CN100499390C - Echo suppressor, echo suppression method using normalization minimum mean-square calculation - Google Patents

Abstract

The invention relates to an echo eliminator and echo elimination method by using NLMS algorithm. Said NLMS algorithm changes self-adaption mode among slow mode, radical mode and inhibit mode according to the difference of states of echo eliminator. This realizing mode does not need correlative information, it is simple and can saves much time. NLMS algorithm is in the radical mode when communication by telephone is began to ensure quick convergence, then to switch to a slow mode after the convergence is finished to return lower residual echo.

Description

Echo eliminator and utilize the method for echo cancellation of normalization minimum mean-square calculation

Invention field:

The present invention relates to echo eliminator and utilize the method for echo cancellation of normalization minimum mean-square calculation.

Background technology:

It is the method for undesirable echo in a kind of extraordinary cancellation in communication systems that echo is eliminated.Fig. 1 shows the block diagram of the transmission network that utilizes a kind of traditional echo eliminator.This echo eliminator 200 is connected to a digital network 100 and a hybrid network 230 by one-way passage 203 and 205.The dotted line that connects in 203 and 205 is used to refer to: describedly is connected possible long enough, thereby may causes annoying echo-signal.Such echo produces in hybrid network 230, and the latter is connected to phone 52 by path 202.

Figure 2 illustrates a kind of widely used adaptive algorithm (adaptation algorithm) that is used for correcting back to the factor vector (coefficient vector) in the wave excluder 200.This adaptive algorithm is a kind of normalization LMS (NLMS) algorithm 210, and it is provided by following formula:

$h_{k+1}=h_k+\frac{\mathrm{\α}}{{\left|\right|x_k\left|\right|}^2}\·e_k\·x_k---\left(1\right)$

Wherein

$e_k=x_k^T\·(g-h_k)+n_k---\left(2\right)$

$d_k=x_k^T\·g+n_k---\left(3\right)$

Wherein, h _kBe the factor vector 212 of echo eliminator, x _kBe input vector, e _kBe residual echo, d _kBe the benchmark echo, 0＜α＜2 are scalars of a control stability and convergence rate, ‖ x _k‖ ²Be input vector x _kMould side, g is an actual ghosts path factor vector 214, n _kBe additive noise (or the adjacent speech under the situation of conversation at the same time).

Echo eliminator 200 will have good performance, and the selection of α is crucial.Little α value can guarantee that mistuning (misadjustment) chance under the stable state is little, also needs for noise passivity (noiseinsensitivity).But little α value can make convergence rate reduce.Big α value can realize restraining faster and better follow-up control usually, and cost then is under stable state, to have higher extra mean square error.Therefore, realize suitable self adaptation in NLMS algorithm 210, suitably selecting α is exactly a subject matter.

In existing document, carried out a large amount of effort, with the automatic adjusting (comprising the detection that echo path changes and converses simultaneously) of controlling described convergence rate.Be published in volume COM-26 the 5th phase 647-653 page or leaf (IEEE Trans.On Communications in May, 1978 " IEEE communication journal " at D.L.Duttweiler, Vol.COM-26, No.5, pp.647-653, May 1978) on " 12 channel digital echo eliminator " (Atwelve-channel digital echo canceller) in, proposed to be used for the Geigel algorithm of conversing simultaneously and surveying.This Geigel algorithm is taken d _kIntensity and a current value x of current instantaneous value sample _{Max, k}Compare.If d _kIntensity be higher than x at least _{Max, k}6 decibels, then judge to have conversation simultaneously.This Geigel algorithm is simple and quick.But, if d _kIntensity be lower than x in the communication process at the same time _{Max, k}6 decibels, this Geigel algorithm is just surveyed less than the existence of conversation simultaneously.In addition, this Geigel algorithm also disturbs responsive to near-end noise.

Except above-mentioned Geigel algorithm, also invented other based on relevant method, be used for automatic adaptive and the detection of conversation detection/echo path variation simultaneously." the NLMS algorithm of variable step size " (" A variable step size NLMS algorithm " at F.Casco etc., IEICE Trans.Fundamentals, Vol.E78-A, No.8, pp.1004-1009, August1995) in, a kind of LMSFIR adaptive filter algorithm of variable step size has been proposed.In this algorithm, the adjusting α of step-length is by residual echo e _kWith benchmark echo d _kBetween correlation control.But the validity of this algorithm is just verified with the white noise input.K.Fujii and J.Ohga are at " monitoring the method for conversation simultaneously by the fluctuation of monitoring echo path " (" Double-talkdetection method with detecting echo path fluctuation ", Electronicsand Communications in Japan, Part 3, Vol.78, No.4, pp.82-93,1996) in, proposed to utilize benchmark echo d _kAnd estimate standard cross-correlation between echo (estimated echo)

$y_k(x_k^T\·h_k)---\left(4\right)$

Distinguishing conversation simultaneously changes with echo path.This method supposition: under the situation that echo path changes, d _kAnd y _kBetween the standard cross-correlation level off to 0, and at the same time under the situation of conversation, corresponding cross-correlation levels off to 1.By observing the situation of described cross-correlation, just conversation simultaneously can be separated with the echo path variation zone.But there are two problems in this method.The first, when echo path ratio on intensity changed, described standard cross-correlation may level off to 1 or greater than 1, rather than 0.Second, at the same time under Tong Hua the situation, in between one section specific sampling date, described near-end is conversed simultaneously and is estimated standard cross-correlation between echo and may level off to 1 (rather than supposition 0), and this makes and is difficult to utilize corresponding cross-correlation to distinguish conversation simultaneously and echo path variation further.

As previously mentioned, many trials were arranged once, wished to design a kind of effective echo eliminator that utilizes cross-correlation.But, carry out adaptive-filtering although carried out many effort based on cross-correlation information, still no evidence shows that it is in the method for purpose that these auto-adaptive filtering techniques can successfully be applied to eliminate with the NLMS echo.

Summary of the invention:

The present invention aims to provide a kind of simply and efficiently technology of not using cross-correlation information.

Basic thought of the present invention is to change the adaptive model of NLMS algorithm between slow mode, radical pattern and prohibited mode along with the difference of echo eliminator state.When beginning to converse, the MLMS algorithm is a radical pattern, to guarantee quick convergence.After convergence was finished, the NLMS algorithm just switched to a kind of slow mode, only returned lower residual echo under this pattern.In case detect or change the enhancing of the residual echo cause by conversation simultaneously or by echo path, just earlier current adaptive-filtering factor is remembered, under radical pattern, upgrade then.During short detection delay, eliminate echo, for should from the benchmark echo, deducting which kind of echo discreet value (to the response of memory filtering (retained filter), or to the response of quick self-adapted filtering (aggressively adapted filter)) judgement, make according to the Geigel algorithm.If detected conversation simultaneously at short notice, then will remember filtering output (taps) with generating residual echo with the Geigel algorithm.Otherwise, just quick self-adapted filtering output is used for exporting residual echo.In order more accurately to judge conversation has simultaneously taken place or echo path changes, described Geigel survey of short duration during in, described quick self-adapted filtering current residual echo that generates and the residual echo that memory filtering generates are compared.If the short-term averaging of the residual echo that is generated by described quick self-adapted filtering always is lower than described memory filtering output, then can judges the echo path variation has taken place.System is upgraded with radical pattern output then, switches to the radical pattern self adaptation, causes new convergence up to reaching.Otherwise, just can judge that the enhancing system of described residual echo causes because of conversation simultaneously.With memory filtering output current filtering output is upgraded then, selected and keep for use inhibition (self adaptation is frozen) pattern, in the scope before error signal (residual echo) is reduced to variation once more.

According to an aspect of the present invention, provide a kind of method for echo cancellation here, it is characterized in that, comprising: carry out a kind of normalization minimum mean-square calculation, this algorithm be a kind of carry out under than the radical pattern of rapid convergence initialized; When convergence is finished under described radical pattern, described normalization minimum mean-square calculation is switched to a kind of slow mode that is used for low residual echo; Detect uprushing of described residual echo, and judge that it still is to be changed by echo path to cause that uprushing of this residual echo conversed by both party; If the uprushing to be changed by echo path of described residual echo cause,, then keep a kind of prohibited mode if then described normalization minimum mean-square calculation is switched to uprushing to converse by both party and causing of described radical pattern and described residual echo.

According to another aspect of the present invention, a kind of echo eliminator is provided here, it is characterized in that, comprise: a normalization minimum mean-square processor, be used to carry out a kind of normalization minimum mean-square calculation, this normalization minimum mean-square processor can a kind of than rapid convergence radical pattern and a kind of slow mode of low residual echo between switch, wherein said normalization minimum mean-square calculation is to carry out initialized according to described radical pattern computing.

Description of drawings:

Fig. 1 shows a kind of normal transmission network that uses the normal echo arrester.

Fig. 2 shows a kind of common normalization minimum mean-square (normalized least mean squared, the NLMS) algorithm that is used in the normal echo arrester shown in Figure 1.

Fig. 3 shows the performed roughly flow process of echo eliminator among a kind of embodiment of the present invention.

Fig. 4 illustrates in greater detail the performed flow process of echo eliminator among a kind of embodiment of the present invention.

Embodiment:

The solution roughly that the self adaptation of the circuit echo that the hybrid network 230 of telephone network shown in Figure 1 produces is eliminated is described below.When a new procedure of adaptation begins, perhaps in echo path, have after the big variation, described convergence factor α is set at a big relatively value, to guarantee to have the fastest convergence rate (radical pattern self adaptation).After convergence was finished, α switched to a less value, with the echo that further reduces to return, and reduced the susceptibility (slow mode self adaptation) of 200 pairs of additive noises of echo eliminator.When taking place to converse simultaneously, just forbid the renewal (prohibited mode self adaptation) of adaptive-filtering factor.

Below with reference to flow chart shown in Figure 3, make adaptive model accurately conversion between radical pattern, slow mode and prohibited mode of NLMS algorithm 210 in illustrating how during short as far as possible.

Below in conjunction with Fig. 3 the present invention is made a general description.At first, begin to make a phone call (step 302 of Fig. 3).The NLMS algorithm 210 of echo eliminator 200 enters radical pattern self adaptation state, to guarantee quick convergence (step 304 of Fig. 3).Then, NLMS algorithm 210 judges whether convergence finishes (step 306 of Fig. 3).If not, 210 of NLMS algorithms return step 304.If 210 of NLMS algorithms change the slow mode self adaptation over to, to reduce echo/noise (step 308 of Fig. 3).Then, NLMS algorithm 210 is just monitored residual echo, sees if there is uprush (step 310 of Fig. 3) of residual echo.If do not have, NLMS algorithm 210 just returns step 308, keeps the slow mode self adaptation.If monitored uprushing of residual echo, just in of short duration time-delay, export this residual echo (step 312 of Fig. 3) according to the Geigel algorithm.Then, NLMS algorithm 210 is done one and is judged more accurately, determines whether to have after described of short duration time-delay the situation (step 314 of Fig. 3) of conversation simultaneously and echo path variation.NLMS algorithm 210 is then judged (step 316 of Fig. 3) to the reason of uprushing in the residual echo.If the reason that residual echo is uprushed is to converse simultaneously, NLMS algorithm 210 just starts the prohibited mode self adaptation, and this pattern is forbidden the renewal (step 318 of Fig. 3) of factor.After step 318, NLMS algorithm 210 is judged in step 320: whether convergence is finished once more.If not, NLMS algorithm 210 just returns step 318.If NLMS algorithm 210 just judges whether conversation finishes (step 322 of Fig. 3).If conversation finishes, NLMS algorithm 210 just advances to step 324, stops its processing procedure.If conversation does not finish, NLMS algorithm 210 just advances to step 308, restarts the slow mode self adaptation.

In step 316, be that echo path changes if NLMS algorithm 210 is judged the reason that residual echo uprushes, NLMS algorithm 210 just enters the radical pattern self adaptation once more, to guarantee factor convergence (step 326 of Fig. 3) fast.Equally, NLMS algorithm 210 judges in step 328 whether convergence is finished.If not, NLMS algorithm 210 just returns step 326.If NLMS algorithm 210 just judges whether conversation finishes (step 322 of Fig. 3).Be similar to described prohibited mode self adaptation, if NLMS algorithm 210 determines that conversation finishes (step 322 of Fig. 3), NLMS algorithm 210 promptly is terminated (step 324 of Fig. 3).If do not finish as yet in step 322 conversation, NLMS algorithm 210 just advances to step 308, restarts the slow mode self adaptation.

Below in conjunction with Fig. 4 the present invention is described in detail.At first, starting speech phase (step 402 of Fig. 4), for guaranteeing quick convergence, convergence factor α is set to α _a(α _aValue can elect 0.5 as), and under radical pattern, begin adaptive process (step 404 of Fig. 4).When residual echo dropped to a particular value, promptly decidable had reached convergence state, thereby described adaptive process is transformed into slow mode (step 406 of Fig. 4, α=α _S, α wherein _SBe 0.04).Judge with following discriminant whether convergence finishes (step 406 of Fig. 4):

If | x _k| ²C|e _k| ², then convergence is finished; Otherwise not.

In step 406, absolute value sign || represent the interior energy (short timewindowed energy) of window in short-term.This in short-term the size of window be chosen as 64 sampling, C can be 1000.Note, in adaptive process,, should forbid carrying out self adaptation if there is not the remote signaling input.

After convergence was finished, 210 pairs of residual echo levels of NLMS algorithm were monitored.In case | x _k| ²＜C|e _k| ², that is to say uprush (step 408 of Fig. 4) that has detected residual echo, just show or have simultaneously and converse perhaps have echo path to change.Then, NLMS algorithm 210 stores (step 412 of Fig. 4) with current sample time and current adaptive-filtering output.Simultaneously, this filtering output adapts under radical pattern.The memory of current key parameter and the adaptive startup of filtering output radical pattern are as described below:

If the residual echo that is produced by quick self-adapted filtering is expressed as e_agg=d _k-h _k ^tX _k, h wherein _kBe quick self-adapted filter vector, and the residual echo that is produced by the filtering output of being got up by memory is expressed as e_freeze=d at sample time k _k-h_freeze ^TX _k, wherein h_freeze is the filter vector of memory.Cause if the enhancing of residual echo is changed by echo path, so, during one section specific radical pattern self adaptation after, | e_agg| should be always less than | e_freeze|.But, if the enhancing of residual echo is caused by conversation simultaneously, so, in a certain predetermined observation time limit, at least for several times sampling, | e_agg| should be greater than | e_freeze|.According to this discrimination standard, after one section specific short time period, just can determine the reason that residual echo strengthens.Change if reason is an echo path, just upgrade current filtering output immediately, and adaptive process is continued under radical pattern with described quick self-adapted output.Otherwise if reason is to converse simultaneously, then (" freezing ", filtering frozen) are exported and are upgraded current filtering output immediately, and make described self adaptation be transformed into prohibited mode with previous memory.The present invention does not eliminate near-end and converses simultaneously.

Strengthening and distinguish conversation simultaneously and echo path in residual echo has between changing one to detect and postpone.Even the performance in order to ensure circuit in this of short duration detection timing period still can allow the people accept, the present invention still will use traditional Geigel algorithm.Although there are many other methods to be suggested the solution detection problem of conversing simultaneously, and the Geigel algorithm has some intrinsic restrictions, and this kind algorithm has still obtained using widely in many business-like echo eliminator products.After the enhancing that detects residual echo, the Geigel algorithm is used for judging whether following formula satisfies (step 414 of Fig. 4):

|d _k|>0.5x _max，k (5)

If | d _k| greater than x _{Max, k}Half, think then conversation taken place simultaneously that residual echo output is (step 416 of Fig. 4) that the filter response to previous memory (freezing) subtracts and produces.Otherwise, just think the variation that echo path has taken place just from described benchmark echo, to deduct described quick self-adapted filtering output response (step 418 of Fig. 4) to generate residual echo output.In the past one section specific during delta (delta can be tens of or hundreds of sampling, for example 128 sampling) (step 420 of Fig. 4) afterwards, on a predetermined observation time limit obs_win relatively | e_agg| and | e_freeze| (this observations time limit also can be hundreds of samples, for example 384).For will be owing to the fluctuation of the residual echo of echo path in changing and issuable side effect is taken into account, use the following standard of distinguishing:

Being illustrated in the obs_win with dt_num | e_agg| is greater than | the number of times of e_freeze|.If dt_num is less than a certain predetermined little threshold value dt_threshold (dt_threshold for example can be set at 4 or 5) (step 422 of Fig. 4), just further judging is the variation that echo path has taken place.Use described quick self-adapted output undated parameter (step 424 of Fig. 4) then, and adaptive process is continued with radical pattern.If the relation that do not wait in the step 422 does not satisfy, then with the previous current parameter (step 426 of Fig. 4) of parameter update of remembering (freezing).Adaptive process just is transformed into prohibited mode then.Above-mentioned self adaptation state continues, and sets up until monitoring following formula once more:

|x _k| ²>C·|e _k| ² (6)

This formula means: or reached new convergence state, or conversation finishes (step 428 of Fig. 4 or step 430) simultaneously.Which kind of situation no matter, adaptive process is all returned (step 432 of Fig. 4) to slow mode (step 408 of Fig. 4, α=α _S), whole process repeats, until finishing conversation (step 434 of Fig. 4).

Above the realization in conjunction with the illustrated NLMS algorithm 210 of Fig. 3 and Fig. 4 does not need relevant information.This kind implementation is simply and not time-consuming.Previously described realization to NLMS algorithm 210 has high convergence rate high stability.The echo eliminator 200 of the aforementioned NLMS algorithm 210 of the execution of Shi Xianing is also simple and quick thus.

Another advantage of the present invention is that it can be handled between different initial sampling dates adaptively.This point is different with method for distinguishing, and the latter need set a constant baseline.Therefore, the present invention has extra flexibility, can deal with multiple systems adaptively.

Another advantage of the present invention is between the renewal of the accurate differentiation of conversation and echo path variation and response filtering output, to have only one section very of short duration delay at the same time.Therefore, in the present invention, be reduced owing to detecting the negative interaction that postpones to cause.

In addition, utilized the advantage of traditional Geigel algorithm at embodiment illustrated in fig. 4.Specifically, in the embodiment shown in fig. 4, the Geigel algorithm is used for simply and quickly detecting at short notice conversation simultaneously and echo path changes.After between this of short duration detection period, conversation and echo path simultaneously change are done further accurately to distinguish, to guarantee being to converse at the same time or circuit all has good performance under the situation of echo path variation.

Foregoing, especially the present invention who is illustrated in conjunction with embodiment shown in Figure 4 has provided several specific parameters.But the present invention should not be limited to these parameters (as more usually describing in conjunction with Fig. 3 institute), and for the ordinary skill in the art, the correct at an easy rate of these parameters can obtain the beneficial effect identical with the present invention equally.

In addition, the present invention above is being to describe to flow chart shown in Figure 4 in conjunction with hardware shown in Figure 2 and Fig. 3.But, the aforementioned function that belongs to echo eliminator 200 can or be downloaded in the echo eliminator 200 of packing into by product (article of manufacture) or as (have or carrier free) transmitting signal (propagatedsignal) (for example passing through the internet), as complete computer program or transmitting signal, perhaps with the form of code segment.

Above, described NLMS algorithm 210 also is described to a clear and definite independent entity of echo eliminator 200.But as known for one of ordinary skill in the art, this NLMS algorithm 210 can integrate with other entity, makes its function be implemented on a card and/or the chip piece with the form of hardware or software.Such change is routinely to those skilled in the art.

The present invention obviously can be changed in many aspects as described.Such variation should not be considered as breaking away from the spirit and scope of the present invention, and all these modifications that it will be apparent to those skilled in the art that all comprise within the scope of the appended claims.

Claims (6)

1. a method for echo cancellation is characterized in that, comprising:

Carry out a kind of normalization minimum mean-square calculation, this algorithm be a kind of carry out under than the radical pattern of rapid convergence initialized;

When convergence is finished under described radical pattern, described normalization minimum mean-square calculation is switched to a kind of slow mode that is used for low residual echo;

Detect uprushing of described residual echo, and judge uprushing of this residual echo converse by both party cause or change by echo path and to cause;

If the uprushing to be changed by echo path of described residual echo cause, then described normalization minimum mean-square calculation is switched to described radical pattern and

If the uprushing to converse by both party of described residual echo causes, then keep a kind of prohibited mode.

2. method according to claim 1 is characterized in that described method is irrelevant.

3. method according to claim 1 is characterized in that, described detection step comprises:

It is constant to keep the adaptive-filtering factor in the moment of uprushing of detecting residual echo,

Under described radical pattern, upgrade described adaptive-filtering factor,

In given period, judge roughly described residual echo be uprush converse by both party cause or by echo path change cause and

After the described given period, accurately judge uprushing of described residual echo converse by both party cause or change by echo path and to cause.

4. method according to claim 3 is characterized in that, described rough determining step is carried out the Geigel algorithm.

5. method according to claim 3, it is characterized in that, described accurate determining step comprises: in a given observation window, the residual echo that the adaptive-filtering factor of determining to upgrade under described radical pattern is produced is greater than the number of times of the residual echo that is produced by the described adaptive-filtering factor that remains unchanged.

6. method according to claim 5, it is characterized in that, if the residual echo that the adaptive-filtering factor that upgrades under described radical pattern in a given observation window is produced greater than the number of times of the residual echo that is produced by the described adaptive-filtering factor that remains unchanged more than several times, the uprushing to converse by both party and cause of described residual echo then, otherwise the uprushing to change and cause of described residual echo by echo path.

Images