|
Abstract:
Speech signals exhibit a huge redundancy from an information content
point of view. In this talk I would like to present two problems where
speech redundancy, and its dual concept, speech sparsity play a key
role in solving them. A sparse representation allows one to hypothesis
a strong disjointness of multiple speech signals. This key observation
is further used to solve the blind source separation problem in
degenerate cases (when more sources are present than the number of
sensors). Another way of exploiting redundancy is by building
prior stochastic models compatible with the sparse representation
assumption. Then such prior knowledge can be used in a Bayesian
framework to perform signal estimation. I will present some issues and
results related to the noise reduction problem.
Radu Balan graduated from Polytechnic Institute of Bucharest, and
University of Bucharest, Romania, with Bachelor degrees in EE (Control
Theory) 1992, and Physics (Theoretical Physics) 1994, respectively. Then
he obtained his PhD degree from Princeton University (Applied and
Computational Mathematics) in 1998. After one year of postdoc at IBM
T.J.Watson and IMA (Minneapolis), he joined Siemens Corporate Research
in Princeton, NJ, where he currently is a Senior Research Scientist, in
the Department of Real-Time Vision and Modeling. At Siemens he has been
responsible for audio and signal processing applications in hearing
aids, mobile phones, and sensor fusion; he is also involved in analysis
and control of WLAN systems. For the past two years he has been also an
adjunct professor at Princeton University, teaching a graduate course in
Wavelets and Time-Frequency Analysis.
Bio: