SteerDuplex Enables Instruction-Following in Full-Duplex Speech Models
September 10, 2026
SteerDuplex is a Moshi-based full-duplex speech model fine-tuned via two-stage reinforcement learning to support steerability. It allows users to shift conversational attributes like tone, persona, and speaking rate during low-latency, interruptible dialogue.
HOW THIS AFFECTS YOU
●
builderYou can build more natural voice assistants that respond to stylistic instructions in real-time.
●
designerThis enables new UX patterns for highly expressive and steerable conversational agents.