Interspeech'2005 - Eurospeech
An objective of the paper was to determine a set of low-dimensional feature spaces that provide high emotion recognition rates. Candidates for target feature spaces were randomly drawn from a broad pool of speech signal parameters that comprised both commonly used characteristics and newly introduced features. As a result, several four-dimensional feature spaces that provide the highest emotion classification rates (68%) on Polish language database, which we used in experiments, were identified.
Bibliographic reference. Cichosz, Jaroslaw / Slot, Krzysztof (2005): "Low-dimensional feature space derivation for emotion recognition", In INTERSPEECH-2005, 477-480.