Ircam - Centre Georges-Pompidou Equipe Analyse/Synthèse



 CUIDADO Workpackage WP 2.1: AUDIO FEATURE EXTRACTION
 
 

Descriptor selection and design (PCM, A/S, UPF)


Scientific knowledge in sound perception (timbre spaces) and cognition (textual) extended by experimental studies.

According to the signal categories, different levels of abstraction, ranging from samples description to semantic description such as textual attributes.

Since some of these descriptors can relate to other ones, a hierarchical description scheme is required (allowed cost).

A descriptor can be a combination of other ones giving vectors of description (e.g. positioning in multi-dimensional spaces).

Select the best suited high-level descriptors for the Sound Palette and the Web monitoring system:


Use of the information theory (selection of descriptors supporting the maximum information). Classes are the different index values to be attributed to signal events. A first stage will consist in computing mutual information for sets of sound features (temporal, spectral,Ö) and sets of indexes. Then features with maximal mutual information are selected and descriptors are based on these features.
 
 
 



 
Page remise à jour le 14 Janvier 2001