Fact · Voice and AI
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.
Checked against its source on October 17, 2026
In the source's own words
“reasoning models systematically over-delete fluent content, revealing a bias toward semantic abstraction over structural fidelity”
Conversational Speech Reveals Structural Robustness Failures in SpeechLLM Backbones, Maria Teleki et al., arXiv, 2026, Abstract.
Open the original- What it measures
- The models were scored on how they edit conversational transcripts that contain interjections, edits and parentheticals.
- What it does not show
- A preprint; the abstract reports the pattern and the full paper has the tables. It concerns models given conversational input, not one dictation app.