following formula:

$\mathrm{\&mu;}\left(n\right)=\frac{1}{\mathrm{\&delta;}+x_L^T\left(n\right)*x_L\left(n\right)}|1-\frac{\sqrt{{\widehat{\mathrm{\&delta;}}}_d^2-{\widehat{\mathrm{\&delta;}}}_{\hat{y}}^2}}{\mathrm{\&xi;}+{\widehat{\mathrm{\&delta;}}}_e^2}|\frac{\mathrm{\&eta;}\left(n\right)}{\mathrm{\&eta;}\left(n\right)+{\mathrm{\&gamma;}}_{ed}\left(n\right)}---1.6$

In formula, δ, ξ are a constant.