Affective encoding in the speech signal and in event-related brain potentials

A number of perceptual features have been utilized for the characterization of the emotional state of a speaker. However, for automatic recognition suitable objective features are needed. We have examined several features of the speech signal in relation to accentuation and traces of event-related brain potentials (ERPs) during affective speech perception. Concerning the features of the speech signal we focus on measures related to breathiness and roughness. The objective measures used were an estimation of the harmonics-to-noise ratio, the glottal-to-noise excitation ratio, a measure for spectral flatness, as well as the maximum prediction gain for a speech production model computed by the mutual information function and the ERPs. Results indicate that in particular the maximum prediction gain shows a good differentiation between neutral and non-neutral emotional speaker state. This differentiation is partly comparable to the ERP results that show a differentiation of neutral, positive and negative affect. Other objective measures are more related to accentuation than to emotional state of the speaker.