Full-duplex spoken language models (FD-SLMs) are designed to listen and speak at the same time, enabling more natural, interruptible conversation. However, according to a new preprint on arXiv, these models often lag behind half-duplex spoken language models (HD-SLMs) in general capabilities, which typically operate in alternating turns.
To address this, the paper introduces Dialogue-to-Duplex (D2), a framework for building high-capability full-duplex models from half-duplex backbones. The abstract does not detail the inner workings of D2, but the framing suggests that the strengths of half-duplex models can be transferred to the full-duplex setting rather than training full-duplex systems from scratch.
The work highlights a central trade-off in spoken language AI: real-time interactivity versus broad competence. If D2 succeeds, it could point toward full-duplex systems that do not have to sacrifice general performance for conversational fluidity. As with any preprint, the claims have not yet been peer-reviewed.