First page Back Continue Last page Overview Text
However, it’s still possible that these people aren’t having a conversation together, but rather talking on phones nearby each other. To determine whether the two speakers are actually talking we take the mutual information between the audio streams. If they are synched – essentailly the voicing segments should be the noisy compliments of each other.
To detect an interruption – we look for high-energy, non-correlated wav forms (essentially two people speaking at the same time) followed by a speaker transition. Preliminary results show this works quite well.