DuplexPO uses RL to decouple conversational timing from reasoning in AI voice models

Recent full-duplex spoken dialogue models have made strides in human-like interaction—responding with low latency, inserting backchannels, and handling barge-ins. However, these conversational improvements often come at the cost of weaker reasoning and instruction-following, creating an apparent trade-off. A new paper from researchers argues this trade-off isn't fundamental. They introduce DuplexPO, a reinforcement learning framework that separates conversational dynamics from semantic reasoning. DuplexPO learns a temporal decision policy from human dialogue data, optimizing when the model speaks without compromising what it says. The framework uses a Factorized Conversational Dynamics Reward (FCDR) for fine-grained credit assignment on turn initiation, backchanneling, yielding, and participation, then optimizes the policy with a GRPO-style objective.

In experiments, DuplexPO substantially improved full-duplex behaviors including timely backchannels, smooth turn-taking, and barge-in handling, while preserving strong reasoning and instruction-following. The researchers found that improvements in dynamics-oriented metrics directly translated to better user experience, suggesting that optimizing conversational timing as a standalone objective can promote more natural interaction. This decoupling approach could unlock more fluid and intelligent voice assistants that feel both responsive and smart, without forcing an unnatural trade-off between speed and substance.