The sound of cooperation and deception when stakes are high

This paper examines whether acoustic features of the human voice contain information about cooperative and deceptive behavior in high-stakes environments. We analyze two naturalistic datasets in which speech is produced under meaningful incentives: (i) the British TV game show “Golden Balls” and (ii) courtroom testimonies from “Real-life Trial”. Using machine-learning models estimated separately by speaker gender, we find that voice-only models consistently outperform benchmark predictions, and Shapley-based decompositions show that vocal cues contribute a substantial and nonredundant share of predictive performance relative to text. Logistic regression analyses identify specific acoustic markers, such as pitch, intonation, and harmonics-to-noise ratio, associated with cooperative and deceptive behavior that differ across contexts and genders. Finally, cross-context prediction shows voice features have limited transferability. Our results highlight voice as a scalable and informative source of process data for behavioral economics and field research.