Analysis of Voice Stress in Call Centers Conversations

This paper presents how voice stress is manifested in the acoustic and phonetic structure of the speech signal. Out of few thousand authentic Police 997 emergency phone calls, a few hundred were selected automatically according to their duration (dialogs shorter than 3-4 seconds were omitted). Finally, 45 speakers were chosen for acoustic evaluation, the basis for selection being a perceptual assessment. Basic statistical measurements for stressed and neutral speech run over the database showed the relevance of the arousal dimension in stress processing. The MDVP analysis confirmed statistical significance of such parameters as fundamental frequency and pitch variation, noise-to harmonic-ratio, subharmonics and voice irregularities for stress detection. In case of highly stressful conditions a systematic over-one-octave shift in pitch was observed. Linear Discriminant Analysis based on 9 acoustic features showed it is possible to categorize the following classes: male – stressed and neutral, female stressed and neutral speech.