TurnGuide: Enhancing Meaningful Full Duplex Spoken Interactions via Dynamic Turn-Level Text-Speech Interleaving

TurnGuide is a novel approach for enhancing Full-Duplex Speech Language Models (FD-SLMs) by implementing dynamic turn-level text-speech interleaving, which allows for improved modeling of conversational turn-taking. This method addresses the challenges of integrating discrete text tokens into continuous audio streams, enabling more natural and coherent spoken interactions while maintaining the acoustic flow. Experimental results indicate that TurnGuide achieves state-of-the-art performance in generating semantically meaningful speech across various turn-taking scenarios, making it a valuable advancement for practitioners working on real-time spoken dialogue systems.