A Shaky Voice Is Not Always a Dodge: Benchmarking Textual and Vocal Evasion Detection in Earnings Calls
Computation and LanguageSound
Summary
The authors look at how company leaders sometimes avoid giving clear answers during earnings calls, not just by the words they use but also by how they say them. They created a dataset called DualEvasion that labels both the text and the speaker's confidence from audio to study this. Their tests show that current AI models find it hard to recognize when someone sounds unconfident, often missing the context of the speaker's usual behavior. Using comparisons to each speaker's normal tone helps a bit, but the models still don't do as well as humans. This work highlights the challenge of combining speech and text to detect evasive communication.
evasion detectionearnings callstextual transcriptsvocal confidencemultimodal modelsacoustic cuesspeaker baselinequestion-answer pairsannotated datasetaudio analysis
Authors
Mirae Kim, Seonghun Jeong, Youngjun Kwak
Abstract
Existing approaches to evasion detection in earnings calls focus on textual transcripts, treating evasion as a single-dimensional phenomenon. We argue that evasion in spoken communication is inherently multidimensional: beyond what executives say, how they say it carries independent and complementary information. To study these dimensions jointly, we introduce DualEvasion, a benchmark for evasion detection across text and audio in earnings call Q&A. The benchmark contains 505 annotated question-answer pairs from 60 earnings calls, each with two independent labels: textual evasion (direct vs. evasive) and vocal cues operationalized as speaker confidence (confident vs. unconfident). Our experiments show that state-of-the-art multimodal models struggle to detect vocal confidence, particularly on unconfident responses. Our analysis suggests these models interpret acoustic cues in isolation rather than relative to each speaker's baseline. Providing speaker-level references yields modest improvements, but a substantial gap with human performance remains.