ISCA Archive SpeechProsody 2014
ISCA Archive SpeechProsody 2014

Encoding and decoding confidence information in speech

Xiaoming Jiang, Marc Pell

This study aims to investigate the perceptual-acoustic correlates of vocal confidence. Statements with different communicative functions (e.g., stating facts, making judgments) were spoken in confident, close-to-confident, unconfident and neutral voices. Statements with preceding linguistic cues (e.g. I’m positive, Most likely, Maybe, etc.) or no linguistic cues were presented to sixty listeners in a perceptual study. The listeners were asked to judge whether statements conveyed some level of confidence, and if so, they were asked to evaluate the level of confidence of the speaker. The results demonstrated that the intended levels of confidence varied in a graded manner in the perceptual rating score; the more confident the statement intended to be, the higher the rating. In general, the neutral voice was judged to be more confident than the close-to-confident voice, but less than the confident voice. The presence of a linguistic cue tended to increase ratings of confident voices but decrease ratings of voices in the less confident voice conditions. To evaluate how specific prosodic cues are used to encode and decode confidence information, acoustic analyses were performed on the stimuli without the linguistic cue based on the mean perceptual rating of speaker confidence for each item. Results showed that statements rated as confident versus unconfident differed in the mean and the variance of fundamental frequency (f0) as well as speech rate, with confident statements exhibiting lower mean f0, smaller f0 variance, and faster speaking rate than unconfident statements. The perceived level of confidence was differentiated in the mean fundamental frequency in a parametric way, the lower the level of confidence, the higher the mean f0. Confident voices were also distinct from the other three conditions in terms of mean and range of amplitude (i.e., loudness). These findings shed light on how linguistic and paralinguistic cues reveal confidence-related information to listeners during speech.