CN1212555A - Direction transform echo canceller and method - Google Patents

Abstract

An echo canceler converts a remote terminal input voice to its linear predictive remainder. At a coefficient generating circuit, an updating direction to the coefficient of the echo canceler is formed in two stages. In the first stage, the updating direction is formed in the region of the linear predictive remainder. On the second stage, the updating direction is converted from the remainder region to an audio region, so that an updating method to be used for adapting the coefficient can be provided.

Description

Direction transform echo canceller and method

The present invention relates to Echo Canceller, relate to the rapid convergence Echo Canceller more specifically.

Some intercommunication systems have different passages to the signal that in the opposite direction sends.In such system, the signal in passage may be reflected in another passage.These are commonly called the signal of the reflection of echo signal, disturb required signal of communication.As a result, people have developed Echo Canceller to suppress these reflected signals.

Echo Canceller uses sef-adapting filter to estimate to reflex to the echo signal in one of signalling channel.The estimation of this echo is deducted from the signalling channel that comprises this echo signal, does not have echo or echo to be subjected to the signal that presses down substantially to produce.

Because lowest mean square (LMS) sef-adapting filter has better simply structure, and calculation stability is efficient, they are widely used in the Echo Canceller with the estimated echo signal.Yet when being used for acoustic echo such as videoconference and hands-free cellular telephone communication and eliminate using, the shortcoming of LMS sef-adapting filter is slow convergency factor.

To be them be no more than 30dB (decibel) to the inhibition of echo signal to another shortcoming of LMS sef-adapting filter usually.In typical acoustic applications, echo signal is very strong, and may be the same with required signal of communication strong.In such environment, sef-adapting filter is convergence fast, simulating quick variation echo passage, and wishes that very echo signal is suppressed 40dB at least.If these two conditions can not be satisfied, big echo evaluated error will occur, and these mistakes will cause the serious decline of required signal of communication signal quality.

Adapt to soon with sef-adapting filter than Echo Canceller with the LMS filter as other algorithms such as recursive least-squares (RLS) and affine projection algorithms, but they or unstable, or calculate expensive.Calculate expensive filter and need big circuit or a lot of treatment facility.Unsettled sef-adapting filter will produce erratic estimation, and can't realize therefore guaranteeing that the user can be by calling out the clearly sufficiently stable operation of communication.

In the environment of hands-free mobile phone and hands-free phone meeting and so on, the sef-adapting filter of convergence especially fast, stable, high inhibition and efficient calculation.Echo passage in this environment stands quick variation, strong echo and high background noise.Therefore, need the sef-adapting filter convergence of Echo Canceller fast, and suppress echo signal more than 40dB.Also wish Echo Canceller energy efficient calculation, so that it can be realized in having the less circuit of low-power consumption feature.

The purpose of this invention is to provide a kind of direction transform echo canceller and method, it has such as suffering can to stablize in acoustic echo disturbs, wherein echo path changes fast and echo signal the is strong environment and the filter of efficient calculation.

The invention provides a kind of method of upgrading direction that produces in Echo Canceller, described Echo Canceller comprises sef-adapting filter, and described method is characterised in that and may further comprise the steps:

In the linear prediction residue field, calculate and upgrade direction;

Described renewal direction is transformed in the speech domain from the linear prediction residue field, is used for the speech domain renewal direction that described adaptive filter coefficient upgrades with generation.

Advantage of the present invention is: can be easy to realize, stable, function is strong and it is efficient to calculate.

Brief Description Of Drawings

Fig. 1 is the circuit diagram of describing according to the Echo Canceller of prior art;

Fig. 2 describes when only having two coefficients to be L=2, the echo that the inaccuracy of the coefficient of LMS self-adaptive echo eliminator is estimated to cause estimate etc. the example of mean square error curve;

Fig. 3 describes when only having two coefficients to be L=2, echo that the inaccurate estimation of the coefficient of LMS self-adaptive echo eliminator causes estimate etc. another example of mean square error curve;

Fig. 4 is a circuit diagram of describing Echo Canceller;

Fig. 5 is depicted in the linear prediction residue field and upgrades direction and only use with two coefficients, promptly under the situation of L=2, the caused echo of the inaccurate estimation of the coefficient of self-adaptive echo eliminator estimate etc. the example of mean square error curve;

Fig. 6 describes only to use with two coefficients by upgrading direction in the speech domain that the renewal direction converts in the linear prediction residue field, be under the situation of L=2, the caused echo of the inaccurate estimation of the coefficient of self-adaptive echo eliminator estimate etc. the example of mean square error curve;

Fig. 7 is a flow chart of describing to produce improvement direction.

Echo Canceller conversion far-end voice sample is its linear prediction residue sampling, and the renewal direction of the coefficient of Echo Canceller builds in the linear prediction residue field.Then, the renewal direction in the residue field is converted into the renewal direction in the speech domain.The coefficient self adaptation of Echo Canceller is carried out with the renewal direction of the conversion in the speech domain.Calculate the renewal direction by conversion, the renewal direction that can obtain to optimize from the linear prediction residue field to speech domain.This speed that makes Echo Canceller be adapted to the echo passage improves greatly, thereby allows it to suppress more noise, and maintenance simultaneously is stable, efficient, function is strong.

Traditional lowest mean square (LMS) Echo Canceller 100 is shown in the device 101 of Fig. 1.Device 101 comprises microphone 104 and the loud speaker 102 that is used for hands-free operation.Speaker path comprises the decoder 106 and digital-to-analogue (D/A) transducer 108 of the receiver output that is connected to transceiver 120.Microphone channel comprises modulus (A/D) transducer 114, combiner 112 and the encoder 116 of the transmitter that is coupled to transceiver 120.Therefore signal x (n), y (n) and e (n) are digital signal.Transceiver 120 is coupled to the antenna 122 in the radio telephone.

Device 101 can be the hand-free kit that is connected to mobile radiotelephone, also can be radio telephone or public conference phone.Person of skill in the art will appreciate that Echo Canceller 100 can realize in the bidirectional communication apparatus that numeral and/or analog circuit are arranged.Therefore decoder 106 and encoder 116 only are used to describe purpose.If if transceiver 120 is in analog cellular telephone or simulation land line phone or installs 101 analog outputs that are connected to digital cellular telephone or digital land line phone that then decoder 106 can be an A/D converter.Perhaps, decoder 106 can be as the digital decoder in the cellular digital device of numeral.If if transceiver 120 is in analog cellular telephone or simulation land line phone or installs 101 analog input ends that are connected to as the digital device of digital cell phone or digital land line phone that then encoder 116 can be a D/A converter.Perhaps, encoder 116 can be as the digital encoder in the cellular digital device of numeral.

Describe Echo Canceller 100 referring now to Fig. 1 to Fig. 3.Current sampling instant is n, far-end voice sample x (n) is the output of loud speaker 102, and near-end audio signal y (n) receives from microphone 104, in this manual, signal x (n) and y (n) are synchronous, i.e. digital to analog converter 108 and the identical clock of analog to digital converter 114 usefulness.Near-end audio signal y (n) comprises adjacent speech s (n), echo t (n) and near-end noise N (n).Echo t (n) is the part that is reflected back toward microphone 104 among the signal x (n) by loud speaker 102 output.Noise N (n) can be, for example the ambient noise in the automobile cab.

In the following description, when the coefficient that is located at sef-adapting filter carried out self adaptation, adjacent speech s (n) was zero (being that adjacent speech s (n) does not exist).When near-end and far-end speech all exist, i.e. alleged two-way call usually, self adaptation will have to stop.Be used to detect this situation and prevent and under this condition, carry out adaptive two-way call detector, for known in the art, for easy, no longer meticulous here description.Therefore, said, near-end audio signal y (n) only comprises echo t (n) and noise N (n).

Echo Canceller 100 comprises that coefficient produces circuit 109 and LMS sef-adapting filter 110 comes the analog echo passage.Estimate that by the echo of LMS sef-adapting filter 110 generations z (n) makes up with letter y (n) in combiner 112, to remove or the inhibition echo signal.Echo evaluated error e (n) is that echo estimates that z (n) and near end signal y's (n) is poor.The coefficient of LMS sef-adapting filter 110 produces circuit 109 by coefficient and upgrades based on the far-end voice sample and the echo evaluated error that receive.

At moment n, the coefficient of LMS sef-adapting filter 110 is W (n)=[W ₀(n) W ₁(n) ... W _L-1(n)] ^TWherein L is a filter length, subscript T (i.e. () ^T) transposition of representative vector.L far-end voice sample that receives be X (n)=[x (n), x (n-1) ..., x (n-L+1)] ^TThe echo of LMS sef-adapting filter 110 estimates that z (n) is:

z(n)=X(n) ^TW(n) (1)

The echo evaluated error e (n) of LMS sef-adapting filter 110 is:

e(n)=y(n)-z(n) (2)

Establish an equation under coefficient W (n) basis of LMS sef-adapting filter 110 and be updated: $W(n+1)=W\left(n\right)+\frac{\mathrm{\μ}}{\left|\right|X\left(n\right){\left|\right|}^2}\left(n\right)X\left(n\right)------\left(3\right)$

Wherein μ is an adaptive step, and || X (n) || ²=X (n) ^TX (n).

R (n)=E{X (n) X (n) ^T}={ rij (n) | i, j=0,1 ..., L-1} is the autocorrelation matrix of far-end voice sample.E{ ^*Be statistical average.Because far-end voice signal right and wrong stably, R (n) becomes the LXL matrix when being, the positive eigenvalues { λ that becomes when having L _i(n) i=0,1,, become eigenvector { v in the time of L-1} and L _i(n) i=0,1 ..., L-1}.S (n)=E{y (n) X (n) }=[s ₀(n) s ₁(n) ... s _L-1(n)] ^T=[E{y (n) x (n) } E{y (n) x (n-1) } ... E{y (n) X (n-L+1) }] ^TIt is the cross-correlation vector of far-end voice sample and near-end audio signal.

When depending on L, the performance of sef-adapting filter becomes the value of eigenvalue and eigenvector.Known μ is that fix and necessary according to following principle selection:

0＜μ＜2/λ _max(n) (4)

K wherein _Max(n) be the dominant eigenvalue of R (n).Because far-end speech right and wrong stably, λ _Max(n) great dynamic range is arranged.Yet, be to keep the stability of sef-adapting filter, all constantly μ all must be very little with the scope that remains on equation (4) in.Being used to keep the little μ of stability is a slow reason of LMS sef-adapting filter convergence.

The optimization coefficient of LMS sef-adapting filter 110 is W ₀=[W ₀W ₁..., W _L-1] ^TCommon W ₀Change in time is very slow, therefore can be looked at as constant vector here.The LMS sef-adapting filter is provided by following formula with the echo evaluated error that current coefficient W (n) produces:

ε (n)=y (n)-z (n)=y (n)-X (n) ^TW (n) (5) wherein y (n) is the near end echo signal.Below consider the situation that y (n) is made up of echo and noise.In following analysis, establish near-end noise and do not exist.Wherein, W (n) and y (n) and x (n) uncorrelated (this hypothesis is ballpark when W (n) restrains, so W (n) is looked at as constant vector).The echo estimated mean-square of LMS self-adaptive echo eliminator is defined by following formula: ξ=E{ ε ²(n))=E ([y (n)-X (n) ^TW (n)] ²}

=E{y(n) ²}+E{W(n) ^TX(n)X(n) ^TW(n))-2E{y(n)X(n)W(n) ^T}

=E{y(n) ²}+W(n) ^TE{x(n)X(n) ^T}W(n)-2E{y(n)X(n)}W(n) ^T

=E{y(n) ²}+W(n) ^TR(n)W(n)-2S(n)W(n) ^T (6)

Because W ₀Be the optimization solution of W (n), and E is the function of W (n), so the slope of ξ is at W ₀Be 0, it establishes an equation under satisfying: $\∂\mathrm{\ξ}\∂W\left(n\right)=0$  2R (n) W ₀-2S (n)=0  W ₀=R (n) ^-1S (n). with (7) formula substitution (6) formula, we can obtain the least mean-square error that echo is estimated: ξ with establishing an equation down (7) _Min=E{y (n) ²}-S (n) ^TW ₀(8) though ξ _MinChange in time, but that it changes in time is very slow, so it is counted as steady state value here.With (8) and (7) substitutions (6), can confirm following relationship: ξ=ξ _Min+ [W (n)-W ₀] ^TR (n) [W (n)-W ₀] (9)

The LMS sef-adapting filter drives W (n) as possible to its true value W ₀, or drive ξ to ξ _MinTwo coefficients are being arranged, and promptly under the situation of L=2, the convergence state of sef-adapting filter can be understood, wherein W (n)=[W ₀(n) W ₁(n)], W ₀=[W ₀W ₁].The convergence state of sef-adapting filter existing with reference to shown in Fig. 2 and Fig. 3 etc. all square curve interpretation explain.By setting that ξ equals different steady state values and with W ₀(n) and W ₁(n) as the axle curve plotting.These curves are at W ₀(n)-W ₁(n) ellipse in the plane, the steepness of the wherein oval main shaft of the eigenvector of R (n) definition, and eigenvalue definition error surface.Therefore, bigger eigenvalue has long axle, and the length of the eigenvalue of matrix R (n) definition axle.

Because voice signal is astable, the ellipse that becomes when being etc. the shape of mean square error curve.The LMS sef-adapting filter upgrades direction with gradient direction 11 among Fig. 2 and 13 conducts of the gradient direction among Fig. 3.The gradient direction is only simple and reliable direction of upgrading.As can be seen, in most of the cases, comprise these two matrix R (n) the time become the example of eigenvalue, the optimization direction 12 shown in direction 11 relative Fig. 2 alters a great deal, renewal direction 13 alters a great deal with respect to the optimization direction 14 shown in Fig. 3.

Shape and size etc. the mean square error curve change in time.Therefore, the LMS sef-adapting filter is usually with departing from the renewal direction of optimizing direction greatly.This is in the application that acoustic echo is eliminated and so on, the convergence of LMS sef-adapting filter slow and echo signal can not be suppressed down 40dB another reason.

Fig. 4 illustrates the system configuration of improved Echo Canceller 200.Echo Canceller 200 available microprocessors, digital signal processor, microcomputer, computer or any other suitable circuit are realized.Echo Canceller 200 comprises linear prediction circuit 202, is connected with reception far-end voice signal X (n), input signal, and at output 203 output linear predictor coefficient and residual signal d (n).The linear prediction circuit can have any suitable linear prediction circuit to realize.Because have many fast algorithms to can be used for estimating the coefficient of linear prediction error filter, use this filter more useful.The preferred fast algorithm that uses calculating that can be efficient, stable.Selection has the linear prediction error filter of these features to guarantee that error concealment device 200 is stable with efficiently.

Linear prediction analysis can provide point-device speech parameters to estimate.To the speech of saying, linear prediction system is encouraged by impacting series excitation, and for voiceless sound, linear prediction system is encouraged by random noise sequences.The parameter of linear prediction system slowly changes in time.Relation between voice sample s (n) and excitation u (n) is by following The Representation Equation: $s\left(n\right)=-\underset{k=1}{\overset{p}{\mathrm{\Σ}}}{\mathrm{\α}}_{k}s(n-k)+\mathrm{Gu}\left(n\right)------\left(10\right)$

Wherein p is the level of time varying digital filter, and G is a gain factor; { α _kShu K=1,2 ..., p} is the coefficient of time varying digital filter.

The linear prediction of voice sample produces the estimation of current voice sampling from the linear combination of previous sampling.Wherein Be the linear prediction of s (n), Can predictive filter coefficient { α be arranged with following _kShu k=1,2 ..., the linear prediction symbol of p} obtains: $\stackrel{\mathrm{\Λ}}{s}\left(n\right)=-\underset{k=1}{\overset{p}{\mathrm{\Σ}}}{\mathrm{\α}}_{k}s(n-k).-------\left(11\right)$ The predicated error of the linear prediction symbol of equation (11) is: $e\left(n\right)=s\left(n\right)-\stackrel{\mathrm{\Λ}}{s}\left(n\right)=s\left(n\right)+\underset{k=1}{\overset{p}{\mathrm{\Σ}}}{\mathrm{\α}}_{k}s(n-k).----\left(12\right)$

From equation (12) as can be seen, the predicated error sequence is the output of linear prediction error filter, and its transfer function is: $A\left(z\right)=1+\underset{k=1}{\overset{p}{\mathrm{\Σ}}}{\mathrm{\α}}_k{z}^{-k}.------\left(13\right)$

Estimate linear prediction error filter coefficient { α _kShu k=1,2 ..., the recurrence fast algorithm of p} is known.The linear prediction error filter is a prewhitening filter, because they produce the residual signal that white noise character is arranged.In the present invention, Paul levinson one moral guest (Levinson-Durbin) recursive algorithm is useful especially.This filter has the feature of a plurality of hope, and it is highly suitable for speech prediction, so it is well understood and develops; It calculates efficient, so it can be realized in less circuit.Another key property of this algorithm is to use the linear prediction error filter of this filter stable.

Echo Canceller 200 frame by frames operate on the audio sample of reception.The audio sample that receives represent the remote signaling of subtend loud speaker 102 outputs and the sampling of the near end signal imported from microphone 104.Audio sample to far-end and near end signal is synchronous.Frame size K is usually between 50 to 200 samplings.In case receive the frame from K new audio sample of far-end and near-end, algorithm brings into operation.K far-end voice sample that receives is x (n), x (n-1) ... x (n-K+1), K near end signal sampling that receives is y (n), y (n-1) ... y (n-K+1), wherein n is current sampling instant.

Be in operation, K far-end voice sample that receives, i.e. the linear prediction analysis of input signal sampling is at first carried out in linear prediction circuit 202.Linear prediction error filter factor { α _k| k=1,2 ..., p} can be estimated by various algorithms, as estimating by Paul levinson-De Bin algorithm.The level p that is used for the linear prediction error filter of linear prediction circuit 202 can be in 5 to 15 scope for example.The transfer function A (z) that is used for the linear prediction error filter of linear prediction circuit 202 is shown below: $A\left(z\right)=1+\underset{k=1}{\overset{p}{\mathrm{\Σ}}}{\mathrm{\α}}_k{z}^{-k}.--------\left(14\right)$

The coefficient of sef-adapting filter produces in the circuit 204 at coefficient and upgrades, and this circuit is the coefficient update circuit.Adaptive echo is eliminated the audio sample that receives at K, i.e. input signal sampling is gone up with the execution of coefficient in LMS sef-adapting filter 110 of upgrading.Time index i in the time range of the new audio sample that receives of K, i.e. n≤i≤n-K+1.W (i)=[W ₀(i) W ₁(i) ... W _L-1(i)] ^TBe the coefficient of self-adaptive echo eliminator at moment i, L is the length of self-adaptive echo eliminator, and X (i)=[X (i) X (i1) ... x (iL+1)] ^TBe the vector of L far-end voice sample that receives at moment i.L can less than or greater than K.

The execution echo that the sampling that 200 pairs K of Echo Canceller receives is sampled is one by one eliminated, that is: i=n-K+1, n-K+2 ..., n, Echo Canceller 200:

When i=n-K+1, carry out echo elimination with coefficient W (n-K+1), and update coefficients W (n-K+1) is coefficient W (n-K+2);

When i=n-K+2, carry out echo elimination with coefficient W (n-K+2), and update coefficients W (n-K+2) is coefficient W (n-K+3); And

Continue this operation in each continuous sampling instant, until last sampling instant at i=n, carry out echo elimination with coefficient W (n), update coefficients W (n) is coefficient W (n+1); And wait for that next frame is to repeat this sequence to this frame.

At particular moment i, the echo evaluated error of self-adaptive echo eliminator is:

e(i)=y(i)-X(i) ^TW(i) (15)

Wherein y (i) is the near-end audio sample at moment i.

At moment i, the coefficient update direction of self-adaptive echo eliminator is based upon in the linear prediction residue field.By the far-end voice sample d (i) that the far-end voice sample generation that receives in moment i filtering with linear prediction error filter A (z) receives, d (i-1) ..., L the linear prediction residue sampling of d (i-L+1): $d\left(j\right)=x\left(j\right)+\underset{k=1}{\overset{p}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{k}x(j-k)j=1,i-1,\&hellip;,i-\mathrm{<figure-callout\; id="1"\; label="L\; +"\; filenames="A9811885100192.png"\; state="\{\{state\}\}">L</figure-callout>}+1.---\left(16\right)$ $Q\left(i\right)\frac{\mathrm{\&mu;}}{\left|\right|D\left(n\right){\left|\right|}^2}e\left(i\right)D\left(i\right)-------\left(17\right)$ Self-adaptive echo eliminator coefficient update direction is Q (i) in the linear prediction residue field, wherein:

Wherein D (i)=[d (i), d (i-1) ... d (i-L+1)] ^T, Q (i)=[q ₀(i), q ₁(i) ... q _L-1(i)] ^T, μ is an adaptive step, and || D (i) || ²=D (i) ^TD (i)

Be used for after the linear prediction residue field is carried out adaptive coefficient update direction in acquisition, produce in the circuit 204 at coefficient, the renewal direction in speech domain can be calculated based on A (z) and Q (i).The coefficient update direction of self-adaptive echo eliminator in speech domain is G (i)=[g ₀(i) g ₁(i) ... g _L-1(i)] ^TTo each component of G (i) by following Equation for Calculating G (i): $g_j\left(i\right)=q_j\left(i\right)+\underset{k-1}{\overset{p}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_k{q}_{j-k}\left(i\right)----j=\mathrm{0,1},\&hellip;,L-1\left(18\right)$

Wherein for separating above equation, to j=-1 ,-2 ... ,-p establishes g _j(i)=0.

Equation (18) is represented the filter that different traffic directions are arranged with equation (16).Equation (16) is carried out forward direction linear prediction filtering, so that the far-end voice sample is transformed into the linear prediction residue field from speech domain.Equation (18) is carried out the back to linear prediction, is transformed into speech domain will upgrade direction from the linear prediction residue field.

After obtaining G (i), the coefficient of self-adaptive echo eliminator can directly be upgraded by following formula:

W(i+1)=W(i)+G(i) (19)

Wherein W (i+1) is the coefficient that is used for the renewal of next sampling at moment i+1.

Because being used for the coefficient of the linear prediction error filter of voice signal changes slowly in time at each frame, in fact the coefficient of linear prediction error filter is constant in a frame, becomes the FIR filter when non-so it can be considered to be at a frame neutral line of voice sample.The coefficient of one frame of voice sample is estimated by Paul levinson one moral guest algorithm.And the linear prediction error filter is a prewhitening filter, thus linear prediction residue actual be white noise.So the remaining autocorrelation matrix of linear prediction is the autocorrelation matrix of white noise.

P (i)=E{D (i) D (i) ^T}={ P _Ij(n) | i, j=0,1 ... L-1} is the remaining autocorrelation matrix of linear prediction.If the renewal direction in linear prediction residue field Q (i) is used as the renewal direction in the self adaptation, at moment i, LMS echo estimated mean-square β (i) is obtained by equation (5) to (9):

β(i)=β _min+[V(i)-V ₀] ^TP(i)[V(i)-V ₀] (20)

β wherein _MinIt is the minimum value of β (i).As previously mentioned, because it changes slowly in time, it can be counted as steady state value.V (i) is the coefficient of Echo Canceller at moment i, and it is the convolution of W (i) and (1/A (z)).V ₀Be the optimization solution of previously defined coefficient V (i), and be counted as constant vector slowly because of changing in time.V ₀Be W ₀Convolution with (1/A (z)).

Fig. 5 illustrate the linear prediction residue field etc. the mean square error curve.P (i) has L identical eigenvalue, and etc. the mean square error curve be round.As a result, slope direction 51 is only reliable renewal directions that the LMS filter can be used, and slope direction 51 is optimization directions of any point on circle.V (i) converges to V ₀Speed converge to W faster than W (i) far away ₀Speed.Please note that V (i) is not equal to W (i) and is actually W (i) and the convolution of (1/A (z)).The renewal direction of optimizing can be found to be Q (i) in residue field.

The optimization of W (i) in speech domain is upgraded direction and can be found from the information available of Q (i) and A (z).Utilize following equation to carry out the self adaptation of V (i):

V(i+1)=V(i)+Q(i). (21)

The renewal direction in using the linear prediction residue field, this LMS sef-adapting filter with Fig. 1 is the same.After convergence, coefficient is not equal to real coefficient W (i), but with the convolution of (1/A (z)).Because linear prediction error filter A (z) becomes during right and wrong in a frame of the far-end voice sample that receives, W (i) is the convolution of the A (z) of V (i).Equation (21) and A (z)) convolution obtains:

A (z)  V (i+1)=A (z)  [V (i)+Q (i)] (22)

Wherein  represents the convolution of two filters.Therefore, above-mentioned self adaptation equation obtains to upgrade direction by the renewal direction in the residue field and A (z) being carried out convolution, and the renewal direction of this conversion just is used in the Echo Canceller 200.

The renewal direction of conversion is also pointed to the direction of the optimization in speech domain, such as by following proof:

(1) because Q (i) be the adaptive optimization direction of V (i), so have a scalar η so that:

V(i)+ηQ(i)-V ₀=0；

(2) do not upgrade direction if A (z)  Q (i) is not the optimization of W (i), then do not have scalar ce to exist so that:

W(i)+α[A(z)Q(i)]-W ₀=0；

(3) because A (z) is linear-time invariant filter and W (i)=A (z)  V (i), so above equation can be rewritten as:

A(z)[V(i)+αQ(i)-V ₀]=0；

(4) very clear, have at least a scalar ce=η that above equation is set up.Therefore, A (z) upgrades direction with the optimization of the convolution generation W (i) of Q (i).

According to equation (5) to (9), the echo estimated mean-square of Echo Canceller 200 can by under establish an equation and similarly express:

ε(n)=y(n)-z(n)=y(n)-X(n) ^TW(n)(23)

Wherein y (n) is a near end signal, W (n)=[W ₀(n) W ₁(n) ... W _L-1(n)] ^TIt is the coefficient of Echo Canceller 200.X (n)=[x (n), x (n-1) ..., x (n-L+1)] ^TIt is the far-end voice sample that receives.The echo estimated mean-square of new self-adaptive echo eliminator is defined by following formula: ξ N=E{ ε ²(n) }=E{[y (n)-X (n) ^TW (n)] ²}=ξ _Min+ [W (n)-W ₀] ^TR (n) [W (n)-W ₀] ξ wherein _Min=E{y (n) ²}-S (n) ^TW ₀Be the least mean-square error of Echo Canceller 200, W ₀=[W ₀W ₁..., W _L-1] ^TIt is the optimization coefficient of Echo Canceller 200.R (n)={ X (n) X (n) ^TAs above definition.The same etc. the LMS Echo Canceller of mean square error curve and Fig. 1.Yet, upgrade direction and be modified, as shown in Figure 6.

For two coefficients are arranged, i.e. the situation of L=2, the mean square error curve of new sef-adapting filter is shown among Fig. 6, wherein W (i)=[W ₀(i) W ₁(i)], W ₀=[W ₀W ₁].Different with the situation of the LMS sef-adapting filter of Fig. 1, new sef-adapting filter does not upgrade direction with gradient direction 61 as it.Coefficient produces circuit 204 and determines the coefficient update direction in two steps: at first, it finds the renewal direction in the linear prediction residue field of far-end speech; Then, it is changed this renewal direction and enters speech domain, so that obtain to optimize renewal direction 62.It is total with optimizing the renewal direction, the eigenvector that the far-end speech of therefore avoiding prior art LMS filter to be experienced changes and the adverse effect of eigenvalue that coefficient produces circuit 204.

Echo Canceller (Fig. 2) is restrained far away faster than traditional LMS self-adaptive echo eliminator.And Echo Canceller 200 can suppress acoustic echo more than 60dB, and when being used for the application of acoustic echo elimination, this is especially favourable.

Therefore as can be seen, Echo Canceller 200 (Fig. 2) is with traditional LMS sef-adapting filter 110.Yet coefficient produces the improved renewal direction of circuit 204 usefulness.Upgrade the output that direction is taken from the linear prediction error filter, this filter calculates linear predictor coefficient by the input signal sampling, shown in square frame 700 (Fig. 7).The algorithm that calculates the linear prediction error filter coefficient is carried out easily, is stablized and be quick.Because Paul levinson-De Bin algorithm is known and calculates efficiently,, be useful so use it so that the size that does not need too to increase echo cancel circuit just can realize.Other rudimentary finite impulse response (FIR) filter is known, and can be used to calculate linear prediction information.

Renewal direction in the linear prediction residue field produces from the linear prediction residue, shown in square frame 701.This is converted into linear prediction residue back in the far-end voice sample and realizes.This relates to the FIR filtering of input signal sampling.The level of FIR filter can, for example about 10.Therefore, required calculating strength is little and stable guaranteed.Like this, upgrade direction in the linear prediction residue field (signal of handling by linear prediction filter is in the linear prediction territory) at first.

Then, upgrade direction and from the linear prediction residue field, be transformed in the speech domain, shown in square frame 702.Speech domain is a digital voice, and the territory is changed available FIR filter and realized.The renewal direction that coefficient generation circuit 204 is used in the speech domain is upgraded adaptive filter coefficient.

As everyone knows, the LMS Echo Canceller is simple, function strong, it is efficient to calculate.The Echo Canceller 200 preferred LMS filters that use to utilize these characteristics of LMS Echo Canceller, by using improved renewal direction, improve the speed of self adaptation convergence simultaneously.Can reach a conclusion, improved Echo Canceller can easily be realized, stablizes, function is strong and calculate efficient.Simulated experiment shows that Echo Canceller 200 function in the environment of making a lot of noise is also stronger, and the conversion of renewal direction provides the convergence rate of accelerating greatly to reach the noise suppressed that improves greatly that forms owing to restraining fast.Person of skill in the art will appreciate that the advantage of the renewal direction of conversion can be used in the non-LMS Echo Canceller.Therefore, " sef-adapting filter " used herein and " Echo Canceller " are not limited to LMS sef-adapting filter and LMS Echo Canceller.

Claims (10)

1, produce the method for upgrading direction in Echo Canceller, described Echo Canceller comprises sef-adapting filter, and described method is characterised in that and may further comprise the steps:

In the linear prediction residue field, calculate and upgrade direction;

Described renewal direction is transformed in the speech domain from the linear prediction residue field, is used for the speech domain renewal direction that described adaptive filter coefficient upgrades with generation.

2, the method for upgrading direction that in Echo Canceller, produces as claimed in claim 1, it is characterized in that: use Paul levinson-De Bin algorithm to produce the linear prediction residual signal, according to the described renewal direction of described linear prediction residual signal calculating in the linear prediction residue field.

3, the method for upgrading direction that produces in Echo Canceller as claimed in claim 1, it is characterized in that: the step of the described renewal direction of described calculating in the linear prediction residue field comprises: draw the linear prediction residual signal that calculates with linear predictor coefficient.

4, the method for upgrading direction that produces in Echo Canceller as claimed in claim 3, it is characterized in that: described linear prediction residual signal produces with the linear prediction error filter that a plurality of linear predictor coefficients are arranged.

5, the method for upgrading direction that in Echo Canceller, produces as claimed in claim 4, it is characterized in that: described linear prediction residual signal is d (n), and $d\left(n\right)=x\left(n\right)+\underset{k=1}{\overset{p}{\mathrm{\&Sigma;}}}\mathrm{\&alpha;}{k}^{x(n-k)},$

Wherein p is a filter stage, and x (n) and x (n-k) are respectively the current of input signal and sampling in the past, { α _kIt is the linear predictor coefficient of described linear prediction error filter.

6, the method for upgrading direction that produces in Echo Canceller as claimed in claim 3, it is characterized in that: described switch process is included in the described renewal direction of filtering in the finite impulse response filter, to obtain the described renewal direction in speech domain.

7, the method for upgrading direction that produces in Echo Canceller as claimed in claim 3, it is characterized in that: described switch process comprises with described linear predictor coefficient changes described renewal direction to speech domain.

8, the method for upgrading direction that produces in Echo Canceller as claimed in claim 5, it is characterized in that: the step of the described renewal direction of described conversion comprises according to the described renewal direction g of following formula calculating in speech domain _n(i): $g_n\left(i\right)=q_n\left(i\right)+\underset{k=1}{\overset{p}{\mathrm{\&Sigma;}}}\mathrm{\&alpha;k}{q}_n(i-k)$ Wherein L is the length of described sef-adapting filter, and P is the level of described linear prediction error filter, Q (i)=[q ₀(i), q ₁(i) ... q _L-1(i)] ^TBe the described renewal direction in the linear prediction residue field, G (i)=[g ₀(i), g ₁(i) ... g _L-1(i)] ^TBe that described speech domain is upgraded direction, and { α _kIt is the coefficient of described linear prediction error filter.

9, the method for upgrading direction that in Echo Canceller, produces as claimed in claim 1, further comprising the steps of:

Estimate (z (n)) with current coefficient of described sef-adapting filter (W (n)) and far-end voice sample (x (n)) echogenicity;

Produce coefficient of linear prediction wave filter (d (n)) according to described far-end voice sample;

Produce the linear prediction residual signal by described coefficient of linear prediction wave filter;

Described renewal direction from the described Echo Canceller of described linear prediction residual signal calculating the linear prediction residue field; And

Produce the coefficient of the renewal of described sef-adapting filter according to described current coefficient and described speech domain renewal direction.

10, a kind of Echo Canceller is characterized in that its execution any one method of determining according to claim 1-9.

Images