Laurent Daudet: "In the search of the perfect signal representation for audio source characterization"

Laurent Daudet , is Associate Professor at the Université Pierre-et-Marie-Curie (UPMC - Paris 6), France ; Visiting Senior Lecturer at Queen Mary University of London. This talk is sponsored by CIRMMT Research Axis 3.

ABSTRACT:

Significant progress has been made in the last decade for the characterization / classification of audio signals, using powerful
machine learning techniques. In general, these are not applied directly on the audio samples, but on a limited number of time- and/or
frequency-domain features, usually computed from a short-time Fourier transform of the signal.

This talk will present a few experiments where we try alternate, more adaptive, signal processing techniques to dig out the information from
the audio. One such technique uses so-called "sparse representations", i.e. where the signal is seen as a linear combination of a small
number of elementary waveforms, taken from a fixed, very large set - a well-known optimization problem ! Beyond an obvious application in
audio coding, we will demonstrate that this also rearranges the information in a hierarchical way, that can be used to build "scalable
features". Example applications are audio similarity, and audio classification experiments.