Analyzing the Prosodic and Lingual Features of Popular Speakers

If the mission of Technology, Entertainment and Design (TED) is to spread ”great ideas” via talks, then high viewership of such talks is imperative to their success. But what are some stylistic differences between talks with higher viewership than others? To better understand this, we collect a large number of TED talk audio (N=2,065) from YouTube and create two categories: the lower (N1=402) and higher (N2=398) viewership classes. Using an explainable classifier (random forest), we find that prosody is not as effective a modality in predicting viewership classes, but augmenting it with linguistic features results in significantly better predictions. The prediction task is not the main objective of this work, rather it is implemented to help understand what prosodic and linguistic cues are important when differentiating between the delivery styles of low versus high viewership speakers. Although the main prosodic cues between the two classes are statistically significantly different, we found the most influential cue to be the “fraction of all words in the talk captured by LIWC”, a linguistic feature that has been shown to strongly correlate with the use of informal, nontechnical language.