4,068 posters, 219 videos, 101 topics, 69 sessions, 14,365 authors, 6,128 institutions
ePostersLive by SciGen Technologies S.A. All rights reserved.
Alzheimer's Association International Conference 2026
July 12-15, 2026 | London, United Kingdom, and Online
4494
Evaluating Automatic Speech Recognition Systems For Dementia Speech: A Multi-metric Approach
Part of Topic
•ASR is increasingly used for speech-based digital biomarkers, but models are trained on clean, fluent speech.
•Dementia speech is atypical, with disfluencies, pauses, fragments and agrammatism that carry clinical signal.
•Word Error Rate (WER) is surface-level: it counts token errors but ignores syntax, meaning and disfluency.
•WER weights all errors equally, and two systems with identical WER can differ greatly in clinical utility.
•ASR is worse for dementia speech across every metric, even after controlling for task, length, speaking rate and disfluencies (WER β = 0.34, p < 0.001).
•AWS and CrisperWhisper are most reliable (CrisperWhisper WER 16.1% dementia / 6.8% healthy); Google and Meta are weakest.
•System rankings change with the metric: AssemblyAI ranks 6th on WER but 2nd on semantic similarity, so the best ASR depends on the downstream clinical task.
•POS metrics expose grammatical distortions WER misses; SDE is poor across all systems, so disfluencies are rarely preserved.
Comments
Email is required but is not shown publicly.
Loading comments...