DualTurn: Learning Turn-Taking from Dual-Channel Generative Speech Pretraining

Researcher Shangeth Rajaa has introduced DualTurn, a novel AI architecture designed to solve the unnatural turn-taking problem in modern voice AI pipelines. Current systems face a trade-off: end-to-end speech-to-speech models handle conversational flow naturally but lack tool-calling capabilities, while production pipelines that chain Automatic Speech Recognition (ASR), Large Language Models (LLMs), and Text-to-Speech (TTS) rely on clunky silence timeouts. DualTrain bridges this gap through a unique two-stage training process. First, its 0.5B parameters are generatively pre-trained on dual-channel conversational audio, forcing the model to predict both speakers' future audio and implicitly learn the dynamics of dialogue without any labeled data.

This foundational understanding is then fine-tuned to produce interpretable, actionable signals. Instead of just detecting silence, DualTurn continuously monitors both audio channels to anticipate turn boundaries and outputs one of five specific agent actions (like 'hold', 'speak', or 'listen'). The results are striking for a model of its size. On standard benchmarks, it significantly outperforms the Voice Activity Projection (VAP) model on agent action prediction, achieving a weighted F1 score (wF1) of 0.633 versus VAP's 0.389. It also beats a much larger 3.1B parameter audio-text model on word-level turn prediction, with an Area Under the Curve (AUC) of 0.930 compared to 0.880. Crucially, it achieves this with earlier anticipation of turn boundaries, leading to fewer awkward interruptions and more human-like conversational flow. The work, submitted to Interspeech 2026, represents a promising step toward voice agents that can reason and use tools without sacrificing natural interaction.