CN103369162A - Low complexity phone echo self-adaption eliminating method - Google Patents

Abstract

The invention discloses a low complexity phone echo self-adaption eliminating method. The method mainly comprises the steps of: A, far-end signal filtration: obtaining a large-step filtering value y1(n) and a small-step filtering value y2(n); B, convex combination: completing the convex combination of the large-step filtering value y1(n) and the small-step filtering value y2(n) so as to obtain a combined filtering value y(n), wherein y(n)=lambda(n)y1(n)+(1-lambda(n))y2(n); C, echo cancellation: completing the subtraction of a near-end signal d(n) which has echo and is picked by a near-end microphone and the combined filtering value y(n), and then sending the signal back to a far end, wherein a return signal is a total residual signal e(n), and e(n)=d(n)-y(n); D, updating the weight coefficient of filtering device taps; E, updating the weight of the filtering device, i.e. updating a hybrid parameter a(n) through a formula which is simplified by a sign function; F, defining the weight of the filtering device; and G, setting n=n+1, and repeating the steps of A, B, C, D, E and F until the conversation is over. The method has strong identification capability to sparse telephone communication systems, and particularly has quick convergence rate and small steady state error in a transition period; an echo eliminating effect is good; and moreover, the method is low in calculation complexity, low in hardware cost and easy to implement.

Description

A kind of listener's echo self adaptive elimination method of low complex degree

Technical field

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

Background technology

Along with recent advances in technology, the communication technology is growing, and various emerging communications emerge in an endless stream, but topmost communication mode remains voice communication, and the user also more and more is concerned about voice communication quality.Acoustic echo is the main factor that affects voice call quality.Acoustic echo in the communication process refers to that the user hears the sound of oneself repeatedly in communication process, is divided into direct echo and indirect echo, and directly echo refers to that the sound that is sent by loud speaker directly enters microphone without any reflection.The time-delay of this echo is the shortest, speech energy with remote speaker, distance between loud speaker and the microphone, angle, the broadcast sound volume of loud speaker, the factor analysis such as microphone pick-up sensitivity, and enter the set of the echo that microphone produces after the one or many reflection of sound through different paths (such as any object in house or the house) that indirectly echo refers to be broadcasted by loud speaker, the any change of any object in the house all can change the passage of echo, therefore, this echo is to become in multipath, time.The reason that produces acoustic echo is that the sound that loud speaker sends is also picked up and pass back the talker by multiple path by microphone when being heard by recipient owing to having certain space length between the loud speaker, microphone.And people's ear is for echo and sensitivity thereof, and the echo that postpones 10ms in communication process can both be caught by people's ear and perceive, and the echo that surpasses 32ms will be caused great interference to communication.Take hand-free telephone system as example, this system should be equipped with acoustic echo canceller (Acoustic Echo Cancellation), is called for short AEC.The loud speaker of hands-free phone and the space length between the microphone cause can entering microphone and produce acoustic echo through the multipath reflection of room interior from a part of signal of loud speaker, the path that this part signal transmits is exactly the acoustic echo channel, and available rooms impulse response is described.The echo signal of the echo path that the echo elimination will pick out exactly and the signal subtraction in the channel are to balance out echo.Therefore, the effect of echo elimination depends primarily on the performance of echo identification.Because echo disturb to exist postpones, the time characteristic such as change, multipath, these have all caused very large difficulty for the echo identification, conventional method is difficult to obtain good effect.Sef-adapting filter can utilize adaptive algorithm to change filter parameter and structure according to the change of environment, adjusts the sef-adapting filter parameter according to the replacement criteria of its regulation, just can better pick out echo.Wherein, LMS(lowest mean square) filter.It is simple in structure, be easy to realize, has obtained domestic and international researcher's extensive concern.System Discrimination principle based on the LMS filter is: utilize steepest gradient to seek filter tap weight vector w (n), so that filter cost function J (n)=| e (n) | ²Minimum, the input signal of system is x (n), d (n) is reference signal, y (n) is the actual output of LMS filter, and is minimum with the error of y (n) by making d (n), so that the output of adaptive filter algorithm approaches the output of unknown system as far as possible, after the filter convergence, think that namely unknown system obtained identification when both transmission characteristic was basically identical, echo is eliminated.System Discrimination problem formula based on the LMS algorithm is as follows:

Filter output: y (n)=w (n) ^HX (n) (1)

Output error: e (n)=d (n)-y (n) (2)

Weight coefficient upgrades: w (n+1)=w (n)+μ e (n) x (n) (3)

Wherein: 1) w (n) is n filter tap weight vector constantly.2) μ is known as step factor, and it is directly related with final performance, and the excessive fast convergence rate steady-state error of μ is large; The slow steady-state error of the too small convergence rate of μ is little.

Although the LMS filtering algorithm has simple distinguishing feature, there is the intrinsic contradictions of convergence rate and steady-state error, in order effectively to address this problem, adopting two LMS filters to carry out convex combination is a kind of good selection.Simultaneously carry out identification for sparse unknown system, add proportional based on l ₁The IPNLMS algorithm of norm is further accelerated the convergence rate of convex combination LMS filter.So-called Sparse System just refers to that the most of coefficient of impulse response is 0 system, and length can reach hundreds of symbols, only has efficiency factor (nonzero coefficient) seldom.Because general system identifying method does not have the constraint for Sparse System, so these algorithms do not have special advantage over against the Sparse System identification time.Add proportional based on l ₁The IPNLMS algorithm of norm can improve the convergence tracking performance in the convex combination filter, reduce steady-state error.In the application of present Sparse System identification, ripe method has following three kinds:

(1) based on the Sparse System discrimination method of proportional normalization minimum mean-square (PNLMS) algorithm

List of references 1 " Proportionate Normalized Least-Mean-Squares adaptation in echo cancellers " (D.L.Duttweiler, IEEE Transactions on Speech and Audio Processing, vol.8, no.5, pp.508 – 518, Sep.2000.) the method is that normalization minimum mean-square (NLMS) algorithm is added proportional factor, in this algorithm, need not priori, each sef-adapting filter is mixed own suitable adaptive step parameter simultaneously, step-length is obtained by the parameter Estimation of last time, larger parameter can obtain larger step-length, thereby accelerated convergence rate, this algorithm is called the PNLMS algorithm.The method need not prior uncertainty, and the active regions of room echo channel also need not accurate location, but this algorithm is to adopt a maximum value between current estimated value and the fixed constant upgrading Step matrix, seems too dogmatic.

(2) improved proportional normalization minimum mean-square (IPNLMS)

List of references 2 " An improved PNLMS Algorithm " (J.Benesty, IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), vol.2, pp.1881 – 1884, May.2002) this algorithm is called again the IPNLMS algorithm, this algorithm is derived the proportional factor by the l1 norm, generally be better than the PNLMS algorithm for sparse situation, initial convergence speed is compared than the PNLMS algorithm and is improved.But still there are the intrinsic contradictions of sef-adapting filter convergence rate and steady-state error in this algorithm.

(3) the convex combination adaptive echo is eliminated filtered method

List of references 3 " Adaptive Combination of Proportionate Filters for Sparse Echo Cancellation " (J.Arenas-Garcia, A.R.Figueiras-Vidal, IEEE Transactions on Speech and Audio Processing, vol.17, no.6, pp, 1087 – 1098, Aug.2009), the method is that the convex combination adaptive echo is eliminated filter (CIPNLMS) method, by the IPNLMS algorithm being added in the convex combination LMS algorithm, the original convergence rate that not only keeps, also utilize the characteristic of convex combination itself, solved the contradiction between convergence rate and the problem error, reduced steady-state error.But this algorithm is because the internal mix parameter is many, and transfer complex has significantly increased computation complexity, and parameter is numerous, and is higher to the required precision of parameter testing.

Summary of the invention

Goal of the invention of the present invention just provides a kind of listener's echo self adaptive elimination method of low complex degree, and the method is strong to the identification capability of this Sparse System of telephone communication, the fast convergence rate of transition stage especially, and steady-state error is little; The echo eradicating efficacy is good; Computation complexity is low simultaneously, and required hardware cost is low, easily implements.

The present invention realizes that the technical scheme that its goal of the invention adopts is, a kind of listener's echo self adaptive elimination method of low complex degree, and its step is as follows:

A, remote signaling filtering

The remote signaling sampling that far-end is transmitted obtains the centrifugal pump x (n) of the current time n of remote signaling, obtains respectively large step-length filter value y after remote signaling centrifugal pump x (n) is eliminated filter filtering by the convex combination adaptive echo ₁(n), y ₁(n)=w ₁(n) ^HThe long filter value y of x (n) and small step ₂(n), y ₂(n)=w ₂(n) ^HX (n); Wherein, w ₁(n) and w ₂(n) the convex combination adaptive echo that is respectively current time n is eliminated the large step-length filter in the filter, the tap weights coefficient of the long filter of small step, and its initial value is zero, and subscript H represents conjugate transpose;

B, convex combination

With large step-length filter value y ₁(n) and the long filter value y of small step ₂(n) carry out convex combination and obtain combined filter value y (n),

y(n)=λ(n)y ₁(n)+(1-λ(n))y ₂(n)

Wherein, λ (n) is the weight of large step-length filter, and its expression formula is Be hybrid parameter, its initial value is 0;

C, echo cancelltion

The near-end microphone picked up after subtracting each other with the near end signal d (n) of echo and combined filter value y (n) loopback is to far-end again, backhaul signals is total residual signals e (n), e (n)=d (n)-y (n);

D, filter tap weight coefficient upgrade

With near end signal d (n), respectively with large step-length filter value y ₁(n), the long filter value y of small step ₂(n) subtract each other, obtain large step-length residual signals e ₁(n) and the long residual signals e of small step ₂(n), that is:

e ₁(n)=d(n)-y ₁(n)，e ₂(n)=d(n)-y ₂(n)；

Use improved proportional normalization all square filtering method calculate next constantly the adaptive echo of n+1 eliminate filter tap weight coefficient w (n+1):

$w_1(n+1)=w_1\left(n\right)+\frac{{\mathrm{\μ}}_{1}X{\left(n\right)}^T{G}_1\left(n\right)e_1\left(n\right)}{X{\left(n\right)}^T{G}_1\left(n\right)X\left(n\right)+\mathrm{\δ}}$

w ₂(n+1)=w ₂(n)+μ ₂X(n) ^TG ₂(n)e ₂(n)

Wherein: μ ₁Eliminate the step-length of filter for large step-length echo, its value is 0.5～0.8; μ ₂Eliminate the step-length of filter for the long echo of small step, its value is 0.1～0.3; δ is regularization parameter, and its value is 0.001～0.01; The matrix that X (n) consists of to n-L+1 at n constantly for remote signaling x (n), X (n)=[x (n) ..., x (n-L+1)] ^TG ₁(n) and G ₂(n) be respectively large Step matrix and little Step matrix, calculated by following formula

$G_i\left(n\right)=\frac{1-{\mathrm{\κ}}_i}{2L}+(1+{\mathrm{\κ}}_i)\frac{\left|w_i\left(n\right)\right|}{2{\left|\right|w_i\left(n\right)\left|\right|}_1+{\mathrm{\ϵ}}_{\mathrm{ip}}},i=\mathrm{1,2}$

Wherein, || || ₁Expression 1-norm, κ _iProportionality control parameter κ _i∈ [1,1], i=1,2, ε _IpBe regularization parameter, its value is 0.001～0.01;

The weight of E, filter is upgraded

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

a(n+1)=a(n)+μ _asgn[e(n)(e ₂(n)-e ₁(n))]

Wherein sgn represents sign function

μ _aBe a constant, value is 0.01; In the weight expression formula with hybrid parameter a (n+1) the substitution step B after upgrading, obtain the updating value λ (n+1) of filter weight, $\mathrm{\λ}(n+1)=\frac{1}{1+e^{-a(n+1)}};$

The weight of F, filter limits

If λ (n+1)＜0.01 then makes λ (n+1)=0; If λ (n+1)〉0.99, and 4 μ ₁＜0.5, then make λ (n+1) if=0 λ (n+1) 0.99 and 4 μ ₁0.5, then make λ (n+1)=0.99

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

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

(1), the little and fast convergence rate of steady-state error

The combined filter value y (n) of bank of filters output is the estimated value of echo signal, the near-end microphone picked up with the near end signal d (n) of echo be from the signal of near-end loopback to far-end with subtracting each other, this signal is for eliminating the total residual signals e (n) after the echo.The convergence rate of total residual signals is faster, and its echo eradicating efficacy is better.The little steady-state error of Fast Convergent and the long filter of small step by large step-length filter has guaranteed that the whole convex combination adaptive echo after the two convex combination eliminates the little while fast convergence rate of steady-state error of filter.In convex combination filter internal mix parameter is transmitted, by hybrid parameter a (n) symbolization function is simplified, so that a (n) is to parameter μ _aWhen getting identical value, be subjected to the impact of error e (n) little, and then when making convex combination, weight is large in the early stage for large step-length filter, combined filter value y (n) mainly is large step-length filter value, later stage, large step-length filter weight was little, combined filter value y (n) mainly is the long filter value of small step, and the weight variation tendency is relatively stable, be subjected to error e (n) impact little, can more give full play to advantage, especially the transition stage fast convergence rate of the little steady-state error of the Fast Convergent of large step-length filter and the long filter of small step.Its echo is eliminated rapider, residual lower.

(2) computation complexity is low

With the transfer mode of the internal mix parameter a (n) of document 3 be: a (n+1)=a (n)+μ _aE (n) (y ₁(n)-y ₂(n)) λ (n) [1-λ (n)], e (n)=d (n)-y (n) needs 6 multiplication, and computation complexity is 6; The transfer mode of internal mix parameter a of the present invention (n) is: a (n+1)=a (n)+μ _aSgn[e (n) (e ₂(n)-e ₁(n))] only need 3 multiplication, computation complexity is 3, has significantly reduced computation complexity, and required hardware cost is low, easily implements.

The present invention is described in detail below in conjunction with the drawings and specific embodiments.

Description of drawings

Fig. 1 is near end signal d (n) figure.

Fig. 2 is the combined filter value y (n) that the present invention tests output.

Fig. 3 is the weight λ (n)-time graph of large step-length filter of the present invention.

Fig. 4 is the normalization stable state imbalance curve of document 1 and document 2.

Fig. 5 is document 3 and normalization stable state of the present invention imbalance curve.

Embodiment

Embodiment

A kind of listener's echo self adaptive elimination method of low complex degree, its step is as follows:

A, remote signaling filtering

The remote signaling sampling that far-end is transmitted obtains the centrifugal pump x (n) of the current time n of remote signaling, obtains respectively large step-length filter value y after remote signaling centrifugal pump x (n) is eliminated filter filtering by the convex combination adaptive echo ₁(n), y ₁(n)=w ₁(n) ^HThe long filter value y of x (n) and small step ₂(n), y ₂(n)=w ₂(n) ^HX (n); Wherein, w ₁(n) and w ₂(n) the convex combination adaptive echo that is respectively current time n is eliminated the large step-length filter in the filter, the tap weights coefficient of the long filter of small step, and its initial value is zero, and subscript H represents conjugate transpose;

B, convex combination

With large step-length filter value y ₁(n) and the long filter value y of small step ₂(n) carry out convex combination and obtain combined filter value y (n),

y(n)=λ(n)y ₁(n)+(1-λ(n))y ₂(n)

Wherein, λ (n) is the weight of large step-length filter, and its expression formula is

Be hybrid parameter, its initial value is 0;

C, echo cancelltion

The near-end microphone picked up after subtracting each other with the near end signal d (n) of echo and combined filter value y (n) loopback is to far-end again, backhaul signals is total residual signals e (n), e (n)=d (n)-y (n);

D, filter tap weight coefficient upgrade

With near end signal d (n), respectively with large step-length filter value y ₁(n), the long filter value y of small step ₂(n) subtract each other, obtain large step-length residual signals e ₁(n) and the long residual signals e of small step ₂(n), that is:

e ₁(n)=d(n)-y ₁(n)，e ₂(n)=d(n)-y ₂(n)；

Use improved proportional normalization all square filtering method calculate next constantly the adaptive echo of n+1 eliminate filter tap weight coefficient w (n+1):

$w_1(n+1)=w_1\left(n\right)+\frac{{\mathrm{\μ}}_{1}X{\left(n\right)}^T{G}_1\left(n\right)e_1\left(n\right)}{X{\left(n\right)}^T{G}_1\left(n\right)X\left(n\right)+\mathrm{\δ}}$

w ₂(n+1)=w ₂(n)+μ ₂X(n) ^TG ₂(n)e ₂(n)

Wherein: μ ₁Eliminate the step-length of filter for large step-length echo, its value is 0.5～0.8; μ ₂Eliminate the step-length of filter for the long echo of small step, its value is 0.1～0.3; δ is regularization parameter, and its value is 0.001～0.01; The matrix that X (n) consists of to n-L+1 at n constantly for remote signaling x (n), X (n)=[x (n) ..., x (n-L+1)] ^TG ₁(n) and G ₂(n) be respectively large Step matrix and little Step matrix, calculated by following formula

$G_i\left(n\right)=\frac{1-{\mathrm{\κ}}_i}{2L}+(1+{\mathrm{\κ}}_i)\frac{\left|w_i\left(n\right)\right|}{2{\left|\right|w_i\left(n\right)\left|\right|}_1+{\mathrm{\ϵ}}_{\mathrm{ip}}},i=\mathrm{1,2}$

Wherein, || || ₁Expression 1-norm, κ _iProportionality control parameter κ _i∈ [1,1], i=1,2, ε _IpBe regularization parameter, its value is 0.001～0.01;

The weight of E, filter is upgraded

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

a(n+1)=a(n)+μ _asgn[e(n)(e ₂(n)-e ₁(n))]

Wherein sgn represents sign function

μ _aBe a constant, value is 0.01; In the weight expression formula with hybrid parameter a (n+1) the substitution step B after upgrading, obtain the updating value λ (n+1) of filter weight, $\mathrm{\λ}(n+1)=\frac{1}{1+e^{-a(n+1)}};$

The weight of F, filter limits

If λ (n+1)＜0.01 then makes λ (n+1)=0; If λ (n+1)〉0.99, and 4 μ ₁＜0.5, then make λ (n+1) if=0 λ (n+1) 0.99 and 4 μ ₁0.5, then make λ (n+1)=0.99

G, make n=n+1, repeat the step of A, B, C, D, E, F, until end of conversation;

Emulation experiment

In order to verify validity of the present invention, carried out emulation experiment, and the contrast of carrying out with existing document 3 algorithms.

The remote signaling x (n) of emulation experiment is 0.5 single order autoregression (AR(1) for limit) colourful signal.Near end signal d (n): be high 2.5m in the room, wide 3.75m, 20 ℃ of long 6.25m temperature, near-end after loud speaker is play, is that the 8000Hz sampling order is 512 to pick up out altogether the near end signal d (n) of 60000 moment point with microphone by sample frequency with the remote signaling that receives in the room in the quiet closed room of humidity 50%.

Above-mentioned remote signaling x (n) and corresponding near end signal d (n) are carried out echo elimination with the inventive method and existing three kinds of algorithms.The concrete value of the parameter of the whole bag of tricks such as table 1.

Fig. 1 is near end signal d (n) figure, and Fig. 2 is the combined filter value y (n) that the present invention tests output.

Can find out in conjunction with Fig. 1, Fig. 2, the combined filter value y (n) that the present invention tests output and near end signal d (n) are almost identical, also are that the present invention is almost identical to estimated value and the echo signal of echo signal, and its soundproof effect is good.

Fig. 3 is the weight λ (n)-time graph of large step-length filter of the present invention.As shown in Figure 3, the weight λ (n) of 3 large step-length filters prolongs in time and moves closer in 0.

Fig. 4 is the normalization stable state imbalance curve of document 1 and document 2.As can be seen from Figure 4: in the situation that coloured input and step parameter value are identical, document 2 is faster than document 1 convergence rate, restrains the latter two steady-state errors basic identical.

Fig. 5 is document 3 and normalization stable state of the present invention imbalance curve.As can be seen from Figure 5: in the situation that coloured input and step parameter value are identical, the present invention that computation complexity is low is faster in the transition stage convergence rate than document 3; Restrain the latter two steady-state errors basic identical.

The approximate value of optimized parameter of each algorithm of table 1 experiment

Claims (1)

1. the listener's echo self adaptive elimination method of a low complex degree, its step is as follows:

A, remote signaling filtering

The remote signaling sampling that far-end is transmitted obtains the centrifugal pump x (n) of the current time n of remote signaling, obtains respectively large step-length filter value y after remote signaling centrifugal pump x (n) is eliminated filter filtering by the convex combination adaptive echo ₁(n), y ₁(n)=w ₁(n) ^HThe long filter value y of x (n) and small step ₂(n), y ₂(n)=w ₂(n) ^HX (n); Wherein, w ₁(n) and w ₂(n) the convex combination adaptive echo that is respectively current time n is eliminated the large step-length filter in the filter, the tap weights coefficient of the long filter of small step, and its initial value is zero, and subscript H represents conjugate transpose;

B, convex combination

With large step-length filter value y ₁(n) and the long filter value y of small step ₂(n) carry out convex combination and obtain combined filter value y (n),

y(n)＝λ(n)y ₁(n)+(1-λ(n))y ₂(n)

Wherein, λ (n) is the weight of large step-length filter, and its expression formula is

A (n) is hybrid parameter, and its initial value is 0;

C, echo cancelltion

The near-end microphone picked up after subtracting each other with the near end signal d (n) of echo and combined filter value y (n) loopback is to far-end again, backhaul signals is total residual signals e (n), e (n)=d (n)-y (n);

D, filter tap weight coefficient upgrade

With near end signal d (n), respectively with large step-length filter value y ₁(n), the long filter value y of small step ₂(n) subtract each other, obtain large step-length residual signals e ₁(n) and the long residual signals e of small step ₂(n), that is:

e ₁(n)＝d(n)-y ₁(n)，e ₂(n)＝d(n)-y ₂(n)；

Use improved proportional normalization all square filtering method calculate next constantly the adaptive echo of n+1 eliminate filter tap weight coefficient w (n+1):

$w_1(n+1)=w_1\left(n\right)+\frac{{\mathrm{\μ}}_{1}X{\left(n\right)}^T{G}_1\left(n\right)e_1\left(n\right)}{X{\left(n\right)}^T{G}_1\left(n\right)X\left(n\right)+\mathrm{\δ}}$

w ₂(n+1)＝w ₂(n)+μ ₂X(n) ^TG ₂(n)e ₂(n)

Wherein: μ ₁Eliminate the step-length of filter for large step-length echo, its value is 0.5～0.8; μ ₂Eliminate the step-length of filter for the long echo of small step, its value is 0.1～0.3; δ is regularization parameter, and its value is 0.001～0.01; The matrix that X (n) consists of to n-L+1 at n constantly for remote signaling x (n), X (n)=[x (n) ..., x (n-L+1)] ^TG ₁(n) and G ₂(n) be respectively large Step matrix and little Step matrix, calculated by following formula

$G_i\left(n\right)=\frac{1-{\mathrm{\κ}}_i}{2L}+(1+{\mathrm{\κ}}_i)\frac{\left|w_i\left(n\right)\right|}{2{\left|\right|w_i\left(n\right)\left|\right|}_1+{\mathrm{\ϵ}}_{\mathrm{ip}}},i=\mathrm{1,2}$

Wherein, || || ₁Expression 1-norm, k _iProportionality control parameter k _i∈ [1,1], i=1,2, ε _IpBe regularization parameter, its value is 0.001～0.01;

The weight of E, filter is upgraded

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

a(n+1)＝a(n)+μ _asgn[e(n)(e ₂(n)-e ₁(n))]

Wherein sgn represents sign function

μ _aBe a constant, value is 0.01; In the weight expression formula with hybrid parameter a (n+1) the substitution step B after upgrading, obtain the updating value λ (n+1) of filter weight, $\mathrm{\λ}(n+1)=\frac{1}{1+e^{-a(n+1)}};$

The weight of F, filter limits

If λ (n+1)＜0.01 then makes λ (n+1)=0; If λ (n+1)＞0.99, and 4 μ ₁＜0.5, then make λ (n+1) if=0 λ (n+1)＞0.99 and 4 μ ₁＞0.5, then make λ (n+1)=0.99

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

Images