Full-duplex speech models need RL-based alignment beyond standard training to handle natural conversation dynamics—pauses, turn-taking, and interruptions—without degrading response quality.
This paper improves full-duplex speech models (which listen and speak simultaneously) by using reinforcement learning to optimize four key conversational behaviors: pauses, turn-taking, backchanneling, and handling interruptions. Rather than just maximizing word prediction accuracy, the method trains models with specific reward signals for each interaction type, while preserving response quality.