Research

My field of interest mostly covers audio indexing and statistical learning. My Master internship and subsequent PhD, attended jointly at TELECOM ParisTech and the french broadcast station RTL, were dedicated to audio classification, i.e. the partitioning an audio excerpt according to its acoustical content. My PhD thesis "Classification automatique de flux radiophoniques par Machines à Vecteurs de Support" discusses the application of Kernel Methods (such as Support Vector Machines in particular) to the problem of speech/music classification. It explores the issue of tuning the kernel and its parameters, and selecting relevant features for a given problem, as well as ways to introduce the temporality on a classic bag-of-frames scheme, by combining successive SVM outputs.

After lecturing my PhD and until now, my postdoctoral work at IRCAM focused on audio fingerprinting. Audio fingerprinting consists in robustly characterizing chunks of audio signal in order to recognize it among a database of several thousands or millions of tracks. The AudioPrint technology, developed and patented by IRCAM, relies on a double-nested FFT that models the local variations of sub-band spectral energies.