Supra‐segmental Features
Björn Wolfgang Schuller, Anton M. Batliner · 2013
The general principle of supra-segmental features is to obtain a single, fixed length feature vector which describes a sequence of low-level descriptors (LLDs) of possibly variable length. Common methods to obtain a supra-segmental feature vector are: mapping of the LLDs to a single vector by applying functionals to the LLD time series, stacking of low-level feature frames optionally followed by a dimensionality reduction, for example, by principal component analysis. Functionals are relations which map a (time) series of input values of arbitrary length to a single output value. Moreover, functionals can be applied hierarchically, that is, another set of functionals can be applied to a series of supra-segmental feature vectors.