RÉSUMÉ. Nous présentons les principaux travaux menés dans le projet Sample Orchestrator, destiné au développement de fonctions innovantes de manipulation d'échantillons sonores. Celles-ci se fondent sur des études consacrées à la description des sons, c'est-à-dire à la formalisation de structures de données pertinentes pour caractériser le contenu et l'organisation des sons. Ces travaux ont été appliqués à l'indexation automatique des sons, ainsi qu'à la réalisation d'applications inédites pour la création musicale -synthèse sonore interactive par corpus et aide informatisée à l'orchestration. Le projet a aussi comporté un important volet consacré au traitement de haute qualité des sons, à travers plusieurs
RÉSUMÉ ÉTENDU EN ANGLAISThe Sample Orchestrator R&D project was aimed at the development of innovative functions for the content-based manipulation of sound samples. Its main objectives and advances are summarized hereinafter.
Sound description and automatic indexingThis objective concerns the development of new heuristics for managing sample databases. The sound description issue involves a top-down process for elaborating representations of the sound world relevant to the human cognition. Recent studies define three levels for sound similarity: semantic, causal and acoustic. Causal relates to the sound production in terms of sources and actions. For each homogeneous sound set, a specific timbre space, with dimensions directly correlated to acoustical attributes, can be obtained. A first study consisted in performing a meta-analysis of timbre spaces resulting from several studies related to different sound corpora. A second study was dedicated to an experience of categorization of a given corpus of environmental sounds targeted to the materials of the sources. The results show a good correlation between the experimental data and the actual categories. Three labeled corpora were elaborated from these experiences: Materials, Reduced materials and Onomatopoeia.The second stage, automatic indexing, aimed at building a signal model of the classes in order to perform the automatic classification of new sounds from the analysis of their signals. First, an extensive set of audio descriptors, combining several models (spectral, harmonic, perceptual) was elaborated and implemented as the IrcamDescriptor library. For the learning phase of classification, an algorithm provides a selection of descriptors for a given set of classes. After applying an LDA, various supervised classification methods were evaluated. The best results were obtained with an SVM, with accuracies and f-measures greater than 70% for the 3 formerly produced corpora.A last study concerned the categorization of sound morphologies. 6 classes emerged from an experience on intensity profiles and very good results were obtained for the automatic classification with a binary decision tree operating on selected descriptors.
High-quality sound processingThis objective consisted in extending the phase vocoder model for high-quality sound processing. The model...