The Removal of Acoustic Echo

vii Introduction Acoustic echo is the delayed and possibly distorted version of the transmitted sound reflected back from the loudspeaker to the microphone. Attempts have been made to reduce echo by using directional microphone. This solution reduces the acoustic echo by placing the loudspeaker in the null of the microphone, thereby eliminating the direct path of the acoustic coupling between them. The performance of this system is higly sensitive to the positioning of the microphone and the speakers. Attempts have also been made to the acoustic treatment of the room. In this thesis we analyze echo cancelers based on L.M.S. and N.L.M.S. algorithms. The echo canceler first tries to estimate the echo path and then generates a replica of echo. The estimated echo is then subtracted from the original signal. Adaptive filtering is required to obtaine a good replica of echo, since the echo path is usally unknow and time varying. The recommendation ITU-T : G.167 ( shown in Appendix 5) specifies the performance requirements of acoustic echo control devices. The performance of the echo canceller depends on the choice of adaptive filter algorithm. In this thesis we illustrate the performance of the adaptive filtering algorithm L.M.S. and N.L.M.S. by using simulation results obtained in Matlab and real implementation on ADSP 2181 Devolopment Board from ANALOG DEVICE. We used three sets of input signal for the simulation : white Gaussian noise , speech file , speech + echo file , this last obtained from a real situation. Frequently , many developers uses Echo Return Loss as the performance index of the algorithms. E.R.L. is estimated as the ratio between the energy of the original echo to the energy of the residual echo. But this is possible if and only if echo is modeled as an allpass system. The energy of the signal is computed as the 128- points moving average of istantaneus square amplitude. In our thesis , we used a real signal for both simulation and implementation. That’s why it was impossible an estimation of original echo and residual echo. So, our parameter index is M.S.E. ( mean square error ) computed as the energy loss from the original input signal after the adaption process : The simulation was based on Matlab 5.2 classroom limited version . LMS and NLMS algorithms codes are shown on Appendix 1 and Appendix 2 respectively. While a good graphic simulation was possible by comparing MSE ,some memory problems was encountered while reading the sound file (`echosim.wav`). So that , the simulation was working only with a limited number of samples ( around 10000 – 15000 samples) for each signal input. In the other hand this number of samples are quite enough for an approximately evaluation of Output Estimation Error and Mean Square Error. 22 2 2 _ _int_128 _int_128 outin outout inin sslossenergye saveragemovingpos saveragemovingpos