The theory of sound-types is a framework for sound analysis and synthesis designed
to represent and manipulate signals at a quasi-symbolic level. The basic idea is
to describe sounds by means of classes of equivalence and probabilities. Conceptually,
the analysis is implemented with the following steps:
- atomize: divide a sound in small overlapping chunks called atoms
(this can be done by windowing or by using more complex techniques such as atomic decomposition);
- make classes: compute a set of low-level features for each atom and project it
onto a feature-space; apply any kind of clustering algorithm (such as GMM) to find the principal c
lusters of atoms in the space;
- compute probabilities: apply any kind of sequential analysis (such as HMM)
to estimate the probabilities that a cluster is followed by another cluster in the original signal.
For more information please see
the related papers.
THIS PAGE IS UNDER CONSTRUCTION: MORE MATERIAL WILL BE READY SOON