Evaluating the Safety Tolerances of Speech Recognition AI for Look-Alike/Sound-Alike (LASA) Clinical Nomenclature under Acoustic Stress

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This study evaluates how open-source automatic speech recognition (ASR) models handle critical medical terminology when exposed to background noise, simulating the conditions of clinical transcription software used in hospitals. Software developers typically evaluate ASR models using an unweighted average Word Error Rate (WER). However, in high-acuity healthcare environments, aggregate metrics can hide dangerous, localized errors, such as the incorrect substitution of Look-Alike/Sound-Alike (LASA) medication names [1]. Using an automated Python testing pipeline, OpenAI’s Whisper-Tiny model [2] was tested against clinical handover audio files overlaid with three tiers of Gaussian white noise (0%, 10%, and 30% noise amplitude configurations). The experimental results demonstrate a non-linear performance collapse under severe acoustic stress. Rather than showing a gradual decrease in accuracy, the model introduced critical phonetic mutations at low noise levels (e.g., "humalog" transcribed as "humilog") and entered infinite autoregressive repetition loops under high noise (30%), causing localized error rates to exceed 1200%. These findings indicate that standard commercial evaluation metrics fail to isolate safety risks in clinical deployment, highlighting the need for specialized validation frameworks.