Creation and characterization of an emotional speech database

Abstract

Paralinguistic features of speech communicate emotion in the human voice. In addition to semantic content, speakers imbue their messages with prosodic features comprised of acoustic variations that listeners decode to extract meaning. Psychological science refers to these acoustic variations as affective prosody. Most studies of affective prosody obscure semantic content, although the resulting stimuli are less representative of naturally occurring emotional speech. The presented works details the creation of a naturalistic emotional speech database on which both acoustical analysis and a listening study were conducted. To this end, 55 adults were recorded speaking the same semantic content in happy, angry, and sad voices. Based on prior acoustic analyses of affective prosody, classic parameters were extracted including pitch, loudness, timing, as well as other low-level descriptors, and compared the acoustic features of each emotion with published evidence and theory. Preliminary results indicate that this naturalistic speech samples yielded acoustic features that are congruent with prior experimental stimuli of anger and happiness, but was less consistent with sadness. The results of the listening study indicated that listeners discriminated the intended emotions with 92% accuracy. The dataset therefore yielded a database of emotionally salient acoustical information for further analyses. [Work supported by NIH-R21-104547.]