Towards Morphological Sound Description using segmental models

Julien Bloit; Nicolas Rasamimanana
DAFx-2009 - Como
We present an approach to model the temporal evolution of audio descriptors using Segmental Models (SMs). This method yields a signal segmentation into a sequence of primitives, constituted by a set of user-defined trajectories . This allows one to consider specific primitive shapes, model their duration and to take into account the time dependence between successive signal frames, contrary to standard Hidden Markov Models. We applied this approach to a database of violin playing. Various types of glissando and dynamics variations were specifically recorded. The results show that our approach using Segmental Models provides a segmentation that can be easily interpreted. Quantitatively, the Segmental Models performed better than standard implementation of Hidden Markov Models.