EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

When Students Are Misheard: Role-Aware Robustness of LLM-Based Classroom Discourse Analysis to ASR Errors

Christopher Hale

International Journal of Business and Social Science Studies · 17.09.2026 · 6 (2) · 255-237

DOI: 10.66795/ijbsss74 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Business and Social Science Studies, online erschienen am 17.09.2026. Verfasst von Christopher Hale. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Large language models (LLMs) are increasingly used to code classroom discourse, yet they are developed on human transcripts while deployed systems must rely on automatic speech recognition (ASR). We quantify how recognition errors propagate into LLM-based talk-move analysis through a controlled acoustic-to-analytics protocol built on the TalkMoves corpus. We re-voice 910 held-out teacher and student utterances and their preceding turns with a multi-speaker synthesizer and child-like vocal-tract scaling, render them through close-talk, teacher-lapel, and room-microphone conditions with statistical room acoustics and classroom babble, and transcribe them with three open ASR engines, which yields nine configurations with word error rates (WER) from 5.9% to 86.8%. An on-premise classifier built on a compact instruction-tuned LLM loses about 5.6 (teacher) and 4.9 (student) points of F1 retention per 10 points of WER. With a lapel microphone it keeps 95.9% of its teacher-move F1 but only 86.0% of its student-move F1, and moves that hinge on short cue words or on repeating a peer's words degrade most: under a room microphone, pressing for reasoning is detected in only 5.2% of cases once its question word is misrecognized. Posterior fusion across the three recognizers raises macro-F1 over the average engine by up to 6.6 points, and calibrated error-channel noise-aware training, fitted on 250 transcribed utterances, adds 4.3 points of student-move F1 under a room microphone when combined with fusion. The results support role-stratified auditing of discourse analytics and investment in student-side audio capture.

Zitieren

APA

Hale, C. (2026). When Students Are Misheard: Role-Aware Robustness of LLM-Based Classroom Discourse Analysis to ASR Errors. International Journal of Business and Social Science Studies, 6(2), 255-237. https://doi.org/10.66795/ijbsss74

Vancouver

Hale C. When Students Are Misheard: Role-Aware Robustness of LLM-Based Classroom Discourse Analysis to ASR Errors. International Journal of Business and Social Science Studies. 2026;6(2):255-237. doi:10.66795/ijbsss74

BibTeX
@article{Hale2026,
  author = {Hale, Christopher},
  title = {When Students Are Misheard: Role-Aware Robustness of LLM-Based Classroom Discourse Analysis to ASR Errors},
  journal = {International Journal of Business and Social Science Studies},
  year = {2026},
  volume = {6},
  number = {2},
  pages = {255--237},
  doi = {10.66795/ijbsss74}
}
Kostenlos starten →