Stable Performance Under High-Severity Disfluency

Develop methods that enable Speech Large Language Models (Speech-LLMs) to maintain stable performance under high-severity disfluency conditions.

Background

Through fine-grained severity analyses, the paper shows that many models handle mild disfluency but degrade sharply at higher severity levels, particularly for phoneme-level repetition and long-context disruptions. This reveals robustness limitations that persist despite improvements in components such as ASR and LLM reasoning.

The authors explicitly identify achieving stability at high severity as an unresolved challenge, underscoring the need for targeted modeling, training objectives, and system design to ensure reliable performance under severe disfluency.