Integrating acoustic micro-disfluencies and emotional context for robust Alzheimer’s dementia detection using transformer-based models

Alzheimer’s dementia poses significant challenges for early detection due to its subtle impact on speech and emotional patterns. Existing methods often lack the ability to capture both micro-level speech disfluencies and macro-level emotional-contextual dynamics. To address this, we propose a model integrating Temporal Acoustic Micro-Disfluency Patterns (TAMP) and Emotion-Contextual Acoustic Memory Fusion (ECAMF) features with GRU, multi-head attention, and a Transformer encoder. Evaluated on the ADReSSo dataset, the model achieved 90.2% accuracy, 89.6% precision, 88.7% recall, and an F1-score of 89.1%, significantly outperforming traditional and pre-trained methods. This approach offers an efficient and robust solution for early Alzheimer’s dementia detection, emphasizing both acoustic and emotional features in a unified framework, while demonstrating improved sensitivity to nuanced speech and emotional patterns.