Multilingual Noise-Aware Alzheimer's Disease Detection from Speech Using Hyperbolic Multi-View Fusion
Abstract
Alzheimer's disease (AD) is a progressive neurodegenerative disorder that affects speech and language abilities, making speech analysis a promising non-invasive tool for early screening and monitoring. Most existing speech-based AD detection systems are developed using single-language datasets and controlled recording conditions, which limits their applicability in multilingual and noisy environments. This study presents a multilingual noise-aware speech-based AD detection framework evaluated across English, Chinese, and Spanish. The proposed framework uses pretrained speech encoders as frozen feature extractors, applies controlled noise augmentation at clean, 10 dB, and 15 dB conditions, and introduces HyJS-Fuse, a hyperbolic Jensen–Shannon divergence-based multi-view fusion framework. HyJS-Fuse maps complementary speech representations into a Poincaré ball, fuses them through Möbius addition, and optimizes the fused prediction using a supervised Jensen–Shannon divergence objective with optional branch-consistency regularization. Experimental results show that the proposed framework consistently outperforms single-encoder and concatenation-based baselines across languages and acoustic conditions. The best clean-condition results reach 98.27% accuracy and 96.89% F1 on English, 95.06% accuracy and 93.82% F1 on Chinese, and 91.02% accuracy and 89.24% F1 on Spanish. These findings demonstrate that geometry-aware multi-view fusion is effective for robust multilingual AD detection from speech.