TAMPA 鈥� A new study co-written by 最新天美传媒 accounting assistant professor Sung-Yuan (Mark) Cheng introduces a novel way to identify signs of emotional strain among corporate leaders 鈥� not through surveys or interviews, but through the sound of their voice.
Published in the Journal of Accounting Research, the article, 鈥�,鈥� shows how artificial intelligence can analyze subtle vocal patterns in CEO earnings calls to estimate whether a leader is showing vocal indicators associated with depressive symptoms.
The research does not diagnose clinical depression but instead uses machine learning to detect signals that are statistically related to depression in clinical voice data.
鈥淥ur model is informative, but far from perfect,鈥� said Cheng, a faculty member in the Lynn Pippenger School of Accountancy at the USF Muma College of Business. 鈥淚t performs better than chance in clinical settings, but it still produces false positives and false negatives. It is useful for large-sample research, not for diagnosing individuals.鈥�
Cheng noted that the idea for the study originated from his co-author, Nargess Golshan at Indiana University, who listened to a podcast discussing how some medical professionals can use patient voice recordings to help diagnose whether patients have certain mental health issues.
He emphasized that applying the model to CEOs should be viewed as exploratory, because real-world speech differs from clinical voice recordings, which is collected from patients in a healthcare setting.
The team trained its model on using audio from patients who completed professional mental-health assessments. The machine-learning system, based on a large audio model originally developed by Google, learned patterns in speech that tend to show up when people experience depressive symptoms.
Cheng designed an algorithm to isolate only the CEO鈥檚 voice from more than 14,500 S&P 500 earnings call recordings, which also include chief financial officer comments and analyst questions.
鈥淭hat was a key challenge,鈥� he said. 鈥淓arnings calls include many speakers, and we needed only the CEO鈥檚 portion.鈥�
Once the model was trained, the team applied it to audio from CEOs to generate a continuous measure of depressive signals 鈥� a statistical estimate based on vocal patterns.
The study uncovered several notable patterns:
- Depressive signals appear relatively common among CEOs, which the researchers believe reflects the intense pressures of top leadership roles.
- CEOs show higher vocal depressive signals when their companies face greater uncertainty, such as lawsuits, volatile stock returns or disappointing financial results.
- CEOs with stronger depressive signals often receive higher total compensation and pay-performance sensitivity.
Cheng said this compensation pattern could have several explanations. One is that boards, whether consciously or not, may offer stronger financial incentives to retain or support a CEO who appears to be under strain.
Another is that high-pressure corporate environments can both increase emotional burden and lead boards to use stronger incentives to keep a company on track. He added that the study cannot determine intent or causality.
Our model is informative, but far from perfect. It is useful for large-sample research, not for diagnosing individuals."
Mark Cheng
The researchers also found that female and older CEOs appear less likely to show depressive signals. Cheng said that may reflect differences in who reaches the CEO role. For example, women or older executives who make it to the top may be exceptionally resilient. It could also reflect model bias, since voice-based AI can perform differently across demographic groups depending on how well those groups were represented in the training data.
鈥淭hese patterns are descriptive, not definitive,鈥� Cheng said.
One of the most surprising findings came from company performance 鈥� despite higher depressive signals, CEOs did not perform worse on average.
Cheng urged caution in interpreting this result: 鈥淎 lack of statistical difference doesn鈥檛 prove there is no effect. It could reflect measurement noise or the fact that organizations buffer variation in their leaders鈥� emotional states. Depressive signals in voice don鈥檛 automatically translate into lower performance.鈥�
Co-author Nargess Golshan, of Indiana University, noted in an interview with Fortune that the research highlights the importance of discussing mental health in leadership, a topic that still carries stigma.
鈥淲e want to start a conversation and help executives be aware of it, and help companies support their leaders,鈥� she said.
The findings offer a new way to understand the emotional pressures facing executives, while also showing how AI can analyze forms of information, like voice, that were previously difficult to study at scale.
He hopes the research encourages boards and companies to take mental health seriously as part of responsible governance and long-term organizational health.
The paper has captured national attention from outlets including The Economist and Fortune, signaling the growing interest in how AI can support mental-health awareness in high-pressure environments.
鈥淓xecutives are still human,鈥� Cheng said.