ParaBridge: Bridging Paralinguistic Perception and Dialogue Behavior in Speech Language Models

In three linesParaBridge is an on-policy self-distillation method that teaches Speech Language Models to use paralinguistic cues (tone, emotion, noise) in dialogue. On Qwen3-Omni-thinking, it raises VoxSafeBench SAR from 14.6% to 40.3% and improves EchoMind from 3.27 to 3.92, while preserving general abilities.

Summary generated by Claude — human-verified