AI-assisted laughter detection software v. human coders: a new way to collect objective data about stand-up comedy performances?

Audience laughter is central to the interaction between stand-up comedians and their audiences, but systematic human coding of laughter is labour intensive. To address this challenge, this paper evaluates the performance of an AI-assisted tool that measures the laughter found in recordings of stand-up comedy performances, the StandApp Comedy App. We present a comparison of the laughs per minute data from our hand-coded study and the data produced by the app, discuss some of the current limitations of the app, and explain why, despite these limitations, we believe the application provides valuable information to both performers and people who teach stand-up comedy. The comparison between human coding and AI-assisted analysis produced broadly similar results. Specifically, there was only a .51 difference between AI and human coding at Time 1 beginning of semester rates (derived by subtracting 4.05-3.54 laughs per minute [in seconds]), and a .36 difference at T2 end of semester rates (derived by subtracting 9.29-8.93 laughs per minute [in seconds]). These negligible or small mean differences were also supported by non-significant t-test results, Bland–Altman, and ICC analyses. Within the context of analysing live audience responses, where variability and interpretive judgement are common, this level of discrepancy appears relatively small and suggests that AI-assisted methods may provide an efficient and reasonably accurate approximation of human-coded outcomes for comedy teachers and performers.