When silence is gold

Audiovisual source separation is a fascinating approach to source extraction. Several algorithms have already been proposed for extracting speech sources from audio mixtures by exploiting audiovisual coherence. One of the main property of speech signals is that they are highly non-stationary: there are periods during which speakers do not produce sounds. In this work, the audiovisual coherence is used to estimate such silent periods which are then useful to extract corresponding speech signals.