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

Researchers have detailed ParaBridge, a method designed to make speech language models reliably adjust their responses to paralinguistic cues such as tone of voice or background noise, rather than ignoring them in open-ended dialogue. The work, posted to arXiv on June 9, targets a known shortcoming in current speech language models (SLMs): they can recognize non-lexical information but frequently fail to let it shape their replies [1]. The authors note that a simple inference-time instruction scaffold can narrow this perception-behavior gap, indicating the relevant cues are already latent in the model. However, such scaffolds prove brittle when faced with multi-turn conversations or competing instructions [1]. ParaBridge addresses this by converting the scaffold into stable model behavior through an on-policy self-distillation process. During training, the scaffold acts as a temporary privileged view. The scaffold-free model generates its own response, while the scaffolded view supplies dense, full-vocabulary next-token targets along the model's trajectory. This approach teaches the model when non-lexical cues should influence its reply, without requiring curated dialogues, human labels, or external reward models [1]. When applied to the Qwen3-Omni-thinking backbone, ParaBridge lifted the scaffold-free Safe Action Rate on the VoxSafeBench benchmark from 14.6% to 40.3%. On the EchoMind evaluation, the average rating improved from 3.27 to 3.92. General capabilities remained stable: scores on MMAU-Pro, VoiceBench, and GPQA all stayed within 0.4 points of the original model [1]. The paper further reports that ParaBridge generalizes to paralinguistic cues not seen during training, transfers from safety-oriented training to empathy-oriented dialogue, and functions on a different SLM backbone [1]. The research bundle did not yield additional independent context on this specific method, as several supporting entries contained only platform navigation excerpts unrelated to the study [3][4][5], while others addressed unrelated topics such as the United Nations Sustainable Development Goals and transcription factors in molecular biology [6][7].