Séminaire n°8

Similarity in Sparse Domains with Applications to Audio Signals

Intervenant :  Bob Sturm, Post-doc LAM

Contact : boblsturm(at)gmail.com

Date : 07/12/09

Abstract : 
I discuss two problems related to similarity in audio signals, and show how they can be addressed in a sparse domain found using methods of sparse approximation with overcomplete multiscale dictionaries. The first problem is searching for similarity given a query signal. While this problem has been thoroughly addressed in text analysis, time-series analysis, image analysis, and large audio databases, it has not been completely addressed from the perspective of sparse approximation. The fundamental question is: "How can one meaningfully compare two sparse signal models that are made of different elements?" I show, in collaboration with Prof Laudent Daudet, that sparse signal models do provide a scalable way to look for similarity in audio signals, and that they can be very robust to interfering signals. The second problem is signal content classification. This work, in collaboration with Dr. Marcela Morvidone and Prof Laudent Daudet, looks at the automatic recognition and discrimination of musical instruments from real (but monophonic) music recordings using features derived from multiscale representations, one in a sparse domain. We show that the addition of scale information makes the features more discriminative than features computed with a single resolution. This presentation provides an overview of my work accomplished at LAM during my tenure as a Chateaubriand Fellow since March, and ending in December --- at which time I become an assistant professor at University of Aalborg, Copenhagen, in the Department of Media Technology and the Medialogy program.