ParaBridge: Bridging Paralinguistic Perception and Dialogue Behavior in Speech Language Models
- lab CatalyzeX
- lab DagsHub
- lab GotitPub
- lab Hugging Face
- lab ScienceCast
- lab arXiv
- lab arXivLabs
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].
infrastructuremodel-releaseresearch-paperfundingregulation
Background sources we checked (6)
- arxiv.org ↗ Speech carries more information than just words: a child's voice, a fearful tone, or a noisy background should all lead a sufficiently competent spoken-dialogue assistant to different replies. Current Speech Language Models (SLMs) can recognize such paralinguistic cues but often …
- arxiv.org ↗ CatalyzeX Code Finder for Papers (What is CatalyzeX?) [...] DagsHub Toggle [...] DagsHub (What is DagsHub?)…
- arxiv.org ↗ CatalyzeX Code Finder for Papers (What is CatalyzeX?) [...] DagsHub Toggle [...] DagsHub (What is DagsHub?)…
- arxiv.org ↗ CatalyzeX Code Finder for Papers (What is CatalyzeX?) [...] DagsHub Toggle [...] DagsHub (What is DagsHub?)…
- en.wikipedia.org ↗ Sustainable Development Goals (abbr. SDGs) were adopted in 2015 by all United Nations (UN) members for the 2030 Agenda for Sustainable Development. The aim of the 17 global goals is "peace and prosperity for people and the planet", tackling climate change, and working to preserv…
- en.wikipedia.org ↗ In molecular biology, a transcription factor (TF) (or sequence-specific DNA-binding factor) is a protein that controls the rate of transcription of genetic information from DNA to messenger RNA, by binding to DNA sequences. Specificity can be due to sequence motifs, or epigenetic…