WO2002039709A1

WO2002039709A1 - Control system of acoustic echo cancellers for telephone terminals with handset or handsfree

Info

Publication number: WO2002039709A1
Application number: PCT/IT2001/000517
Authority: WO
Inventors: Vojko Pahor; Alberto Carini
Original assignee: Telit Mobile Terminals SpA
Current assignee: Telit Mobile Terminals SpA
Priority date: 2000-11-07
Filing date: 2001-10-12
Publication date: 2002-05-16
Anticipated expiration: 2003-05-07
Also published as: ITTS20000017A0; ITTS20000017A1; AU2001295906A1

Abstract

The invention proposes a variable step-size adaptation control method for acoustic echo cancellation digital filters for handset and hands-free telephone terminals. The variable step-size integrates a double-talk detector, an echo-path change detector and an automatic control of the background noise level. The invention is based on two supplementary adaptive filters with low memory length, which support the acoustic echo cancellation filter control. One supplementary filter monitors the near-end speech activity, while the other monitors the echo-path variations. Such control method provides several benefits: 1. A reliable adaptation control for a broad range of acoustic coupling levels between the loudspeaker/s and the microphone. 2. An adaptation control able to cope also with sudden and strong changes of the echo path. 3. An adaptation control with low computational complexity characterized by a reduced number of easily adjustable parameters. 4. An adaptation control that does not need any a-priori estimate of the acoustic coupling level between the loudspeaker/s and the microphone. 5. An adaptation control that can be applied to any gradient descendent adaptive algorithm.

Description

Control system of acoustic echo cancellers for telephone terminals with handset or handsf ree

Technical Field of the Invention

The present invention relates to the problem of the acoustic echo cancellation in a full-duplex

connection between two telephone terminals wherein a hands-free or handset terminal is

present at one or both sides of the transmission line. Particularly, the invention relates to the

adaptation control of the digital adaptive filter that cancels by subtraction the acoustic echo

signal.

Figure 1 illustrates the acoustic echo generation phenomenon and the acoustic echo

cancellation scheme, performed with a digital adaptive filter, in one side of the transmission

line, which in the present patent is called "near-end side" or "near-end terminal". The digital audio signal x(n) , which comes from the other telephone terminal ("far-end terminal") through

the receive line (RX LINE) of the near-end terminal, is first converted in an analog signal by

means of device 10 (in fig. 1), and then it is amplified and radiated by the loudspeaker 11 (in

fig.1). The acoustic coupling between the loudspeaker and the microphone 12 (in fig. 1) originates the acoustic echo c(t) that is added to the local speech signal s(t) and to the local

noise v(t) . In the transmission line (TX LINE) these analog signals are converted into the

digital signals c(n) , s(n) and v(n) , respectively, by means of the analog-to-digital converter

13 (in fig. 1). The echo cancellation filter 14 (in fig. 1) is a digital adaptive filter with finite impulse response (here called W(n) ) which estimates the acoustic echo c(n) in order to

cancel it by subtraction. Therefore, the output signal of the cancellation filter, y(n) , is

subtracted from the microphone signal d(ή) = s(n) + v(n) + c(ή) . When y(n) is equal or

almost equal to c(n) the acoustic echo signal is removed from the transmitted 'signal e(n) : e(n) = d(n)- y(n) = s(n) + v(n) , [1]

where e(n) is generally called the "estimation error signal."

The acoustic echo estimation process can be viewed as a system identification problem of the

Loudspeaker-Enclosure-Microphone (LEM) system. The LEM system is composed by the

components enclosed in the dashed box of Figure 1.

By modeling the LEM impulse response with a FIR linear filter H(n) of memory length N ,

i.e. by assuming c(n) = ∑ H(n)[k] x(n - k) , [2]

where H(n)[k] indicates the k -th filter coefficient, the acoustic echo is canceled when y(ή) = ∑^ W(n)[Jc] x(n - k) ≡ ^ H(n)[k] x(n - k) , [3]

i.e. when W(ή) approximates H(ή) . Because of the time-varying characteristics of the LEM

system, the filter W(n) has to be continuously updated with an adaptive algorithm in order to

converge to H(ή) . The following general adaptation rule is applied to the digital adaptive filter

14 (in fig. 1) during the far-end speech ( x(n) ) activity periods:

W(n + ϊ) = W(n) + μ(n) - Φ(X(n),E(n)) , [4]

where Φ(y) indicates the generic adaptation rule as, for example, that of the ΝLMS

(Normalized Least Mean Square) algorithm, of the DLMS (Decorrelated Least Mean Square) algorithm or of a APA (Affine Projection Algorithm) of any order, μ(n) is the so-called "step-

size of adaptation", which can also be a time-varying step-size, and X(n) and E(n) are

vectors that collect a sufficient number of successive samples of x( ) and e(n) .

The step-size μ(n) controls the adaptation speed of the cancellation filter W(ή) . The theory

of the adaptive filters establishes that the optimal value of μ(ή) depends on the relationship

between the useful component and the disturbance components of d(n) from the adaptation

viewpoint. From Figure 1 it is evident that c(n) is the useful component of d(n) for the adaptation of W( ) , while v(n) and s(n) represent the disturbance. Moreover, the optimal

value of μ(n) depends also on the convergence state of the filter.

The implementation of an efficient control strategy for the adaptation of the echo canceller in

hands-free and handset telephony is influenced by several environmental factors and it is

affected by the different situations that may appear during the bi-directional full-duplex

conversation:

1 ) Double-talk situation: the far-end speaker and the local speaker are active at the same time (s(n) ≠ 0, c(n) ≠ 0).

2) Echo-path change situation: a sudden variation of the power or of the characteristics of the acoustic echo signal c( ) .

3) The presence of a background noise v(ή) with an unknown and time-varying level.

4) An unknown acoustic coupling between the loudspeaker and the microphone of the

local terminal. (Typically, the acoustic coupling is measured with the mean power rate of c(ή) and x(ή) expressed in dBs).

Description of Background Art

The problem of the adaptation control for acoustic echo cancellation devices has attracted the

interest of researchers especially in the last years. The references that appeared in literature

and in patents deal mainly with the problem of double-talk detection.

The double-talk detectors and more generally the local speech activity detectors may be

subdivided into two categories. The first category comprises all those methods that detect the

start and end of the local speech activity and that generate a two level output indicating the

presence or absence of the local speech. The second category comprises all those methods

that, without taking an explicit decision about the presence of the local speech, control the echo canceller adaptation by appropriately decreasing the step-size during the local speech

activity.

The first category is well represented by the "power comparison methods". In these methods

the output of the double-talk detector derives from some power estimates of the signals

involved in the echo cancellation process. This kind of detector decides about the presence or

absence of double-talk and in case of double-talk detection, it typically inhibits the filter adaptation by setting μ(n) = 0. On the contrary, when no local speech is detected the step-

size μ(n) is set to some predefined value or is computed from some law that manages the

remaining conversation situations.

The principal limitations of the power comparisons schemes are as follows.

1. When the ratio between acoustic echo (c(n) ) power and the local speech (s(n) )

power decreases, the detection reliability also decreases. Generally, these methods can not be applied when the power of c(ή) is equal to or greater than the power of s(n) .

2. These methods require a more or less accurate estimate of the acoustic coupling

between the loudspeaker and the microphone. This estimate is easily known in

handset terminals but it is often unknown with hands-free devices. 3. The local noise v(n) decreases the detection reliability.

Moreover, since the method decides drastically (an ON/OFF decision) about the presence or

absence of the local speech, false or missed detections are often experienced during

operation. The decision is frequently based on a threshold, variable or fixed, and this threshold

critically depends on the acoustic coupling. On the other hand, the detectors based on power

comparisons provide a remarkable simplicity and low computational complexity and, for these

reasons, they are widely employed in handset terminals, wherein the acoustic echo has a moderate power, the acoustic coupling is known and the background noise is limited.

Recently, it was patented a sophisticated and computational intensive method that can be

classified among this category of detectors and that may detect double-talk also in presence of

an acoustic echo to local speech ratio of 25 dB (U.S. Pat. No. 6,049,606). This method assumes that the received signal x(n) , and thus also c(n) , has a 4kHz frequency band (a

8kHz sampling frequency), while the microphone signal is sampled with a double frequency (a

16kHz sampling frequency), which allows power comparisons between 4 and 8kHz, where c( ) is absent.

To the first category of detectors belong also the cross correlation techniques, which extend

the power comparison scheme by adding to the adaptation control a cross-correlation criterion

between various signals. Examples of this type of double talk detectors may be found in the

papers "An Adaptation Control for Acoustic Echo Cancelers", by P. Heitkamper, IEEE Signal

Processing Letters, V0L.4, NO.6, June 1997, and "A Robust Echo Canceler for Acoustic

Environments", by John F. Doherty, IEEE Trans, on Circuits and Systems-ll: Analog and

Digital Signal Processing, VOL.44, NO.5., May 1997, and also in the U.S. Pat. Nos.

5,646,990. Nevertheless, these techniques are affected by a high computational complexity

and a high memory occupation.

The detection methods of the second category typically employ a control law that

approximates the optimal step-size. Different well-known techniques are based on the

following definition of optimal step-size:

„_{oPt (}„, E[(c ((nn))-- .y(n))²]

[5] E[e(n)²]

By equation [5], the optimal step-size control reduces to an adequate estimate of the mathematical expectation of the unknown term c(n)- y(ή) . Among the most popular methods we can cite the "delay coefficients" technique, which however is unable to

discriminate between a double-talk and an echo-path change. We must remark that the

variable step-size double-talk detectors generally modify gradually the echo canceller

adaptation in relation to any disturbing input signal, and thus also in relation to the local background noise v(n) .

The detection techniques for echo changes (i.e. for echo-path changes and for echo power

level changes) have been scarcely researched and documented. These techniques are

necessary mainly in the hands-free devices in order to guarantee a fast re-adaptation of the

digital filter. Indeed, when an echo change is detected the adaptation control increases the

step-size value. Unfortunately, it is difficult to discriminate between the echo change and

double-talk situations; in the first case the step-size should be increased, while in the latter

case it should be reduced. Different methods for discriminating between double-talk and echo

changes can be found in literature. These methods are based on cross-correlation measures

or on probabilistic measures, but they typically have a high algorithmic and computational

complexity.

While the problem of the adaptation control for different background noise v( ) conditions is

often implicitly solved by the variable step-size methods, this problem has to be faced with a

dedicated device in the two-state double-talk detectors. For example, if no double-talk is

deemed to exist, an appropriate value of the step-size of adaptation can be determined from

the long-term power estimate of the microphone signal. However, it remains difficult to discriminate between the local speech s(n) and the local noise v(n) , i.e. it is difficult to

detect double-talk situation especially in very noisy environments.

Summary of the Invention

The proposed control method is based on a variable step-size of adaptation μ(n) that comprises a double-talk detector, an echo change detector and an automatic control for the

background noise level.

The method exploits the delay of the echo signal c(ή) on the TX line with regards to the far-

end speech signal x(n) . The delay is introduced both by the analog-to-digital and digital-to-

analog conversions (blocks 13 and 10 of Fig. 1 , respectively) and by the flying time of the

sound from the loudspeaker to the microphone.

Under this assumption, we can compute the filter response y(n) by employing the delayed

signal x(n - NT) as input of the echo cancellation filter 14 (in fig. 1 ), which in this contest has

memory length N - NE , with NE less or equal to the delay of c(n) when compared to

x( ) . In any case, an artificial delay can be introduced in the TX line immediately after the

converter 13 (in fig. 1) in order to guarantee the desired minimum delay of the signal. The invention is based on two supplementary filters with relatively short length NE (e.g.

NE = 8 samples) that support the echo cancellation filter 14 (in fig. 1) and that, for the role

they play, have been placed in the control block 15 (in fig. 1). The supplementary filters allow

the calculation of two indicators, one that estimates the local speech activity and the other that

estimates the convergence state of the filter 14 (in fig. 1) and, thus, detects echo variations. These filters are not directly applied in the computation of y(n) . Both filters process the most

NT recent samples of x(n) and they have to be adapted with a particular criterion in order to

effectively distinguish between double-talk and echo changes. J^"he indicators, which can be

estimated at any input signal sample, are employed in the formulation of the variable step-size

that controls the adaptation of the cancellation filter 14 (in fig. 1).

In what follows, we describe how the adaptation equations of the three adaptive filters (the

cancellation filter and the two supplementary filters) can be derived in the case of the ΝLMS

algorithm. Let W_B (ή)[k] be the k -th coefficient of the echo cancellation filter of length N - NE at time

n ≥ 0 and let

y_B(«) = -x(n-NT -k) [6]

be the response of the same filter at time n . Let W_A(n)[k] be the k -th coefficient of the first supplementary filter of length NE , which is

employed for echo change monitoring, and let

NT -l y_A(n) = ∑W_A(n)[k] -x(n-k) [7]

be the response of the same filter at time n .

Let W_c(n)[k] be the k -th coefficient of the second supplementary filter of length NE , which

is employed for local speech monitoring, and let

be the response of the same filter at time n .

In order to derive the adaptation rule of the three filters, let us first combine the filters W_A (ή)

and W_B( ) in a single filter W(n) :

W(n) = {W_A(n)[0],...,W_A(n)[NT -l],W_B(n)[0],...W_B(n)[N -NT -ϊ }^τ , [9]

where the notation (-)^r indicates the transposition operation. For the definition of the filter

responses y_A(n) and y_B(n) in equation [7] and [8], respectively, W(ή) provides the

following response to the input signal x(ή) :

N-\ y(n) = W(n)[k]-x(n-k) = y_A(n) +y_B(n) [10]

.1=0 According to the ΝLMS algorithm, the adaptation rule of the filter W(ή) of length N is given

by the following expression μ(n)[k]

W(n+l)[k] =W(n)[k]+^^^-(d(n)-y(n))-x(n-k), k = 0,...,N-l, [11] X(n)^TX(n)

where X (n) = {x(n), x(n -l),...x(n-N + ϊ)}^τ [12]

is the vector that collects the N most recent samples of x(n) and μ(n)[k] is the step-size

of adaptation that can vary at any sample and for any coefficient rc)[&] .

By dividing the filter W(n) in the two parts W_A(n)[k] and W_B(n)[k] and by employing the

definition of y(n) of equation [10], the expression in [11] can be written as follows:

W_A(n+l)[k] =W_A(n)[k]+ ^μ^ .(d(n)-y_A(n)-y_B(n))-x(n-k), k =0,...,NT-l [13]

X(n) X(n)

[14] where we have imposed: μ_A(n) = μ(n)[0] ≡ μ(n)[ϊ\ ≡ ... ≡ μ(n)[NT -l] [15] μ_B(n) = μ(n)[NT] ≡ μ(n)[NT + l] ≡ ... ≡ μ(n)[N - ϊ\ [16]

Equation [14] can be simplified by exploiting the assumption that the echo signal c(ή) , which

is the reference signal for the cancellation filter adaptation, is a delayed version of x(ή) , with

delay of at least NT samples. Thus, at any sample time the optimal value of W_A(n) from the

echo cancellation viewpoint is the null filter. Therefore, by setting the initial conditions W_A(0) = {0,...0}^r and μ_A(n) = 0 , the contribute of y_A(n) to the resulting output y(n) is

null at any time instant and equation [14] may be rewritten as:

W_B(n+l)[k]=W_B(n)[k]+ ^"J -(_.d(n)-y_B(n)yx(n-NT-k), k = 0,...,N-NT-l

X(n) X( )

[17] Note that, under this assumption, the echo cancellation filter adaptation in [17] does not depend on the response of the W _A( ) filter and that the output of the overall system is give

by y(n) = y_B(n), n≥O. [18] With equation [13], we have obtained an exact expression for the adaptation of the W_A(ή)

filter that depends from the response of the overall system W(ή) . Because of the

independence of the cancellation filter W_B(ή) from W_A(ή) , we can adapt W_A(ή) with any

adaptation step-size, without affecting the behavior of the echo cancellation filter. Moreover, due to the independence of W_B(ή) from W_A(ή) , more than one supplementary

filter similar to W_A (ή) can be introduced. Therefore, we have applied again the derivation procedure for obtaining the adaptation expression of the second supplementary filter W_c (n) .

In conclusion, the adaptation of the three filters with the NLMS algorithm can be computed

with the following vector equations

W_A(n + l) = W_A(n) + - μ-_AA(»ⁿ)⁾

X(n ,)\^TTXΓ(/n- ) ^•e_A(ⁿ -^χ _A(ⁿ [19]

W_B(n + l) = W_B(n) - X_B(n) [20]

W_c (n + 1) = W_c (n) + ^μ^ -e_c(n) - X_c (n) [21] X(n) X(n)

where the error signal are defined as: e_B(n) = d(ή) - y_B(n) [22]

e_A(n) = d(n) - y_A (n) - y_B (ή) = e_B (n) - y_A (n) [23] e_c (n) = d(n)- y_c (n) - y_B (n) = e_B (n) - y_c (n) [24]

and the collections of the x(ή) signal samples are defined as:

X_A(n) = {x(n),x(n-l),...x(n-NT + l)}^T [25]

X_B(n) = {x(n-NT),x(n~NT-l),...x(n-N + l)}^T [26] X_c(n) = X_A(n) . [27]

Moreover, we remark that the only filter W_B (n) is employed in the computation of the output

signal y(n) of the filter 14 (in fig. 1 ) and therefore, with reference to Figure 1 , it is y(n) ≡ y_B(n) [28] e(n) ≡ e_B (n) . [29]

We have derived the NLMS adaptation equations of the three filters. In what follows we will introduce the different adaptation modes by defining the adaptation step-sizes μ_A (ή) , μ_B(n) and μ_c( ) .

As we already mentioned, we want to employ the W_c(ή) filter for monitoring the local speech

activity. By using in equation [21] a step-size μ_c(ή) constant and with a relatively high value

(e.g. μ_c (n) = μ_c = 1 ), during the double-talk situations the coefficients of the filter disadapt

rapidly and they assume high absolute values when the local speech power is high. On the

contrary, these coefficients readapt rapidly to values close to zero when the reference signal d(ή) is constituted by the only acoustic echo signal. This behavior suggests adopting as indicator of the local speech activity a norm of the W_c (n) filter:

I NT

IndC(n) = — -E[∑\ W_c(n)[k] , [30] l k=o where q defines the norm type, E[-] indicates the mathematical expectation and | • | indicates

the absolute value operation. The mathematical expectation can be estimated with a temporal average. IndC(ή) is a continuous function of time which is characterized by sharp peeks in

case of local speech activity and by increasing values in case of an increasing power level of

the local background noise.

The experimental test has shown that sudden and strong echo variation affects also IndC(n) . In fact, after an echo variation, also in the hypothesis that the W_c («)[&] filter

coefficients are close to zero, the W_B(n) filter is necessarily disadapted and this situation

determines an increase of the estimation error e(n) ≡ e_B (ή) . By equation [23] and [24] the disadaptation effect is reflected also on the error e_c(ή) and e_A(n) and, by equations [19]

and [21], it is distributed in the successive adaptation steps in the coefficients of the filters W_A(n) and W_c(n) . This behavior suggests the adoption of the filter W_A(n) for monitoring the convergence state of the cancellation filter (and thus for monitoring the echo variations) by

introducing the following indicator:

I NT

IndA( ) = — - E[∑\W_A(n)[k] |^p] , [31] l i=o and by adapting the filter W_A (n) with the same control rule employed for the cancellation filter

W_B(n)

_Ju _k(n) = μ_B(?z)- tt ≥ O, [32]

or with μ_A (n) = max(μ_B (n), μ_{A n} ), n ≥ 0, [33] where μ_{A πύn} is a appropriate little positive constant which allows a fast response to sudden

echo-path changes.

In this way, any disadaptation of the W_B(n) filter determines an increment of IndA(n) that

disappears as soon as the W_B (n) filter readapts.

The two indicators just derived are employed in the adaptation control of the echo cancellation

filter, which is given by the following equation:

(δ + A - IndA(n)) - Px(n) μ_B ⁽n⁾ = (δr + . A . - I ,nd .,A.(;n!)) ■ Px( ,n) + (1 +^j C • I :nd XC(n)) ■ P ^e( ,n) , .

where δ , A and C are some constant parameters that have to be appropriately chosen, Pe(n) is

Ee(n) = max(Ee(n),Ee_min) , [35]

Px(n) and Pe(n) are the powers of the x(n) and e(ή) signals estimated with time

averages and Pe_πάn is some little positive constant that limits the increment of μ_B( ) during

the near-end speech silence periods.

We can easily understand the behavior of [34]. The indicator IndC(n) amplifies the effect of

the denominator term Pe(n) that, in presence of a local speech activity, reduces the adaptation step-size. Furthermore, an increase in the indicator IndA(ή) , which appears in

case of a cancellation filter disadaptation due to a variation of the characteristics of c(n) ,

tends to increase the adaptation constant. The system strength is the good decoupling that

exists between the two indicators, i.e. the high sensibility of the first indicator to the echo

variations with little influence from the local speech activity and the high sensibility of the

second indicator to the local speech with minimal affection of short duration for the echo

changes.

The short duration of the echo change influence on IndC(n) is guaranteed by the high adaptation step-size constant μ_c(n) that, after an initial transitory period, readapts rapidly

the coefficients towards some values close to zero. We can emphasize the discrimination of the two situations by appropriately choosing the parameters p and q in equation [31] and

[30] respectively: e.g. by using an Euclidean norm for IndA(n) , with p = 2 , and with a linear

norm for IndC( ) , with q = 1 .

The fast detection of local speech activity depends from the duration of the time window used for the power Pe( ) estimation. The experimental results suggest the use of short duration

window (for instance 4 milliseconds). In this way, the control procedure is able to perform

adaptation also during the short periods of low near-end speech power, i.e. it is able to adapt

the filter also during the double-talk.

The adaptation algorithm may be further simplified for the NLMS algorithm by estimating Px(n) on a window of length N ,

N where X(n) is given by equation [12]. In this way we simplify the denominator term of the

ratio μ_B( )/X(n)^τ X( ) in equation [20] (and thus in the ratio μ_A(n)/X(n)^τ X(n) in

equation [19]). The set of equations [6], [7], [8], [12] and of the equations from [19] to [35] does not introduce

any novel adaptation algorithm for the acoustic echo cancellation; on the contrary it introduces

an innovative method for controlling a generic adaptive filter. Indeed, the adaptation rule of the

cancellation filter [20] and the transferring function of the same filter [6] possess the

characteristics of a classical NLMS adaptive filer, apart for the normalization term X(n)^TX(n) that is estimated on N observations of x(n) rather than on N-NT

observations. Nothing prevents from substituting in equation [20] the normalization term computed on N samples with that computed on N-NT samples, i.e. X _B(n)^τ X _B(n) .

Nevertheless this modification does not influence the echo control performance while, on the

contrary, it causes a potential increase in the computational complexity and in the memory

occupation of the system.

The variable step-size of equation [34] can be interpreted as an estimate of the optimal step-

size for the NLMS algorithm, reported in equation [37],

By assuming the residual echo (c(n) - y(n)) to be uncorrelated with the local disturbance

s(n) and v(n) , we obtain the expression of equation [38],

E[{c(n) - y(n) ] ^μ°^pt E[(c(n) - y(n)f]₊ E[(s(n) ₊ v(n)f ] ^[∞i

Let us assume in equation [34] that <5 = 0. Then this equation can be written as follows N - lndA(n) - Px(n) ^μB N - IndA(n) - Px(n) + N - (l + C - IndC(n)) - Pe(n)/A When p = 2 , N ^• IndA(n) • Px(n) is the estimate of the undisturbed error signal

E[(c(n) - y(n))²] , while N JJ + C • IndC(n)) - Pe(n)l A is an estimate of the local

disturbance power E[(s(n) + v(n))²] . The parameters C and A have to be properly

chosen in order to match N - (l + C - IndC(n)) - Pe(ή)/ A and E[(s(n) + v(ή))²) for most operating conditions.

Although a particular embodiment of the invention has been illustrated and described it is

apparent that various changes can be introduced. For example, the control method described

by the set of equations ranging from equation [16] to [35] can be employed with any

adaptation algorithm, as the DLMS, the APA of any order or the RLS. The adaptation

equations [19], [20] and [21] can be derived with the same procedure described by equations

[6] to [18], while the variable step-size of adaptation of equation [34] is directly applicable.

Unlike many methods described in literature, the proposed control method does not

necessitate of any estimation of the acoustic coupling level. Moreover, it is able to readapt to

sudden echo level changes of 20dB or more in about 0.5 seconds.

The proposed control method provides also a double-talk detection technique that is reliable in

any situation, also in case of extreme local noise level, that is often experienced in a car

application. The proposed control method is reliable with any local noise level and with any

echo level; it avoids any annoying distortion or modulation of the local noise and it exploits at

the best the masking effect of local noise on echo.

The proposed control method has been fully tested with different adaptation algorithms. The

complete system has been employed both for echo cancellation in a car application and in a

cellular phone handset application with acoustic coupling levels ranging from -25dB up to

6dB.

Brief Description of the Drawings

Figure 1 is a block diagram of a simplified echo cancellation system illustrating the invention.

Figure 2 illustrates the performances of the invention for a high quality acoustic echo

cancellation application in a car hands-free. Detailed Description of the Drawings

Figure 1 illustrates a simplified block diagram of an echo cancellation system for hands-free or handset telephone terminals. It is shown a receive path (RX LINE) with the digital signal x(n)

that comes from the other telephone terminal ("far-end terminal"), a digital to analog converter

10 (in fig. 1) that drives a power amplifier and diffusion system (a loudspeaker or an ear-piece)

11 (in fig. 1). The acoustic coupling between the loudspeaker and the microphone 12 (in fig. 1) originates the acoustic echo c (t) that is added to the local speech signal s(t) and to the local

noise v(t) . In Figure 1 , it is also shown the transmit path (TX LINE) with the microphone 12

(in fig. 1) and the digital to analog converter 13 (in fig. 1). The echo cancellation system (block

14 and 15 in fig. 1) is connected between the RX and the TX path and is constituted by an

adaptive echo cancellation filter (14 in fig. 1) and the adaptation control (15 in fig. 1) that is the

subject of the present invention. The echo cancellation filter provides an estimate of the echo that is subtracted from d(ή) , the digital version of the microphone signal. The resulting signal

is transmitted to the other party.

Figure 2 illustrates the performances of the proposed control method, applied on a DLMS

adaptation algorithm. The acoustic echo canceller was employed in a car equipped with a HI-

FI amplifier and a 5 loudspeakers system. The acoustic echo, plotted in the first graph from the top of Figure 2, is characterized by amplitudes similar to those of the local signal s(ή) ,

plotted in the second graph. Two sudden echo variations are simulated by reducing after 6.5

seconds the echo power of 20db and by taking it to the initial value after almost 13 seconds. Note in the graph of e(ή) that the echo variation around 6.5 seconds comes during a short

double-talk period. Nevertheless the filter readapts in less than a second. Also the sudden

increase of 20db volume, after 13 seconds, is compensated in less than 1 second. Note moreover that the adaptation step-size μ_B (ή) , which assumes continues values between 0 and 1 , detects promptly the double-talk and it is able to exploit for the adaptation the short periods of low s(n) power (see graph of μ_B (n) after 2 seconds).

Best mode for carrying out the invention

An application where the proposed control method provides, in comparison with the other

known methods, a fair margin of improvement for performances, for functional characteristics

and for computational cost is the acoustic echo cancellation for hands-free car-kit devices. At

present, most of the hand-free devices available on the market are offered with proprietary

amplifiers and loudspeakers.

A significant improvement of these devices in terms of cost, of functionality and of overall size

is given by the possibility of employing in the car an echo canceller implemented in the mobile

phone. The mobile phone could be connected to the sound amplification unit of the car radio

system.

Unlike the cancellation systems with proprietary amplifier and loudspeakers, where the

designer can directly tune the canceller because he knows the output power and the echo

power levels, an acoustic echo cancellation device that takes advantage of an extern

amplification system operates in a completely unknown environment. Furthermore, this

cancellation system is not informed of the sudden volume variation the user may introduce in

the diffusion system. From the echo cancellation viewpoint these volume variations

correspond to echo level variations.

Industrial Applicability

The proposed control procedure can be applied in the acoustic echo control of handset,

headset, vehicle-mounted handsfree, desktop operated handsfree and hand-held handsfree

terminals.

Claims

1. A system for the adaptation control of a digital filter for acoustic echo cancellation in

handset and hands-free telephone terminals, wherein said echo canceller, defined by

equation [6], is controlled by two supplementary filters of relatively short length, defined by

equations [7] and [8], wherein the first said supplementary filter is employed for the echo

change estimation and the second said supplementary filter is employed for the local

speech activity estimation.

2. A system for the adaptation control of a digital filter for acoustic echo cancellation as

defined in claim 1 , wherein the control procedure for the adaptation of the echo canceller

14 (in fig. 1) is computed by means of equation [34].

3. A system for the adaptation control of a digital filter for acoustic echo cancellation as

defined in claim 1 , wherein said echo cancellation filter is controlled by two supplementary

filters included in the control block 15 (in fig. 1).

4. A system for the adaptation control of a digital filter for acoustic echo cancellation as

defined in claim 3, wherein the first of the two supplementary filters is suitable for echo

variation estimation.

5. A system for the adaptation control of a digital filter for acoustic echo cancellation as

defined in claim 3, wherein the first supplementary filter, employed for echo variation

estimation, is adapted with the same adaptation step-size of the echo cancellation filter 14

(in fig. 1).

6. A system for the adaptation control of a digital filter for acoustic echo cancellation as

defined in claim 3, wherein the second of the two supplementary filters is suitable for local speech activity monitoring.

7. A system for the adaptation control of a digital filter for acoustic echo cancellation as defined in claim 3, wherein the second supplementary filter, employed for local speech

activity estimation, is adapted with a fixed and relatively high adaptation step.

8. A system for the adaptation control of a digital filter for acoustic echo cancellation as

defined in claim 3, wherein the indicator employed for the echo variation estimation and

for the echo canceller state monitoring is computed by means of equation [31] from the

first supplementary filter of equation [7].

9. A system for the adaptation control of a digital filter for acoustic echo cancellation as

defined in claim 3, wherein the indicator employed for the local speech activity estimation

is computed by means of equation [30] from the second supplementary filter of equation

[8].

10. A system for the adaptation control of a digital filter for acoustic echo cancellation as

defined in claim 3, wherein the supplementary filters contributes to adaptation of the

cancellation filter 14 (in fig. 1) only through the variable step-size of adaptation.

11. A system for the adaptation control of a digital filter for acoustic echo cancellation as

defined in claim 1 , wherein the echo cancellation filter [6] and the two supplementary

filters [7] and [8] are adapted with equations [18]-[27] with the NLMS algorithm or with

similar equations with any other adaptation algorithm.