         
|
GRASP Lab Seminar 2004-2005
October 15, 11:00 AM, Levine Hall 307.
Justinian Rosca
Seimens Corporate Research
Blind Source Separation for Degenerate and Noisy Mixing Models.
Abstract: The area of this talk is audio source separation. Given a superposition of voices (e.g. a board meeting where participants impatiently overlay their speeches or a brainstorming session at a conference) can one extract each source signal and select the useful ones for meeting transcription/speech recognition purposes? The common thread in the Independent Component Analysis (ICA) and Blind Source Separation (BSS) literature is the case of two or more sensors (microphones) and also when the number of sensors D is larger than the number of microphones L. Very difficult estimation problems result when we assume that sources are not punctual, there are more sources than sensors (L>D), or there is diffuse noise (equivalent to the presence of an infinite number of sources). In real environments we have a combination of all these factors. In this talk I will review outstanding solutions to the problem of source separation in the presence of noise and when the number of sources can be larger than the number of sensors. I will highlight classes of assumptions that make the problem tractable and discuss several approaches. I will focus on a class of assumptions that can lead to effective algorithms, which are related to sparse representations of signals. The specific sparseness assumption used in this work is that the maximum number of statistically independent sources active at any time and frequency point in a mixture of signals is small. This is shown to result from an assumption of sparseness of the sources themselves, and allows us to solve the maximum likelihood formulation of the non-instantaneous mixing source estimation problem. The solutions obtained are applicable to an arbitrary number of microphones and sources, but work best when the number of sources simultaneously active at any time frequency point is a small fraction of the total number of sources. I will also present a live demonstration of audio source separation.
Biography: Justinian Rosca received his M.S. and Ph.D. in Computer Science from the University of Rochester, in 1994 and 1997, respectively. Previously, he received his Dipl.-Ing. in Computers and Control Engineering from Bucharest Polytechnic. He is currently program manager at Siemens Corporate Research in Princeton NJ, where he has grown an internal Siemens audio project into the Audio and Signal Processing Program, with focus on statistical signal processing. Dr. Rosca's research interests include fundamentals of speech representation and processing, blind signal separation, stochastic estimation and inference, adaptive principles in stochastic search and optimization, and probabilistic inference in artificial intelligence. Dr. Rosca has given tutorials on stochastic search techniques at the Genetic Programming and Genetic and Evolutionary Computation conferences in 1998 and 1999. Within Siemens he has been contributing to next generation technology for hearing aids and mobile phones. Dr. Rosca has more that two dozen patent applications, and more than 70 reviewed publications. He co-authored a book on solved problems in higher mathematics. He is presently on the editorial board of the Journal for Genetic Programming and Evolvable Hardware and serves as a member of committees of various conferences in the areas of machine learning and signal processing. He organized the Sparse Representations in Signal Processing workshop at NIPS 2003.
