dreu 2012

I am working on a project to segment transcripts of dialogues into "high stakes" segments, in which the speakers are talking about matters important to them, vs. "low stakes" segments, which contain small talk. This project is part of Babel, a project to quickly develop speech recognition software for languages where large amounts of high-quality tagged data is not available. It will assist in keyword search of the transcripts produced; since people speak more clearly in high-stakes segments, their speech should be easier for the speech recognition system to recognize, so search results from high-stakes segments should be weighted more heavily.