Automated Dementia Detection from Speech Using Disfluency Features

Abstract

Dementia is a major and growing public health concern that affects millions of individuals worldwide. Early detection has moved from traditional diagnostic methods to the use of predictive modeling approaches, Machine Learning (ML) and subsequently Deep Learning (DL) and more recently transformer-based models. The use of speech has rapidly emerged as an effective modality in multimodal dementia detection. However, most existing methods fail to effectively capture both how something is said (from the audio) and what is said (from the transcribed text). Yet, both components are critical for accurate and reliable diagnosis. Therefore, in this work, we propose Pause-based Transformer with Disfluency Markers, a novel approach that explicitly encodes speech disfluencies as special tokens, allowing the Transformer to attend to these discriminative features. The evaluation shows an improved of accuracy outperforming traditional baselines.