Crosstalk Reuction Usin a New Aaptive Noise Canceller ZAYED RAADAN an ALEXANDER POULARIKAS Electrical an Computer Enineerin Department The University of Alabama in Huntsville Huntsville, AL 35899 USA Abstract: This paper introuces a new aaptive noise canceller () to improve the system performance in the presence of crosstalk. The propose consists of three microphones an two aaptive filters that automatically ajust their impulse responses throuh least mean-square (LS) alorithms. Two microphones are use to represent the oriinal speech sinal an the reference noise input. The thir microphone is use to provie a sinal that is processe throuh the first aaptive filter to cancel the sinal crosstalk leakin from the primary input into the reference input. The propose is simulate usin ifferent noise power levels for both stationary an nonstationary noise environments. Simulation results, carrie out usin a real speech, clearly emonstrate the sinificant achievements of the propose in minimizin the sinal istortion an reverberation. Key-Wors: Aaptive filterin, Crosstalk reuction, LS alorithm, Noise cancellation, Sinal leakae. Introuction An important operation in voice communication systems involves the extraction of noise from the esire speech. This problem arises in many situations, such as airplanes, helicopters, an automobiles where acoustic noise is ae to speech. Althouh the sinle microphone approach for noise cancelin can be achieve usin Wiener an Kalman filterin, the two-microphone approach usin aaptive filterin is a more powerful technique for that purpose. The strenth of the aaptive noise cancellers lies in the fact that no prior knowlee of the speech sinal or the corruptin noise is require. However, a correlation between the noise that corrupts the speech an the noise in the reference input (aaptive filter input), is necessary for the aaptin least mean-square (LS) alorithm to remove the noise from the primary input sinal. A typical aaptive noise canceller (), shown in Fi., is compose of two inputs: primary input an reference input. The primary input consists of the oriinal speech sinal, S, corrupte by an aitive noise v. The noise source is represente by, an the transmission path from the noise source to the primary input is represente by the low pass filter, h. The input to the aaptive filter is the reference sinal that is correlate with v, but uncorrelate with S. The effectiveness of the epens on how much v an are correlate. The filter weihts w are aapte by means of an LSbase alorithm to minimize the power in the output sinal. This minimization is achieve by processin via the aaptive filter to provie an estimate of v, ( y = vˆ ), an then subtractin it from to et e. Thus, at the kth iteration: e(k) = S(k) + v(k) y(k) () any two-microphone s have been propose in the literature [-5] usin LS-base alorithms that alter the step-size of the upate equation to improve the trackin ability of the alorithm an its spee of converence as well. In all these s, it was assume that there are no sinal components leakin into the reference input. The presence of these sinal components (also calle sinal crosstalk or sinal leakae) at the reference input is a practical concern because it causes cancellation of a part of the oriinal speech sinal at the input of the,
S h v w y e recovere speech. It is assume that microphones an 3 are place farther apart such that there is no sinal crosstalk leakin from the first into the secon. any LS-base aaptation alorithms coul be use in the s incluin the stanar an normalize LS alorithms [8], [9]. However, we prefer usin the LS aaptation alorithm which was also in one of our previous works [] for its superiority over other alorithms. In that alorithm, the weiht upate recursion is iven by Fi. : A conventional with no sinal leakae. an results in severe sinal istortion an low sinal to noise ratio at the output of the. The manitue of this istortion epens on the sinal to noise ratios at the primary an reference inputs. Several techniques were propose in the literature to enhance the system performance in this case of sinal leakae (see [6], [7]). Hih computational complexity is associate with these alorithms. In the present work, we propose a new that uses two aaptive filters an three microphones instea of two as in a typical. The thir microphone provies a sinal that is an attenuate replica of the esire sinal. That sinal is processe throuh the first aaptive filter to cancel the sinal components leakin into the reference input. The secon aaptive filter is use to cancel the noise at the input of the. Propose Fiure shows a block iaram of the propose. The first microphone represents the speech sinal an the secon microphone represents a mixture of noise an sinal components leakin from the first microphone throuh a channel with impulse response h 3. These sinal components cause istortion in the recovere speech at the output of a conventional. To solve this problem we introuce a thir microphone to provie a sinal that is an attenuate replica of the oriinal speech. This sinal is processe by the first aaptive filter (w ) to prouce a crosstalk-free noise at its output. This noisy sinal, with almost no leakae of the speech, is processe throuh the secon aaptive filter to cancel the noise at the input of, an accorinly prouces the where w (k +) = w (k) + k e (k) = n= α + e(k) e(k) v (k) () e (k n) (3) is the square norm of the error vector e(k), estimate over its entire upate lenth k, α is an aaptation constant, an v is the input of the filter an is replace by v 3 in the first aaptive filter an by v in the secon. ε is a small positive number, ae to avoi a ata over-flow error when e(k) becomes too small []. This propose alorithm was shown to have a small number of computations []. The performance of the aaptive noise canceller may be escribe in terms of the excess mean-square error () or misajustment. The at the k th iteration is efine by L (k) = L j= ee (k j) (4) where ee (k) = e(k) S(k) is the excess (resiual) error, k is the sample (iteration) number, an L is the number of samples use to estimate the. The effect of L is just to smooth the plot of. The steay-state ( ss ), estimate by averain (k) in (4) over k after the alorithm has reache steay-state conition, is efine by K ss = ( ) (k) (5) K P = k P
S e h h3 v y h v 3 v w y w Fi. : Propose for sinal leakae problem where K is the total number of samples of the speech sinal, an P is the number of samples after which the alorithm reaches steay-state conition. The misajustment, a normalize mean-square error, is efine [8] as the ratio of the steay-state excess SE to the minimum SE. ss = (6) SE min where SE min equals the power of the oriinal clean speech sinal, S, averae over samples at which the alorithm is in steay-state (k P) an is iven by K SE min = ( ) S (k) (7) K P = Computer simulations were accomplishe by usin a real speech an ifferent noise power levels for both stationary an nonstationary noise environments. The simulations show performance superiority of the propose in ecreasin sinal istortion, reverberation an consequently, proucin small values of. The propose is simulate usin a real speech an ifferent noise power levels for both stationary an non-stationary noise environments. Computer simulations show performance superiority of the propose in ecreasin sinal istortion, reverberation an consequently, proucin small values of. k P 3 Simulation Results The simulations of the propose were carrie out usin a male native speech sayin soun eitin just ets easier an easier an sample at a samplin frequency of.5 khz. The number of bits per sample is 8 an the total number of samples is 33, or 3 sec of real time. The simulation results are presente for stationary an nonstationary environments. For the stationary case, the noise was assume to be zero mean white Gaussian with three ifferent variances as shown in Table. For nonstationary case, the noise was assume to be zero mean white Gaussian with a variance that increases linearly from min =. to three ifferent maximum values σ max as emonstrate in Table. In all simulations, the followin values of parameters were use: L=, P=, ε =., N =N =, α =, an α =.. The values of α were selecte as a compromise between fast rate of converence an oo trackin capability with most important concern to have a hih rate of converence in the first aaptive filter (α =) an oo trackin capability in the secon (α =.). The impulse responses of the three IIR low pass filters use in the simulations, are: h =[.5.5.], h =[.5.4.], an h 3 =[3..3]. Fiure 3 illustrates the performance of our propose in cancelin the sinal leakae at the output of the first aaptive filter for the case in which σ =. as shown in Table. From top to bottom, that fiure shows the oriinal speech (S),
Table : Comparison of the ss an of the propose an conventional s for stationary noise case. Stationary white zeromean noise =. =. =. Table : Comparison of the ss an of the propose an conventional s for nonnonstationary noise case. Nonstationary white noise min =. max =. max =. max =. Conventional Conventional Propose Propose.7 53.9 45.9.3 3.7 4. 38..6 6. 4.6 3.9 6.4.4 57. 48.3..5 55.7 4..6 4.5 35.6 38..5 S V3 V S.5 -.5.5.5.5 3.4. -. -.4.5.5.5 3.. -. -..5.5.5 3 Fi.3: Cancellation of crosstalk at the output of the first aaptive filter of the propose. From top to bottom: Oriinal clean speech (S), noise corrupte with crosstalk (V 3 ), an cross-talk- free noise (V ). See.5 Fi.. ( σ =., Table ). -.5.5.5.5 3 combination of noise an sinal leakae (v 3 ), an the error sinal of the first aaptive filter (v ) which is the noise free of sinal leakae. A comparison of the propose with the conventional for both stationary an nonstationary noise environments is shown in Tables an. The aaptation constants of the LS alorithms use in both s were selecte to achieve a compromise between small an hih initial rate of converence for a wie rane of noise variances. From these Tables, improvements of up to 6B in ss of the propose over the conventional one were achieve. It is worthwhile to note that if the noise variance increases, the performance of the conventional becomes better as illustrate in Tables an. This is expecte because increasin noise power results in a S - e S - e -.5.5.5 3.5 Propose -.5.5.5.5 3.5 Conventional -.5.5.5.5 3 Fi.4: Performance comparison between the propose an conventional s. From top to bottom: Oriinal clean speech (S), noise corruptin speech (), resiual (excess) error of the propose, an resiual error of the conventional ( max =., Table ).
in B - - -3-4 -5-6 -7 Propose Conventional -8.5.5.5 3 Fi.5: of the propose an conventional s ( max =., Table ). less sinificant effect of the sinal leakae at the reference input. Fiures 4 an 5 provie more illustrations of the sinificant achievements of the propose over the conventional one for the nonstationary noise case in which max =. (Table ). From top to bottom, Fi.4 shows the speech sinal (S), the noise corruptin speech (), an the resiual error (S e) of the propose an (S-e) of the conventional. The effect of increasin the variance of the noise on the processe speech is clearly shown in the secon plot of Fi.4 (plot of ). Fi.5 shows the plot of for both s. The improvements of the propose are clearly evient. 4 Conclusions In this paper we presente a new that corrects the cross talk leakin from the primary channel to the reference channel. The propose consists of three microphones an two aaptive filters that use LS-base aaptation alorithms. The leakin sinal is cancelle by the first aaptive filter usin an attenuate replica of the speech sinal provie by a thir microphone. Accorinly, the output of the first aaptive filter is a crosstalk-free noise an this noise is cancelle throuh the secon aaptive filter. Compare with the conventional in both stationary an nonstationary noise environments, the propose emonstrates superior performance in minimizin sinal istortion, reverberation an. References: [] S. Ikea an A. Suiyama, An aaptive noise canceller with low sinal istortion for speech coecs, IEEE Trans. On Sinal, Processin, vol. 47, pp 665-674, arch 999. [] J. E. Greenber oifie LS alorithms for speech processin with an aaptive noise canceller, IEEE Trans. On Speech an Auio Processin, vol. 6, pp 338-35, July 998. [3] W. A. Harrison, J. S. Lim, an E. Siner, A new application of aaptive noise cancellation, IEEE Trans. Acoust.., Speech, Sinal Processin, vol. 34, pp. -7, Jan. 986. [4].J. Al-Kini an J. Dunlop, A low istortion aaptive noise cancellation structure for real time applications, in Proc. IEEE ICASSP, 987, pp. 53-56. [5] S. F. Boll an D. C. Publisher, Suppression of acoustic noise in speech usin two microphone aaptive noise cancellation, IEEE Trans. Acoust.,, Speech, Sinal Processin, vol. ASSP- 8, pp. 75-753, 98. [6] G. irchanani. R. L. Zinser, an J. B. Evans, A new aaptive noise cancellation scheme in the presence of crosstalk, IEEE Trans. Circuits Syst.., vol. 39, pp. 68-694, 99. [7] V. Parsa, P. A. Parker, an R. N. Scott, Performance analysis of a crosstalk resistant aaptive noise canceller, IEEE Trans. Circuits Syst., vol. 43, pp. 473-48, 996. [8] S. Haykin, Aaptive filter theory, Prentice-Hall, Upper Sale River, NJ,. [9] S. Haykin an B. Wirow, Least-ean-Square Aaptive Filters, Wiley, NJ, 3. [] Z. Ramaan an A. Poularikas An aaptive noise canceller usin error nonlinearities in the LS aaptation Proc. IEEE Southeastcon, Greensboro, North Carolina, arch 4. [] N. J. Bersha, Behavior of the ε-normalize LS alorithm with Gaussian inputs, IEEE Transactions On Acoustics, Speech, Sinal Processin, vol. ASSP-35, pp. 636-644, ay 987.