When Voice Adds to Earnings-Call Sentiment

Investors increasingly use natural language processing (NLP) sentiment to extract signals from earnings calls. David Pope, author of the June 2026 study “What NLP Sentiment Can’t Hear—How Vocal Delivery Improves Earnings-Call NLP Sentiment,” makes the point that words alone do not tell the whole story. Vocal delivery can add economically meaningful information beyond the words themselves.

Pope looked at 41,395 Russell 3000 earnings-call observations from July 2020 through September 2025 and asked a practical question: Once text sentiment is already positive or negative, does vocal delivery improve the signal? The answer is yes, and the strongest evidence is on the downside.

Why the Study Matters

Text-based sentiment analysis has become a standard tool because it is scalable and consistent. But earnings calls are live events, not static documents. Management is speaking under pressure, and voice may reveal whether the language sounds controlled, confident, strained, or merely scripted.

That distinction is important because investors often assume that a positive or negative sentiment score is enough. This paper argues that it is not. Voice can act as a second filter that helps investors rank the quality of the same textual signal.

What Pope Examined

Pope sorted earnings-call observations into quintiles by textual sentiment and then into quintiles by vocal measures. The key voice variables were:

· Vocal Strain: Intended to capture a broad strained-delivery state rather than nervousness alone. It increases when management sounds more nervous, uncertain, or activated, and decreases when management sounds more assertive, positive in vocal affect, and balanced in delivery.

· Balanced Delivery: Captures whether management’s vocal delivery is more controlled, measured, and stable. Balanced Delivery is used as an independent check on the Vocal Strain result. If negative text is more concerning when delivered under strain, it should also be more concerning when delivered with lower Balanced Delivery.

· Vocal Valence: Captures the positive or negative affective quality of the speaker’s vocal delivery. In the positive-text tests, Vocal Valence is used to examine whether favorable language becomes more informative when it is delivered with a more positive vocal tone.

The main design held text constant and asked whether returns changed when delivery changed. For negative text, the study compared high versus low Vocal Strain and low versus high Balanced Delivery. For positive text, it asked whether high Vocal Valence improved the return profile of positive language.

Key Findings