The Sound of Confidence: Can Machines Perceive Nonverbal Signals in Vocal Delivery?

Abstract

This paper examines whether artificial intelligence can approximate human perception of vocal confidence in managerial speech and whether this nonverbal communication cue affects capital market outcomes. The paper develops DeepConfidence, a domain-specific deep learning model that measures perceived vocal confidence from raw audio recordings of earnings conference calls. The model combines self-supervised pre-training on 500,000 earnings call sentences with supervised fine-tuning on 5,000 human-labeled sentences classified as confident, neutral, or unconfident. DeepConfidence outperforms heuristic measures based on interjections, a wav2vec 2.0 benchmark without domain-specific pre-training, and state-of-the-art generative large multimodal models. Applying the model to 21,347 earnings conference calls from US-listed firms, the paper documents that managers sound substantially more confident during prepared presentations than during interactive Q&A sessions. This difference becomes particularly pronounced during periods of heightened uncertainty, such as the Covid-19 pandemic. In capital market tests, vocal confidence during prepared remarks is associated with significantly positive stock price reactions around the call, incremental to earnings news, linguistic content, and other vocal cues. Overall, the findings suggest that domain-specific AI models can capture salient nonverbal signals in managerial communication and that vocal confidence constitutes an economically meaningful disclosure cue.