Full-duplex spoken language models—systems that can speak and listen simultaneously—must decide at every frame whether to take the floor. That decision shapes a conversational personality, yet each set of model weights encodes exactly one such style. Human speakers, by contrast, bracket rather than share a single style, suggesting a need for flexibility.
The paper introduces a reference-guided approach to turn-taking style transfer. Instead of retraining a model from scratch, the method uses a reference speaker's style to guide the model's behavior, allowing the same underlying weights to shift their turn-taking patterns. The abstract notes that existing approaches fall short, but the details of the proposed technique are not fully described in the available text.
If successful, this could make spoken AI more adaptable—letting a single model adopt different conversational personalities on demand, rather than being locked into one. The work points toward more natural, human-like interaction in full-duplex systems.