Abstract

Automatic Speech Recognition (ASR) plays an important role in speech-basedautomatic detection of Alzheimer's disease (AD). However, recognition errorscould propagate downstream, potentially impacting the detection decisions.Recent studies have revealed a non-linear relationship between word error rates(WER) and AD detection performance, where ASR transcriptions with notableerrors could still yield AD detection accuracy equivalent to that based onmanual transcriptions. This work presents a series of analyses to explore theeffect of ASR transcription errors in BERT-based AD detection systems. Ourinvestigation reveals that not all ASR errors contribute equally to detectionperformance. Certain words, such as stopwords, despite constituting a largeproportion of errors, are shown to play a limited role in distinguishing AD. Incontrast, the keywords related to diagnosis tasks exhibit significantly greaterimportance relative to other words. These findings provide insights into theinterplay between ASR errors and the downstream detection model.