François Pachet (Sony - CSL) "Utiliser les types, la réécriture, les expressions régulières pour construire un générateur d'extracteurs de descripteurs perceptifs." En Anglais: Extracting Perceptive Descriptors from Audio Files: An Evolutionist Approach. Abstract: We will present research conducted at CSL concerning the extraction of perceptive descriptors from audio/music files. These descriptors are key ingredients for the development of content-based access systems to music collections (EMD), sample collections (music composition) or more generally sound scene analysis. We present a language and a system (EDS) that models the design process of a high-level descriptor, as performed by a human processing expert. The system is based on an evolutionist approach in which signal features are evolved automatically and compared agains human perceptive values. We show preliminary but very encouraging results on difficult acoustic description problems (timbre classification, presence of singing voice, subjective energy) on which EDS finds systematically better features that the state-of-the art (e.g. Mpg7). We illustrate how these descriptors can be used in practice, with the MusicBrowser, the first large-scale content-based music access system. The Music Browser integrates editorial, acoustic and cultural descriptors in a single environment. For instance, users can then define their own, local, genre taxonomies, by giving a few music examples, and EDS will generalize from these examples to compute the descriptors on the whole database.