Abstract
Telemedicine systems rely on AI-based Automatic Speech Recogniton (ASR) modules for clinical documentation, yet audio codec selection, a frequently overlooked infrastructure decision, can silently degrade transcription accuracy. Framed as a design science study, this work introduces a codec admissibility framework: a prescriptive Information Systems (IS) artefact that classifies audio codecs by their measured impact on Whisper large-v3 transcription of Polish medical speech. Using 2,000 annotated recordings across seven medical domains, we computed bootstrapped 95% confidence intervals for ΔWord Error Rate (WER) under three normalization strategies. AAC, MP3, Opus, and ADPCM cause no statistically significant degradation, while G.723, Speex, and the neural codec produce significant increases in ΔWER of up to 0.230 (95% bootstrapped Confidence Intervals (CI): [0.181, 0.230]) The resulting codec selection decision table provides concrete non-functional requirements for IS architects designing ASR-integrated telemedicine platforms.
Paper Type
Poster
DOI
10.62036/ISD.2026.46
Analysis of the Efficiency of Polish Medical Terminology Recognition by Whisper ASR System Depending on the Selected Audio Codecs
Telemedicine systems rely on AI-based Automatic Speech Recogniton (ASR) modules for clinical documentation, yet audio codec selection, a frequently overlooked infrastructure decision, can silently degrade transcription accuracy. Framed as a design science study, this work introduces a codec admissibility framework: a prescriptive Information Systems (IS) artefact that classifies audio codecs by their measured impact on Whisper large-v3 transcription of Polish medical speech. Using 2,000 annotated recordings across seven medical domains, we computed bootstrapped 95% confidence intervals for ΔWord Error Rate (WER) under three normalization strategies. AAC, MP3, Opus, and ADPCM cause no statistically significant degradation, while G.723, Speex, and the neural codec produce significant increases in ΔWER of up to 0.230 (95% bootstrapped Confidence Intervals (CI): [0.181, 0.230]) The resulting codec selection decision table provides concrete non-functional requirements for IS architects designing ASR-integrated telemedicine platforms.
Recommended Citation
Zaporowski, S. & Kurowski, A.(2026). Analysis of the Efficiency of Polish Medical Terminology Recognition by Whisper ASR System Depending on the Selected Audio Codecs. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.46