Research Seminars

Less is more: sparse representations for audio

Laurent Daudet
Musical Acoustics Group, D'Alembert Institute for Mechanical Engineering, University Pierre-and-Marie-Curie - Paris 6

Tuesday 2 June 2009, 15:00, Room 105

Abstract

This talk will be focused on signal modeling using sparse decompositions in overcomplete dictionaries, with a strong focus on audio signals. In such models, a signal is approximated by combining a small number of elementary waveforms ("atoms"), taken from a very large collection ("dictionary"). This provides extra flexibility (e.g. apparently avoids time-frequency resolution constraints) but comes with increased complexity over standard Fourier-based analysis. Greedy techniques have however been developed that provide near-optimal decompositions in reasonable computational cost, i.e. applicable on large-scale multimedia databases. After a general overview, I will discuss recent applications that takes advantage of sparsity, combining scalable audio coding with Music Information Retrieval applications.

Bio

Laurent Daudet is Associate Professor at the Pierre-and-Marie-Curie University (UPMC aka Paris 6), France. He's also Visiting Senior Lecturer at Queen Mary University of London. After a physics education at the Ecole Normale Superieure, Paris, France, he received a Ph.D. degree in applied mathematics from the Universite de Provence, Marseille, France, in 2000. In 2001 and 2002, he was an EU Marie Curie Post-doctoral Fellow with Prof Mark Sandler at the Centre for Digital Music at Queen Mary University of London. Since 2002, he has been working at UPMC where he joined the Musical Acoustics Laboratory (LAM), now part of the D'Alembert Institute for mechanical engineering. He is author or coauthor of over 70 publications on various aspects of digital audio signal processing. His research focuses mainly on applications of sparse signal processing for the analysis and synthesis of audio signals.