Gender and Tone in Recorded Economics Presentations: Audio Analysis with Machine Learning

Abstract

This paper develops a replicable and scalable method for analyzing tone in economics seminars to study the relationship between speaker gender, age, and tone in both static and dynamic settings. We train a deep convolutional neural network on public audio data from the computer science literature to impute labels for gender, age, and multiple tones, like happy, neutral, angry, and fearful. We apply our trained algorithm to a topically representative sample of presentations from the 2022 NBER Summer Institute. Overall, our results highlight systematic differences in presentation dynamics by gender, field, and format. We find that female economists are more likely to speak in a positive tone and are less likely to be spoken to in a positive tone, even by other women. We find that male economists are significantly more likely to sound angry or stern compared to female economists. Despite finding that female and male presenters receive a similar number of interruptions and questions, we find slightly longer interruptions for female presenters. Our trained algorithm can be applied to other economics presentation recordings for continued analysis of seminar dynamics.

Keywords: machine learning, audio analysis, gender, seminar dynamics, tone

JEL Classification: A1, C8, C45