Measuring Vocal Tone in Corporate Disclosures

We examine the usefulness of machine learning approaches for measuring vocal tone in corporate disclosures. We document a substantial mismatch between the widely adopted actor-based training data underlying these approaches and speech in corporate disclosures. We find that existing models achieve near-perfect vocal tone classification within their training domain. However, when tested on actual executive speech during conference calls, their performance declines to chance levels. We thus introduce FinVoc2Vec, a deep learning model that adapts to audio recordings of conference calls and classifies the vocal tone of executive speech significantly more accurately than chance. FinVoc2Vec estimates are associated with future firm performance and can be used to construct profitable stock portfolios. Throughout our analyses, estimates from previous vocal tone models are largely unrelated to firm performance. Our findings emphasize the importance of a domain-specific approach to voice analysis in accounting and finance.