MULTI-CHANNEL ECHO CHANCELLATION WITH ADAPTIVE FILTERS HAVING SELECTABLE COEFFICIENT VECTORS

_!_ 2088558 1 TITLE OF THE INVENTION 2 "Multi-Channel Echo Cancellation With Adaptive Filters Having Selectable 3 Coefficient Vectors" 4 BACKGROUND OF THE INVENTION Field of the Invention 6 The present invention relates generally to echo cancellers, and more 7 specifically to a multi-channel echo canceller for a teleconferencing 8 system and the like. 9 Description of the Related Art A multi-channel echo canceller is described in a paper "Compact 11 Multi-channel Echo Canceller with a Single Adaptive Filter Per Channel", 12 Akihiro Hirano et al, Proceedings of the 1992 IEEE International 13 Symposium on Circuits and Systems, San Diego, California, May 10-13, 14 1992. This paper addresses to the convergence problem that is associated with adaptive filter coefficients when received signals have 16 strong cross-correlation. The known echo canceller is provided with a 17 single set of filter coefficients to reduce acoustic echoes at the near and 1 8 far ends of the system as well as transmission echoes by taking into 19 account all possible microphone-loudspeaker combinations at the near end and all possible acoustic paths from a talker to microphones at the 2 1 far end. The known echo canceller estimates an inter-channel time 2 2 difference between the propagation delays of the received signals and 2 3 couples one of the signals having a smaller propagation delay to the 24 adaptive filters to produce an echo replica from each adaptive filter. The 2 5 echo replica is subtracted from a corresponding signal to cancel the echo 2 6 contained therein. The adaptive filters are controlled with residual 27 echoes so that they reduce to a minimum. The filter coefficients of each 2 8 adaptive filter define a transfer function that is converged to an optimum 2 9 value for a particular talker at the far end in response to the selected 3 0 signal. Therefore, the transfer function of each adaptive filter tends to offset from the optimum value in response to a talker's movement. However, there is a noticeable amount of delays in the adaptive filters for adapting to the changing acoustic parameters at the far end. SUMMARY OF THE INVENTION It is therefore an object of the present invention to provide a multi-channel echo canceller that can quickly adapt itself to changing acoustic parameters at the far end of a communication channel. According to the present invention, there is provided an echo canceller for a teleconferencing system having a plurality of transmit and receive channels inter¬ connecting separated conference rooms, each of the conference rooms having a set of microphones and a set of loudspeakers. The echo canceller is located at each conference room and comprises a plurality of subtracters connected respectively in the transmit channels for respectively receiving a transmit signal from the respective microphone and cancelling an echo contained in the transmit signal with a cancelling signal, and a plurality of adaptive filters associated respectively with the subtracters. Each adaptive filter has a plurality of vectors of filter coefficients. A time difference between propagation delays of distant signals is estimated, and a distant signal having the largest content of echo components or a minimum propagation delay is applied to the adaptive filters and one of the vectors is identified according to the estimated time difference and the selected signal. Each adaptive filter varies its filter coefficients of the identified vector with a correction term proportional to the output of the associated subtracter and filters the selected signal using the coefficients of the identified vector to derive an echo replica, which is supplied to the associated subtracter as the cancelling signal. In response to the talker's movement at the far end of the communication channel, the filter coefficients of each adaptive filter are quickly switched from one vector to another. Therefore, the echo canceller of this invention can quickly adapt itself to changes in the far-end acoustic parameters. In accordance with the present invention, there is further provided an echo canceller for use in a conference room having at least two microphones and at least two loud¬ speakers connected by at least two transmit channels and at least two receive channels to corresponding loudspeakers and microphones of a distant conference room, comprising at least two subtracters for respectively receiving transmit signals from the microphones and cancelling an echo contained in each of the received transmit signals with a cancelling signal; inter-channel time difference estimator means for receiving distant signals from said receive channels and estimating a propagation time difference between the received distant signals to produce time-difference estimate signal; control means responsive to the time-difference estimate signal for selecting one of the received distant signals having a largest - 3a - content of echo components and identifying one of a plurality of vectors; and at least two adaptive filters associated respectively with said subtracters and connected to said control means for receiving said selected distant signal, each of said adaptive filters having a plurality of said vectors each comprising a set of filter coefficients, each of the adaptive filters selecting the filter coefficients of the vector identified by said control means for filtering said selected signal to derive an echo replica and supplying the echo replica to the associated subtracter as said cancelling signal, the filter coefficients of said identified vector of each adaptive filter being variable according to an output signal from the associated subtracter. In accordance with the present invention, there is further provided a method for operating an echo canceller located in a conference room having at least two microphones and at least two loudspeakers connected by at least two transmit channels and at least two receive channels to corres¬ ponding loudspeakers and microphones of a distant conference room, the echo canceller comprising at least two subtracters connected respectively in said transmit channels for respect¬ ively receiving transmit signals from the microphones of the conference room and cancelling an echo contained in each of the transmit signals with a cancelling signal, and at least two adaptive filters associated respectively with said subtracters, each of said adaptive filters having a plurality of vectors each comprising a set of filter coefficients, the - 3b - filter coefficients of each vector being variable according to an output signal from the associated subtracter, each of said adaptive filters filtering an input signal applied thereto using one of said vectors to derive an echo replica and supplying the echo replica to the associated subtracter as said cancelling signal, the method comprising: a) receiving distant signals from said receive channels and estimating a propagation time difference between the received distant signals to produce a time-difference estimate signal; b) selecting one of said distant signals from said receive channels having a largest content of echo components according to the time-difference estimate signal and supplying the selected distant signal to said adaptive filters as said input signal; and c) identifying one of said vectors of said adaptive filters according to said time-difference estimate signal and said selected distant signal and causing each of said adaptive filters to use the identified vector. BRIEF DESCRIPTION OF THE DRAWINGS The present invention will be described in further 2 0 detail with reference to the accompanying drawings, in which: Fig. 1 is a block diagram of a teleconferencing system incorporating multi-channel echo cancellers of the present invention; Fig. 2 is a block diagram of an inter-channel time difference estimators unit of a multi-channel echo canceller according to a first embodiment of the present invention; Fig. 3 is a block diagram of a controller of the echo P 71024-225 _3c- 2088558 canceller; Fig. 4 is a block diagram of a variable coefficient adaptive filter; Figs. 5A and 5B are flowcharts describing a sequence of operations performed by the calculator of the adaptive filter for generating an echo replica; Fig. 6A and 6B are flowcharts describing a sequence of operations performed by the calculator of the adaptive filter for updating filter coefficients; Fig. 7 is a block diagram of an inter-channel time difference estimators unit according to a second embodiment of the present invention; Fig. 8 is a block diagram of the controller of the echo canceller according to a modified embodiment of the present invention; and Fig. 9 is a timing diagram associated with the modification of Fig. 8. DETAILED DESCRIPTION Referring now to Fig. 1, there is shown a telecon- 2 0 ferencing system embodying a multi-channel echo canceller of a first embodiment of the present invention. The teleconfer¬ encing system comprises a plurality of loudspeakers 11~1M and microphones 21~2M positioned in a conference room A and a like plurality of loudspeakers 1' ~1'M and microphones 2'1~2'u positioned in a distant conference room B. The loudspeakers l A NE-498 1 and microphones 2 are formed into a plurality of sets, or channels 2 corresponding respectively to a plurality of sets of loudspeakers 1 ' and 3 microphones 2'. The speaker and microphone of each set in conference 4 room A are connected to a multi-channel echo canceller 100. The outputs of echo canceller 100 are connected by way of a transmission 6 medium 3 to the corresponding set of microphone and speaker in 7 conference room B, and are further connected to an identical multi- 8 channel echo canceller 100' to which signals from microphones 1' are 9 also applied. The outputs of echo canceller 100'are connected through transmission medium 3 to the speaker/microphone sets of conference 11 room A, and further to the echo canceller 100. Therefore, the input 12 signals to all loudspeakers 1 and 1 ' are supplied to the echo canceller 100 13 and 100', respectively, to form a plurality of closed loops for purposes of 14 cancelling inter-channel echoes. Echo canceller 100 at the site of conference room A comprises an 16 inter-channel time difference estimators unit 101 to which the M receive- 17 channel inputs from conference room B are terminated in parallel to 18 connections to loudspeakers 1] ~1 m• The same receive-channel inputs 19 are also applied to inputs of a channel selector 102. A controller 103 is connected to the outputs of inter-channel time difference estimators unit 21 101 to provide a channel selection signal to selector 102. Variable 22 coefficient adaptive filters 104i~104(vi are associated respectively with 23 subtracters 105i~105m and with the transmit channels or microphones 2 4 2-|~2m to receive a speech signal input from the channel selector 102 to 2 5 generate a signal which is a replica of an echo contained in the signal 2 6 applied to the associated subtracters 105. 27 Subtracters 105i~105|vi are connected in the transmit channels to 2 8 receive speech signals from corresponding microphones 2-\~2m as well 29 as the echo replicas from the associated adaptive filters 104i~104m, 3 0 respectively. The outputs of the subtracters 105i~105m, each containing NE-498 .5. 2088553 1 a speech signal and an undesired residual echo, are transmitted through 2 transmission medium 3 and respectively coupled to loudspeakers I'i-I'm 3 at site B, and returned through echo canceller 100' to site A and applied 4 to the inter-channel time difference estimators unit 101 and channel selector 102. The outputs of the subtracters 105]~105|vi are also applied 6 to the corresponding adaptive filters 104i~104m as feedback signals to 7 adaptively control their filter coefficients. 8 Each adaptive filter 104 has a plurality of vectors of filter coefficients. 9 In each adaptive filter, one of its filter coefficient vectors is selected in response to a coefficient vector select command from controller 103. 11 Controller 103 further applies a coefficient update enable or disable 12 command to each adaptive filter 104. 13 As shown in Fig. 2, the inter-channel time difference estimators unit 14 101 comprises (M - 1) inter-channel time difference estimators 200i~200m_i, each being connected to receive speech signals from a 16 corresponding pair of adjacent receive channels. The function of each 17 inter-channel time difference estimator 200; is to estimate the 1 8 propagation delay time difference between signals transmitted through 1 9 the /-th and (/' + 7)th receive channels of the corresponding pair. The 2 0 estimation of the propagation time difference between the signals of 21 adjacent channels is achieved by the use of two adaptive filters 22 associated respectively with the adjacent channels. The adaptive filter is 2 3 typically of a tapped-delay line filter structure that consists of a set of 24 delay elements, a set of multipliers connected to the delay-line taps, a corresponding set of adjustable tap weights and a summer for adding 2 6 the multiplier outputs. It involves detecting the difference between the 2 7 output of each adaptive filter and the input of the other adaptive filter to 2 8 produce an error signal, adjusting the tap weights of each adaptive filter 2 9 in a recursive manner by updating the present estimate of each tap 3 0 weight with a correction term proportional to the error signal at that time, NE-498 .6. 208855S 1 and detecting a maximum value of the tap weights (or filter coefficients) 2 of the adaptive filters. If an adaptive filter receives a signal of shorter 3 propagation delay and produces an estimate of a signal of longer 4 propagation delay, one of its filter coefficients corresponds to a delay closest to the wanted difference and such a filter coefficient has a 6 maximum absolute value. Conversely, if it receives a signal of longer 7 propagation delay and produces an estimate of a signal of shorter 8 propagation delay, its filter coefficients have lower absolute values. 9 To this end, each inter-channel time difference estimator 200; comprises an adaptive filter 201 having an input port connected to the 11 /-th receive channel and L taps respectively connected to absolute-value 12 conversion circuits 202. The output port of the adaptive filter 201 is 13 connected to the negative input terminal of a subtracter 203 whose 14 positive input terminal is connected to the (/' + 7)th receive channel. The output of the subtracter 203 is applied to the control port of the adaptive 16 filter 201. The outputs of all absolute-value conversion circuits 202 are 17 supplied to a decision circuit or maximum value detector 207 where a 18 maximum of the input absolute values is detected. In like manner, the (/ 19 + J)th receive channel is connected to the input of an adaptive filter 204 whose L taps are respectively connected to absolute-value conversion 21 circuits 205 whose outputs are connected to the maximum value detector 22 207. The output port of adaptive filter 204 is applied to the negative 2 3 input of a subtracter 206 whose positive input is connected to the /'-th 24 receive channel. Each of the adaptive filters 201 and 204 is controlled by the output of the corresponding one of the subtracters 203 and 206 so 2 6 that the output of each subtracter is reduced to a minimum. 27 Using the detected maximum absolute value, the maximum value 28 detector 207 of each inter-channel time difference estimator 200j 2 9 produces an output which represents an estimate of the propagation 3 0 time difference between the signals of the /'-th and (/' + 7)th channels. In NE-498 _7. 2088558 1 this way, a set of (M - 1) propagation time differences are produced and 2 applied to the controller 103. 3 As illustrated in Fig. 3, the controller 103 includes a minimum 4 propagation delay detector 300 which receives the outputs of inter- channel time difference estimators unit 101 to detect a minimum time 6 difference value, selects one of the signals on the receive channels having 7 a minimum propagation delay, and supplies a channel select signal to 8 the channel selector 102 to cause it to pass the selected signal to all 9 adaptive filters 104 as a signal having a largest content of echo components. 11 The channel selection signal is applied to a coefficient vector selector 12 301 to which the outputs of inter-channel time difference estimators unit 13 101 are also applied. Using the channel selection signal as an 14 identification of the detected minimum propagation delay and the time difference estimates, coefficient vector selector 301 determines the 16 amount of time by which each of the receive signals is delayed with 17 respect to the signal having the minimum propagation delay by 18 calculating the following Equations (1) and (2). 19 If the time difference between the /-th signal and any of arbitrarily chosen /-th receive signal is denoted as fy;, then the time difference 21 between signals (/- and k-th) of any possible pairs is given by the relation 22 t/£ = tjj + tj. Therefore, if / < /, the time difference tj,j for a set of time 23 differences t/, tj, rj,4, , tM-1tM is given by: j 2 4 k=i+l and if / >/, the time difference tj,; is given by: i ti,j= I>*-U (2) 2 6 k=j+l 27 Therefore, if the /-th signal is identified as having a minimum propagation 28 delay time, it can be said without loss of generality that the time difference t. . assumes one of integers 0,1, t™™ in the 1, n iLicix M discrete time domain. Since there is a maximum of (t +1) max permutations if M integers are selected (provided that repeated selection is allowed), there is a maximum of (t +1) sets of time differences (t. t. _ t. ....,t. M) for a signal having minimum propagation delay. Therefore, each variable coefficient adaptive filter 104. (where M i=l,2, M, Fig. 1) has (t x+l) vectors of filter coefficients corresponding respectively to the time differences (t. t. t. _ ...t. M). 1,X, 1 , £. , 1,J, 1,1 Using the output of minimum propagation delay detector 3 00 as a variable i and successively incrementing an input variable j, coefficient vector selector 301 calculates the following Equation: M 1-1 S(t +1)J t. .+1 .max 1,3 O) to select one of the (t +1)M filter coefficient vectors and v max supplies a coefficient vector select command to all adaptive filters 104 -104,,. 1 M Controller 103 further includes an update control circuit 302 to supply an enable command to all adaptive filters to cause them to constantly update their filter coefficients. The enable control circuit 302 is implemented with a register or the like. Details of each variable coefficient adaptive filter 104., Fig. 1, are shown in Fig. 4. The adaptive filter 104 comprises a calculator 400 for performing convolution a 71024-225 A. - 8a - calculations, a selector 401, a data memory 402 and a plurality of filter coefficient memories 403 -403 . each storing a vector of filter coefficients. Selector 401 selects one of the filter coefficient memories 4 03 in response to the coefficient vector select command from controller 103 to supply the filter coefficients of the selected vector to the calculator 400. Calculator 400 includes a data address generator 404 for generating address data for the data memory 402 and a coefficient address generator 405 for generating address data for the coefficient memories 403. A controller 406 is connected to a program memory 407 NE-498 1 to provide overall control of the calculator through common bus 414 2 using gates 408, 409, 411, multiplier 410, arithmetic and logic unit (ALU) 3 412 and a register bank 413 of registers R-j-Rj. The channel input from 4 selector 102 and the residual echo signal from subtracter 105j are applied to a common bus 414 from which the echo replica of the 6 adaptive filter is taken. 7 Using the control program stored in memory 407, the controller 406 8 operates according to flowcharts shown in Figs. 5A, 5B, 6A and 6B. 9 In Figs. 5A and 5B, controller 406 is programmed to proceed as follows to generate an echo replica. A sequence of data samples from 11 selector 102 are first read in and temporarily stored into register R-\ of 1 2 register bank 413 (step 500) and the samples so stored in register Ri are 13 then transferred to data memory 402 (step 501 ). Register Ri is then 14 initialized to zero (step 502) and a total tap count of the adaptive filter is stored into register R2 (step 503). Data address generator 404 is set to 16 the first storage location of data memory 402 (step 504) and coefficient 17 address generator 405 is set to the first address location of one of the 18 coefficient memories 403 which is selected by the vector select 19 command from the controller 103 (step 505). Gate 408 is enabled and a data sample is read out of the first storage 21 location of the data memory into multiplier 410 via gate 408 (step 506). 22 Gate 409 is enabled and a filter coefficient is read out of the first storage 2 3 location of the selected coefficient memory 403 into multiplier 410 via 2 4 gate 409 (step 507). Controller 406 enables multiplier 411 to perform 2 5 multiplication operations on the input data (step 508) and the result of the 2 6 multiplication is supplied to ALU 412 via gate 411 (step 509). The 27 contents of register R-\ (which are zero in the first pass of loop operation) 2 8 are read and stored into ALU 412 (step 510) and summed with the result 2 9 of multiplication (step 511). The result of summation is stored into 3 0 register R] (step 512) and the address generators 404 and 405 are NE-498 -10 1 incremented to their next storage location (steps 513, 514). The tap 2 count value is then read out of register R2 and supplied to ALU 412 (step 3 515) and the latter is enabled to subtract one from the tap count (step 4 516), the the result of subtraction being stored back into register R2 (step 517). Control proceeds to decision step 518 to determine whether the 6 tap count value of register R2 is equal to zero. 7 If the answer at step 518 is negative, control returns to step 506 to 8 repeat the process by reading a next data sample and a next filter 9 coefficient (steps 506, 507), multiplying them together (step 508), and updating the previous sum in register Rt by adding to it the current 11 multiplication result (steps 509 to 517). When the tap count is reduced to 12 zero, control branches at step 518 to step 519 to output the total sum 13 from register Rt as an echo replica. 14 In Figs. 6A and 6B, controller 406 is further programmed to proceed as follows to update the filter coefficients of a selected coefficient 16 memory 403 when the calculator 400 is supplied with a coefficient 17 update command from controller 103. Coefficient update execution 1 8 starts with step 600 by storing the total tap count of adaptive filter into 19 register Ri. Address generators 404 and 405 are set to the first location 2 0 of data memory 402 and the first location of the selected coefficient 21 memory 403, respectively, (steps 601, 602). Gate 408 is enabled and an 22 error sample is read in from subtracter 105j and supplied to multiplier 2 3 410 via gate 408 (step 603). A step-size value is read out of program 2 4 memory 407 into multiplier 410 via gate 409 (step 604) and multiplier 410 2 5 is enabled to multiply the step-size value with the error sample (step 2 6 605). The result of the multiplication is stored into register R2 (step 606). 2 7 Control proceeds to step 606 to read a filter coefficient from the 2 8 coefficient memory 403 that is selected by vector select command from 2 9 controller 103 and store it into register R3 as an old estimate of a filter 3 0 coefficient. A data sample is read from memory 402 into multiplier 410 NE-498 .„. 2088558 1 via gate 408 (step 608) and the product of the step-size and error 2 sample, now stored in register R2, is read and applied to multiplier 410 3 via gate 409 (step 609) where it is multiplied with the data sample (step 4 610). The result of the multiplication is supplied via gate 411 to ALU 412 (step 611 ) where it is summed with the filter coefficient stored in register 6 R3 (steps 612, 613), producing an updated, or present estimate of the 7 filter coefficient. The updated estimate is stored back into register R3 8 (step 614) and the old coefficient value in the selected coefficient 9 memory 403 is updated with the present coefficient value stored in register R3 (step 615). Both address generators 404, 405 are then 11 incremented to the next storage location (steps 616, 617) and the tap 12 count is read from register R-| into ALU 412 (step 618) where it is 13 decremented by one (step 619) and stored back into register R] (step 14 620). Control proceeds to decision step 621 to check to see if the tap count is reduced to zero. If is answer is negative, control returns to step 1 6 607 to continue the updating process on subsequent data samples and 17 corresponding filter coefficients until the tap count reduces to zero so that 1 8 the old estimates of filter coefficients are all updated. 19 Consider a case of two channels, a time difference between receive channels #1 and #2, the adaptive filter 104] will produce an echo replica 21 sjfn) as given by Equation (4) as follows: N-l Si(n)= £wi,nd(i>n)-r(n-i) (4) 2 2 (=0 2 3 where, n is the sample point, wipn) is the filter coefficient at tap-point 24 / of a vector of adaptive filter 104i which vector is selected in response to 2 5 an inter-channel time difference nd at sample point n (note that, in the 2 6 case of more than two channels, Equation (3) is used to identify the 2 7 vector to be selected), and r(n - i) is the output of selector 102 at sample 2 8 point (n - i), and N is the total number of taps (tap count) of the adaptive 2 9 filter. Likewise, the adaptive filter 1042 will produce an echo replica S2(n) NE-498 .,2- 2088553 1 as given by Equation (5): N-l S2(n)= Zw2,nd(i,n)-r(n-i) (5) 2 i=o 3 where W2/nct(i,n) is the filter coefficient at tap-point / of a vector of 4 adaptive filter 1042. Old estimates of filter coefficients wj//, and W2/nci(i,n) are 6 updated to new coefficients w1/nd(i,n+1) and W2/ncj(ifn+1 ), respectively, 7 as follows: 8 wi,nd(i>n + l) = wlnd(i,n) +fieI(n)-r(n-i) (6) 9 W2,nd(i,n + l) = w2tnd(i,n) +}ie2(n)-r(n-i) (7) where // is the step-size, ei(n) and e2(n) are error samples from 11 subtracters 105i and 1052, respectively. As the filter coefficients 12 wn(i/(i,n) of each selected vector are adaptively updated with a 1 3 correction term proportional to a corresponding error signal en), each 14 filter coefficient assumes a unique value. 1 5 When a talker moves his position with respect to his microphone, the 1 6 acoustic transfer function of the passage of the speech sound varies 17 accordingly and with it the filter coefficients of each adaptive filter are 1 8 instantly switched from one vector to another. Therefore, the multi- 19 channel echo canceller of this invention can quickly adapt itself to changes in delay time differences resulting from talker's movement. 21 A number of modifications are possible without altering the scope of 2 2 the present invention. Inter-channel time difference estimation is also 2 3 achieved by the use of a cross-correlation technique as follows. 2 4 Assume that data samples xrfn) and X2(n) are received on two 2 5 adjacent receive channels at time n with a time difference m 2 6 therebetween. Cross-correlation function Ru(n,m) for the two data 27 samples is given by: 2 8 Rufam) = Eixtin)- X2(n+m)] (8) 2 9 If X2(n) = xn-rifj), the following relations hold: E[(x1(n)-x2(n+m))2] =E[(x1(n)-x1(n-nd+m))2] =E[x12(n)]-2E[x1(n).x2(n+m)]+E[x22(n+m)] =E[x12(n)]+E[x22(n+m)]-2R12(n/in) (9) where E[.] denotes the expected value. Therefore, cross- correlation function R1-(n,ia) is given by: R12 (n/ia) = (l/2) {E[x12 (n) ]+E[x22 (n+m) ] -EtxCnJ-xn-nm))2]} (10) Since E[x (n)] and E[x2 (n)] are both constant if x (n) and x„(n) are of steady values, the following cross-correlation function R «(n,!!!) can be obtained: R12(n,m)=(l/2){C-E[(x1(n)-x1(n-nd+m))2]} where C=E[x 2(n)]+E[x 2(n+m)]. (11) As a result, R _(n,m) is maximum when m=nd. If data samples x (n) and x (n) are of opposite phase to each other, i.e., x.(n)=-x_(n-n,), the following relations hold: E[(x1(n)-x2(n+m))2] =E[(x1(n)-x1(n-nd+m))2] =E[x12(n)]+E[x22(n)]+2R12(n,m) (12) Thus, R1_(n,m) becomes: R12(n,m)=-(l/2){C-E[(x1(n)-x1(n-nd+m))2]} (13) Therefore, R 2(n,m) is minimum and its absolute value is maximum when nt=nd. The inter-channel time difference m between two signals is estimated by detecting when the cross- correlation function R12(n,m) has a maximum absolute value. Fig. 7 is a block diagram of a modified version of A the inter-channel time difference estimators units 101. It comprises (m-1) inter-channel time difference estimators 700 ~700 associated respectively to a pair of adjacent receive channels #1~#M. Each estimator 700. comprises a cross-correlator 701, a plurality of absolute value conversion circuits 710 and a maximum value detector 711. The cross- correlator 7 01 includes a first tapped-delay line formed by unit delay elements 702 connected in series to a first one of the associated pair of receive channels. A plurality of multipliers 703 are connected to the delay-line taps to respectively multiply the successively delayed data samples with a data sample supplied from the other channel of the associated channel pair. The outputs of multipliers 703 are connected to integrators 704, respectively, whose outputs are fed to corresponding ones of absolute value conversion circuits 710. The outputs of integrators 704 respectively indicate the m elements of a cross-correlation function R. . 1(n,m) between the i-th and (i+l)th channels corresponding respectively to time differences m=l,2,....L, where L is the maximum number of the delay-line taps. A similar set of circuits are provided for the other channel. Unit delay elements 705 are connected in series to the other channel to form a second tapped delay line. Multipliers 706 are connected to the second delay-line taps to respectively multiply the successively delayed data samples with a data sample supplied from the first channel. The outputs of multipliers 706 are connected to integrators 707, -«a- 2088558 respectively, whose outputs are fed to corresponding ones of absolute value conversion circuits 710. The outputs of integrators 707 respectively indicate the m elements of the cross-correlation function R. . .(n,m) corresponding to time differences m=-l,-2,....-L, respectively. Additionally, a multiplier 708 is provided for mutually multiplying data samples on the two input channels and an integrator 709 for integrating the multiplier output to produce an output representing the cross-correlation function R. . (n,0) corresponding to time difference m=0. The output of integrator 709 is converted by one of the absolute value conversion circuits 710 and applied to maximum value detector 711. In a practical aspect, each of the integrators 704, 707 and 709 may NE-498 1 be implemented with a transversal filter, or an average circuit, or a 2 recursive integrator. 3 Maximum value detector 711 of each time difference estimator 700j 4 detects a maximum of the absolute value inputs from circuits 710 and supplies a signal to a corresponding input of the controller 103 as an 6 estimate of the propagation time difference between the signals of the 7 i-th and (/' + 7)th channels. 8 In the previous embodiments, adaptive filters 104 quickly adapt 9 themselves to change their filter coefficients in response to changes in delay time differences. However, under noisy environment it is desirable 11 to prevent the echo canceller from quickly responding to changes in time 12 differences. Additionally, it is preferable to disable the updating of filter 13 coefficients with the error signals under such unfavorable conditions. 14 To this end, a modified version of the controller 103 is shown in Fig. 8 1 5 in which a delay unit 800 is connected to the outputs of time difference 16 estimators unit 101 to introduce a unit delay time. A comparator 801 is 17 connected to the input terminals and output terminals of delay unit 800 1 8 for making comparisons between each of the delayed versions of the 19 time difference signals and a corresponding non-delayed version of the 2 0 signals. As shown in Fig. 9, comparator 801 produces an output pulse 21 when there is a mismatch between each of the delayed input signals and 22 the corresponding non-delayed input signal. A timer, or preset counter 23 802 is connected to the output of the comparator 801 to generate a 24 pulse in response to a comparator mismatch output, while it starts 2 5 incrementing its count. When the count reaches a predetermined value, 2 6 the output of timer 802 switches to a low level. The output pulse of timer 27 802 is applied to the minimum propagation delay detector 300 and 2 8 coefficient vector selector 301 as a disable pulse so that they do not 2 9 respond to changes in their input signals for a prescribed interval set by 3 0 the timer 802. If the input signal contains noise, successive insignificant NE-498 .16. 2088558 1 changes may occur in the outputs of the time difference estimators unit 2 101. However, the outputs of the minimum propagation delay detector 3 300 and coefficient vector selector 301 are insensitive to these changes 4 and keep their previous output signals. The modified controller 103 further includes a second timer 803, which may have a different time-out 6 period from that of the first timer 802. Timer 803 also responds to the 7 output of the comparator 801 by generating a pulse which is supplied to 8 an update control 804 whose output is coupled to the adaptive filters 9 104. When the output of timer 803 is low, update control supplies an update enable command to adaptive filters 104, and when the timer 803 11 output is high, it supplies an update disable command instead. Thus, 1 2 under noisy conditions, adaptive filters 104 are prevented from updating 1 3 the contents of its coefficient memories 403 during a prescribed period 14 set by the timer 803.