METISS: Speech and audio data modelling and processing

METISS was an  IRISA joint project-team between INRIA and CNRS.

After a complete lifecycle as an Inria project-team (2001-2012) METISS has been stopped on December 31 2012. A new project-team called PANAMA has been created.

See the web site of PANAMA

Context and objectives

The application domains of the METISS research group are focused on audio signals and include the following topics:

• Audio signal description, segmentation and structuring, in particular for indexing purposes
• Audio scene analysis and source separation
• Automatic information extraction from spoken content (speech recognition / speaker characterisation)
• Advanced models and algorithms for music information retrieval

The scientific framework on which are based our activities relies on probabilistic models, such as Hidden Markov Models and Bayesian Networks, with a focus on robust training and decision. Recent signal processing techniques, such as sparse representations and compressive sensing, are also investigated, as promising ways of approaching description and separation of audio signals.

The main industrial sectors served by our research are the Telecommunication, Internet and Multimedia industry, but also the sectors of Music and Audiovisual production and the sector of Educational Software and Games.

Research axes

• Robust audio classification
• Spoken content processing for multimedia information extraction
• Audio motif discovery
• Robust source separation (under-determined case, diffuse sources, ...)
• Object-oriented music description
• Musical language modelling
• Compressive acquisition of sound fields
• Dictionary learning for sparse representation
• Sparse decomposition algorithms (Matching Pursuit, convex and non-convex optimization)

International and industrial partnerships

• Speech / Speaker Recognition: Parole Project-Team (LORIA, Nancy), Orange-Labs (Rennes, Lannion), Telisma (Rennes), LIA (Avignon), LIUM (Le Mans)
• Music / Source Separation: University of Tokyo (Japan), NTT (Japan), IRCAM (Paris), Telecom Paris Tech’ (Paris), IRIT (Toulouse), LIMSI (Orsay)
• Sparse Representations / Compressed Sensing: EPFL (Switzerland), Queen Mary Univ. (UK), Technion (Israël), Univ. Edimbourg (UK), Jean Le Rond D’Alembert & Jacques Louis Lions Institutes (Univ. Pierre & Marie Curie, France), Langevin Institute (Univ. Diderot, France), LATP (Univ. Provence, France)
• Audio and Multimedia content: Texmex Project-Team (IRISA, Rennes), INA (Paris), Thomson (Rennes), NII (Japan), LTL Lab (Mexico), Exalead (Paris, France), Grass Valley (Rennes)

Scientific leader

• Frédéric Bimbot : Frederic.Bimbot@irisa.fr

+33 2 99 84 75 06

• Administrative assistant: Stephanie.Lemaile@irisa.fr

+33 2 99 84 72 52

About the team

• Web site : http://www.irisa.fr/metiss/
• Activity report : http://ralyx.inria.fr/2008/Raweb/metiss/uid0.html