Mike Mandel (Audience) - Reverberation is all around us, and yet little is understood about its effect
I'm really happy to introduce Mike Mandel, who will be talking about his work to understand reverberation and its effect on source separation algorithms. Reverberation is a tough problem, one that humans with normal hearing pretty much solve without thinking about it. The same can't be said for the elderly or our machines. Mike studied this problem at Columbia, and he's now at Audience (who make the sound chips for the new iPhones).
Who: Mike Mandel
Why: Reverberation is all around us, and yet little is understood about its effect
What: Evaluating Reverberant Source Separation
When: Friday May 11 at 1:15PM
Where: CCRMA Seminar Room, Top Floor of the Knoll at Stanford
Not much reverberation in the Hearing Seminar, but we'll talk about it this week. Should be a good introduction to a hard problem.
"Evaluating reverberant source separation"
While a number of recent algorithms can separate sources from reverberant binaural mixtures, the evaluation of these algorithms is still based on ideas from anechoic source separation. In this talk I will discuss some of the issues that arise when evaluating source separation algorithms with reverberant mixtures. These include a measure of the attenuation of Direct path, Early echoes, and Reverberation of Target and Masker speech (DERTM) as well as the definition of oracle masks based on a similar decomposition of impulse responses. I will also describe a new source separation algorithm, Model-based Expectation Maximization Source Separation and Localization (MESSL), that forms the core of a general probabilistic framework for source separation. An experiment evaluated using DERTM shows that MESSL and another state of the art source separator (Sawada et al., 2007) successfully suppress an interfering speaker's direct-path speech, but are much less successful in suppressing its reverberation. Such suppression slightly improves automatic speech recognition rates, but fails to improve intelligibility for human listeners.
Michael I Mandel uses signal processing and machine learning to model sound perception and understanding. He is currently an Algorithm Developer at Audience, Inc, applying machine learning to the problem of noise suppression in mobile phones. In 2009-10 he was a postdoctoral researcher in the Machine Learning laboratory at the Université de Montréal with Yoshua Bengio and Douglas Eck. He earned his PhD in Electrical Engineering from Columbia University in 2010 in Daniel P W Ellis' Laboratory for the Recognition and Organization of Speech and Audio (LabROSA) and his BSc in Computer Science from the Massachusetts Institute of Technology in 2004.