No abstract

No abstract

For both individual politicians and political parties the internet has become a vital tool for self-promotion and the distribution of ideas. The rise of streaming has enabled political debates and speeches to reach global audiences. In this paper, we explore the nature of charisma in political speech, with a view to automatic detection. To this end, we have collected a new database of political speech from YouTube and other on-line resources. Annotation is performed both by native listeners, and Amazon Mechanical Turk (AMT) workers. Detailed analysis shows that both label sets are equally reliable. The results support the use of crowd-sourced labels for speaker traits such as charisma in political speech, even where cultural subtleties are present. The impact of these different annotations on charisma prediction from political speech is also investigated.

Many paralinguistic challenges have looked at predicting affect, speaker state, or other attributes from short segments of speech of less than a minute. There are situations however, where we want to predict how a user might label a talk or lecture of significantly longer duration. For example, would a viewer find a given talk funny? The question then is how to map long talks to single word labels? In this paper, we rely on the concept of thin slicing, which states that humans make similar judgements on short segments of speech as they do on longer segments. We wish to find short segments that are representative of the talk, which can be used to predict the user label. We explore this concept in order to predict user ratings of TED talks as inspiring, persuasive, and funny. In particular, we pose two questions. The first is how thin can we make our slices? Results show that longer slices, of up to a minute in duration are more useful for the prediction of viewer ratings. We also ask where the best position to slice the video is? We compare the performance of classification based on slices extracted from fixed points to that of slices extracted from salient regions, and find that prediction accuracy can be improved by choosing slices according to the speaker's vocal behaviour or the audience's reactions.

No abstract