TRACE: Temporal Relationship-Aware Conversational Entrainment Detection in Dyadic Speech

A research team has proposed TRACE, a new framework for detecting emotional entrainment in dyadic speech, reporting a 97.01% accuracy rate on a purpose-built dataset called DyadEE, according to a preprint posted to arXiv on 29 June 2026 [1][2]. The work, authored by Sathvik Manikantan Napa Ugandhar, addresses the growing need to understand affective coordination as speech AI agents become more common [2]. Emotional entrainment—the synchronization of emotional states between speakers—is shaped by social relationships and conversational context, the paper states [1][2]. The DyadEE dataset was constructed to support this research, containing both naturally entrained conversations and synthetic interactions where entrainment is deliberately disrupted through partner swapping and emotion resynthesis [2]. TRACE models dyadic interaction as ordered sequences of acoustic embeddings, treating each sample as an interaction trace rather than a collection of pooled utterances [1][2]. The embeddings are derived from a version of OpenAI’s Whisper model that has been fine-tuned for emotion recognition [1][2]. On the DyadEE benchmark, the framework achieved its best accuracy of 97.01% [1][2]. The submission file was listed at 1,289 KB [1]. The preprint appeared on arXiv, an open-access repository that hosts e-prints across physics, computer science, and related fields without peer review [6]. As of late 2024, the platform was receiving roughly 24,000 new articles per month [6]. The paper’s landing page also surfaces a set of experimental community tools developed under the arXivLabs framework, which allows third-party collaborators to build features such as citation explorers and recommender systems directly on the site [4][5]. arXivLabs was formalized in 2020 to provide guidelines for collaborations that share the repository’s values of openness, community, excellence, and user data privacy [5].