Design and evaluation of reduced-size feature sets for the assessment of sincerity in speech

The recognition of states and traits of speakers is a significant issue to investigate, to be able to achieve more useful interactive systems. The sincerity of a speaker is a relevant paralinguistic phenomenon, which have not received too much attention from the affective computing community. In this work, we tackle the problem using novel feature sets proposed for emotion recognition. In addition, bioinspired features (using an auditory signal representation) and other spectral features are also evaluated. Finally, diverse combinations of these reduced-size feature sets are built. The provided standard, complete set with 6373 features is used for comparison purposes. Results show that using the combination of the proposed representations and state-of-art features, it is possible to obtain very small feature sets (less than 3% of the original size) that get comparable correlation measure with respect to the baseline.