Preprint
Conversational Speech Reveals Structural Robustness Failures in SpeechLLM Backbones
Maria Teleki et al., arXiv, 2026
Read the originalRead on October 17, 2026
- What it is
- A study of how the language models behind speech systems handle conversational speech with interjections, edits and parentheticals, which a clean transcript should remove. It measures how models trade off removing disfluencies against deleting fluent content.
- How much weight it carries
- A preprint (version 2, March 2026). I read the abstract, not the full results tables. It tests models on conversational input, so it describes how editing models behave on transcripts, not how a specific dictation app behaves.
Facts taken from this source
- In a 2026 preprint on conversational speech, reasoning models systematically over-deleted fluent content, which the authors read as a bias toward abstraction over fidelity to the transcript.
“reasoning models systematically over-delete fluent content, revealing a bias toward semantic abstraction over structural fidelity” (Abstract)