arXiv Preprint: 'Initiative Vector' Identified in Full-Duplex Voice Models Enables Reference-Guided Style Transfer

Researchers have identified a single late-layer residual-stream direction, termed the "initiative vector," within full-duplex spoken language models that determines when the model takes the floor in a conversation.

While existing methods control conversational personality from the outside, this study reveals an internal quantity that sets how much the model speaks. Experiments show that scaling this vector shifts the speech ratio from a baseline of 21-30% to either 6-12% or 42-61%, with consistent effects across various conversations, model weights, and corpora.

Crucially, a short reference recording is sufficient to transfer a specific speaker's turn-taking style to the model. The transfer is highly reference-specific: swapping in another speaker's reference worsens the match in 93-98% of cases, indicating that style is not frozen into the weights but can be read, shifted, and calibrated per person.

These findings rely on first-party benchmarks and have not yet been independently reproduced. The work highlights a new mechanism for controlling AI voice interaction styles without retraining, though real-world stability in complex environments remains to be verified.