Moment-by-moment tracking of audience brain responses to an engaging public speech: Replicating the reverse-message engineering approach

ABSTRACTPublic speaking engages and entertains audiences. Through neuroimaging, we can examine responses to speeches in real time. Replicating an earlier study, this study carries out two kinds of analyses – forward and reverse correlations. First, we examine how the soundwave carrying the speech relates to brain responses, finding that bilateral auditory cortex responses track with the speech signal's energy. Second, we use the speech-evoked brain responses to reverse-identify salient moments in the speech. Specifically, we focus on the right temporoparietal junction (TPJ), a region associated with social cognition. We find that TPJ-peaks reverse-identify socially engaging content (defined by the ability to evoke laughter). These results demonstrate new ways to study the relationship between story content and the audience responses it evokes.KEYWORDS: Communication neurosciencestorysocial cognitionaudience engagementneuroimaging AcknowledgmentsWe thank the authors of the original study for making the data publicly available. We also thank the creators of the NiLearn and BrainIAK packages for neuroimaging data analysis and the developers of the pandas, seaborn, and Jupyter software packages. We acknowledge support of the high-performance computing cluster at Michigan State University (icer.msu.edu).Disclosure statementNo potential conflict of interest was reported by the author(s).Notes1 Readers may ask: Why not perform a 'forward correlation' as well? This has to do with the fact that while it is easy to quantify the over-time RMSE characteristic, it is more difficult to create a continuous theoretical measure of a story's social content or demand – however defined. In the current study, we will also carry out 'forward correlation analyses' using coded content, but the main goal of this study was to replicate the reverse correlation procedure from the rTPJ, which relies on instances in which the audience responds with laughter to the speakers' story.2 In theory, one could also analyze the sentiment of these words via NLP tools. However, this is not very promising for this story because the narrative-level information develops based on the word context and is not a property of the single words.3 One option could be to simply drop the t-thresholding procedure and work with a fixed set of peaks/troughs (e.g., the top 10) and a fixed window (e.g., ±5 s) around those peaks. An alternative option would be to present results after running parameter sweeps (e.g., varying the t-threshold continuously and keeping track of the results). Together with these refinements, one might then also compute parameters like precision and recall (see Supplementary Results).4 That is, in terms of proximal hierarchical processes (Colley, Citation1961) or different goals, which can differ by the context of the genre (e.g. health: to inform and influence; entertainment: to produce laughter or intellectual stimulation).5 We thank an anonymous discussant for suggesting this potential to us.