InterSpeech 2021

Acoustic-Prosodic, Lexical and Demographic Cues to Persuasiveness in Competitive Debate Speeches
(Oral presentation)

Huyen Nguyen (Universität Hamburg, Germany), Ralph Vente (CUNY Hunter College, USA), David Lupea (NYU, USA), Sarah Ita Levitan (CUNY Hunter College, USA), Julia Hirschberg (Columbia University, USA)
We analyze the acoustic-prosodic and lexical correlates of persuasiveness, taking into account speaker, judge and debate characteristics in a novel data set of 674 audio profiles, transcripts, evaluation scores and demographic data from professional debate tournament speeches. By conducting 10-fold cross validation experiments with linear, LASSO and random forest regression, we predict how different feature combinations contribute toward speech scores (i.e. persuasiveness) between men and women. Overall, lexical features, i.e. word complexity, nouns, fillers and hedges, are the most predictive features of speech evaluation scores; in addition to the gender composition of judge panels and opponents. In a combined lexical and demographic feature model, we achieve an R² of 0.40. Different lexical features predict speech evaluation scores for male vs. female speakers, and further investigation is necessary to understand whether differential evaluation standards applied across genders. This work contributes a larger-scale debate data set in a democratically relevant, competitive format with high external relevance to persuasive speech education in other competitive settings.