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

ParaBridge is a novel on-policy self-distillation method designed to enhance Speech Language Models (SLMs) by effectively integrating paralinguistic cues into dialogue behavior. It improves the scaffold-free VoxSafeBench SAR from 14.6% to 40.3% and raises EchoMind's average rating from 3.27 to 3.92 while maintaining general performance across other benchmarks. This approach allows models to adapt to unseen paralinguistic cues and transfer learning from safety to empathy-oriented dialogues, offering practitioners a robust method for developing more contextually aware SLMs without reliance on curated datasets.