Certain
inverse problems can be solved quite efficiently if the solution is
known to have a sparse atomic decomposition with respect to some basis
or frame in a Hilbert space. One particular example of such an inverse
problem is the so-called cocktail party (or blind source separation)
problem: Suppose we use a few microphones to record several people
speaking simultaneously. How can we separate individual speech signals
from these mixtures?
In this talk, I will describe an algorithm adressing the blind source
separation problem when the number of speakers is larger than the
number of available mixtures. The algorithm is based on the key
observation that Gabor expansions of speech signals are sparse. The
separation is done in two stages: First, the "mixing matrix" A is
estimated via clustering. Next, the Gabor coefficients of individual
sources are computed by solving many q-norm minimization problems of
type {min ||x||_q subject to Ax=b}. Several choices for the value of q
will be compared.