Evaluating Large Language Models Abilities for Addressee, Turn-change, and Next Speaker Prediction in Meetings

Large language models can predict who will speak next in a meeting more accurately than humans, according to a study posted to the arXiv preprint repository on 16 June 2026. The models achieved this despite never being trained on meeting data and without access to audio or visual cues. The research, which has not yet been peer-reviewed, evaluated LLMs on three aspects of conversational turn-taking: detecting whom an utterance is addressed to, predicting whether a speaker change will occur, and identifying the next speaker [1][2]. Experiments used the AMI corpus, a standard dataset of multimodal meeting recordings [1][2]. Text-based LLMs outperformed both supervised models trained specifically for the tasks and human subjects on next speaker prediction [1][2]. The authors noted that the models succeeded even though they lacked domain-specific training and could not process audio or visual information [2]. A multimodal LLM that could access raw audio-visual signals performed better than text-only models on addressee detection and turn-change prediction, but its accuracy remained below that of humans [1][2]. The researchers wrote that this gap indicates the model had difficulty leveraging the additional sensory data [2]. Ablation analyses showed that conversational context was the most important factor for next speaker prediction [1][2]. The study also found that human and LLM prediction patterns were similar, and that both struggled during stretches of conversation with frequent turn changes [2]. The paper appeared on arXiv, an open-access repository that hosts electronic preprints across physics, mathematics, computer science, and other fields [6]. Founded in 1991, arXiv passed the two-million-article milestone by the end of 2021 and now receives roughly 24,000 submissions per month [6]. Submissions are moderated but not peer-reviewed before posting [6]. The study was shared through arXivLabs, a framework launched in 2020 that allows community collaborators to build experimental tools and features on top of the repository’s article pages [4][5]. arXivLabs projects must adhere to the repository’s stated values of openness, community, excellence, and user data privacy [4].