Adaptive cancellation of acoustic echoes during double-talk based on an information theoretic criteria

This paper casts the problem of canceling acoustic echoes during a double-talk event as a blind signal separation problem. Principles of independent component analysis are applied to derive algorithms for updating the echo cancellation filter during double-talk, and the near-end signal is separated out from the microphone signal. This approach simplifies the the structure of acoustic echo cancelers by avoiding the need for double-talk detection. The same updating algorithm is applied under all conditions. The proposed approach is compared with existing methods on live data recorded on a speaker phone.