Events Calendar

Recent Development in Simultaneous Recognition of 2 or More Speakers Recorded on a Single Microphone

Abstract:
Human listeners are able to selectively tune out all but the speaker of interest when multiple speakers are talking, as in a cocktail-party scenario. We describe a new computational method which, for the first time, can accurately transcribe overlapping speakers better than humans can. This is a breakthrough: it is almost unprecedented for computers to outperform people in any field of perception, audio or visual.Bio:
Peder is currently a manager for the speech engines and algorithms group at IBM's TJ Watson Research Center. He is also serving as a chair of the User Interface Technology Professional Interest Community at IBM. Peder has been doing speech recognition at IBM since 1996 and has published over 50 papers in the area. His work in speech has been very mathematical in nature, as is reflected by an Erdos number (http://www.oakland.edu/enp/) of 3 earned while at IBM. Prior to joining IBM, Peder received a PhD in mathematics from the University of Michigan. There he proved that the existence of our universe is reasonable by showing that the Schrodinger equation is well posed and has a unique solution. Presently Peder lives in Westchester, New York, where he is trying to dissuade 6 foot long snakes from taking up residence under his porch.Host: Professor Shrikanth Narayanan

Location: Ronald Tutor Hall of Engineering (RTH) - 320

Audiences: Everyone Is Invited

Contact: Mary Francis