Alzheimer's Disease Classification From Speech Pause Distributions With Context Information

Alzheimer's disease (AD) is known to affect the lengths and frequencies of certain kinds of pauses in speech. Previous studies have used features based on pause lengths for AD classification. We conjecture that in addition to using pause lengths, it is beneficial to incorporate the "context" behind each pause, i.e., what is being said before and after each pause. We propose an AD detection method based on this idea. As part of the proposed method, pause lengths and context are extracted from the raw audio using automatic speech recognition (ASR) and forced alignment. Then, statistical summaries of pause lengths with context information are extracted from the transcripts and used as features for classification. Our results indicate that incorporating the context significantly improves classification performance compared to using pause lengths alone, with classification accuracy of up to 81%. Additionally, the proposed features largely preserve privacy.