|M.Sc Student||Schreibman Amos|
|Subject||Adaptive Stereo Acoustic Echo Cancellation in Reverberant|
|Department||Department of Electrical Engineering||Supervisor||Professor Israel Cohen|
|Full Thesis text|
Acoustic Echo cancelation (AEC) is an important tool which is needed to reduce the echo levels heard by the speaker in communication systems such as full-duplex hands free teleconferencing system. This echo is the result from the loudspeaker to microphone coupling in the near-end room. Adaptive filtering techniques are usually used in order to estimate the near-end room loudspeaker to microphone impulse response. The estimated echo signal is subtracted from the received microphone signal, thus reducing the level of the returning echo to the far-end room.
Stereophonic communication systems provide a more lifelike conversation quality since the participants can use the added spatial information to distinguish the current speaker location.
The current available solutions that reduce the stereophonic echo focus mainly on solving problems caused by the high correlation between the two far-end room microphone signals (the near-end room loudspeakers signals). This high correlation is caused by the single origin of the two signals (the far-end room speaker). Therefore the two signals differ only by the convolution with the far-end room speaker to microphone impulse responses. These solutions assume the existence of a Double Talk Detector (DTD), a device which freezes the adaptive filters adaptation if a double talk interval in the near-end room is detected. However, in the presence of a constant near-end room interference, the existing solutions are unable to converge and therefore do not function properly. Furthermore, due to the placement of both the near-end room loudspeakers and microphone, it is possible that one of the near-end room loudspeaker to microphone impulse responses will consist of higher energy when compared to the second near-end room loudspeaker to microphone impulse response. In such case, due to the use of Normalized Least Mean Square (NLMS) based solutions in the existing solutions, the residual echo of the dominant channel interferes with the weaker channel convergence.
This work suggests the use of a new noise and double talk resilient adaptive algorithm in collaboration with the POWER stereophonic echo cancelation algorithm. To produce a complete solution, the noise and double talk resilient algorithm is able to converge during constant near-end room interference (continuous double talk interval). Additionally, the algorithm produces good results for a low signal to noise ratio (SNR) environment in order to comply with the problem of a weak channel convergence in the presence of a dominant channel.