Beyond manual transcripts: Exploring the potential of automatic speech recognition errors in improving Alzheimer's disease detection.

OBJECTIVE

This study aims to extend the counterintuitive observation that Automatic Speech Recognition (ASR) errors can be beneficial for Alzheimer's Disease (AD) detection. Our objective is to conduct a large-scale investigation to validate this phenomenon and, more importantly, to elucidate the specific mechanisms by which ASR errors can serve as valuable diagnostic clues for distinguishing individuals with AD from Healthy Controls (HC).

METHODS

We employed 18 ASR models, in both their original and fine-tuned versions, to generate 36 sets of transcripts from the ADReSS dataset. We also synthesized speech from both manual and ASR transcripts using a text-to-speech (TTS) model. Knowledge-based features and pre-trained embeddings were extracted and fed into two proposed AD detection models : a self-attention model and a cross-attention-based interpretability model. To uncover the underlying mechanisms, we conducted a multi-faceted set of analyses, including examinations of ASR error types, words affected by ASR errors, linguistic comparisons, attention weight analysis, and case studies.

RESULTS

We demonstrate that transcripts generated by certain ASR models achieve higher AD detection accuracy than gold-standard manual transcripts. This performance gain stems not from errors in general or a high Word Error Rate (WER), but from specific and asymmetric error patterns. Our analyses reveal that these patterns amplify some pre-existing linguistic deficits in AD speech (e.g., disfluencies), thereby increasing the feature-level divergence between the AD and HC groups. Furthermore, we show that these diagnostic clues are effectively preserved when speech is synthesized from ASR transcripts, holding significant implications for data augmentation strategies in AD research.

CONCLUSION

The specific, asymmetric error patterns introduced by certain ASR models enhance the distinction between AD and HC groups by amplifying pathological linguistic deficits associated with AD. This work suggests a paradigm shift for clinical ASR development: optimizing models not merely for transcription accuracy, but for their downstream diagnostic utility.