Research interests

Current research

Statistical signal processing, statistical learning and computational statistics now are my main research areas. In the past (1990s), I also worked in the domain of speech and audio processing. More recently, I have been involved in some aspects of natural language processing. My current research interests include

Monte Carlo methods
Hidden Markov models
Classification of sequences
Online learning
Reinforcement learning

Recent information

Please see my list of publications, which I maintain regularly. Recent works are also generally available from my arXiv's search page.

You may also want to browse through some recent talks: Bayesian Methods for Latent Variable Models (course at Ecole d'été de Peyresq, 2010), Online EM Algorithm for Latent Data Models (University of British Columbia, April 2010), Inside Sequential Monte Carlo Methods (tutorial at IEEE SSP 2009, Cardiff, UK, 2009), An Introduction to Sequential Monte Carlo for Filtering and Smoothing (Worshop AssimilEx, Strasbourg, 2008), Non-Linear Filtering and Sequential Monte Carlo Methods (Réunion des GdR PHENIX & ISIS), Monte Carlo Methods for Cosmological Models (XXIIIrd IAP Colloquium), Adaptive Population Monte Carlo (Recent Advances in Monte Carlo Based Inference workshop, Isaac Newton Institute), Monte Carlo Methods for Bayesian Inference (Cosmology summer school, Cargese), On the use of sequential Monte Carlo methods for approximating smoothing functionals, with application to fixed parameter estimation (SMC'2006), Recursive EM Algorithm with Applications to DOA Estimation (poster, ICASSP 2006), A Markov Chain Monte Carlo Primer (University Pompeu Fabra, doctorate program seminar), Particle Methods for Parameter Estimation in Hidden Markov Models (talk at MTA SZTAKI), Particle Methods for Hidden Markov Models (3ème cycle romand de Statistique et de Probabilités Appliquées, EPFL, see also abstract and handout), Influence du schéma d'échantillonnage sur l'estimation des paramètres de la file M/M/1 (séminaire du projet Metropolis), Continuous-Time Jump MCMC for Model Selection; An Alternative to Reversible Jump Techniques? (séminaire LMC/IMAG), Recursive Computation for HMMs and Particle Approximations (workshop ACI MathSTIC).

More information

What follows is a somewhat more detailed statement of research for the period 1990-2004.

Computational statistics

MCMC and bayesian analysis

See [43], [47], [54], [61], [51], [53] if you are interested in Markov Chain Monte Carlo techniques. [54] discusses a method to approximate some smoothed functions in HMMs (see below) using particle filtering, this is detailed further in these slides. [47] discusses a method proposed by M. Stephens as an alternative to Green's reversible jumps approach (the software used for this paper is available here). [43] is an alternative to MCMC using multiple interacting simulations; it has been generalised by other authors, including Del Moral et al. and Douc et al.

Hidden Markov models

HMMs are now being used for an amazing range of applications. My main contribution in this field is our monograph [39] pubmished by Springer in 2005. The following HMM bibliography is an exhaustive list of references pertaining to HMMs which dates back to March 2001:

You may also be interested in the h2m toolbox.

Speech and audio processing

Spectral envelope estimation: Spectral envelope estimation for sinusoidal models is a topic on which we have worked with E. Moulines and M. Campedel-Oudot [86], [89], [56], [82], [84], [81]. The method which is briefly outlined in the following slides and described in more details in the paper can be tested using the following matlab functions.
Speech recognition: Speech recognition was my main activity during my stay at the CNET (1995-1996) [73].
Voice conversion: Here's what we did in 1995 with Y. Stylianou and E. Moulines (at ENST with support from CNET) [72], [79], [90]. Note that the training method proposed in [72] has been improved by A. Kain & M. Macon (ICASSP'98).
Speaker recognition: Here is my speaker recognition bibliography (HTML file), containing 110 entries (1996 - so it's maybe a little dated now), which corresponds to the report [91] (in French).
Restoration of audio recordings: Most of my PhD (1990-1993) was devoted to the study of background noise reduction techniques based on short-time spectral attenuation [95], [92], [88], [74]. There is also a collection of EMI Classics CDs which we restored with A. Chaigne and J. Laroche [93].