ISCA Archive Interspeech 2024
ISCA Archive Interspeech 2024

Cascaded Transfer Learning Strategy for Cross-Domain Alzheimer's Disease Recognition through Spontaneous Speech

Guanlin Chen, Yun Jin

In our work, we propose a cascaded transfer learning strategy for cross-domain Alzheimer’s disease (AD) recognition through spontaneous speech. This strategy cascaded a pre-trained GPT-3 model as the first-level and a Random Forest Multi-Source Discriminant Subspace Alignment (RF-MDSA) Algorithm as the second-level. The goal of this strategy is to align feature spaces extracted from different corpora to derive a common projection subspace. We conduct experiments on three corpora in Chinese, English, and Spanish. On single corpus experiments, we achieve accuracy rates of 74.6%, 82.4%, and 62.8%, respectively. In cross-domain experiments, we achieve accuracy rates of 71.8%, 72.9%, and 56.1%, respectively. These results suggest that our cascaded strategy improves the model’s generalization capabilities, making it effective for cross-domain AD recognition.