Statistical Unpredictability of F0 Trajectories as a Cue to Sentence Stress

This paper introduces a hypothesis that the perceived sentence stress in speech is related to the unpredictability of prosodic features, thereby capturing attention of the listener. In order to study this idea, a computational model was designed that learns the statistical structure of temporal F0 trajectories from continuous speech data without supervision using n-gram statistics. When the model output is compared to human perception of stress on a set of novel utterances, the lowprobability points of the F0 trajectories show high correlation with the moments of subjective perception of stress. The result gives support to the idea that perceptual attention and unpredictability of sensory stimulus are mutually connected, and suggests that stress perception can be learned with similar statistical learning mechanisms that are considered to play a central role in early word segmentation.