Cross-Dataset Evaluation of Foundation Model-Based Acoustic–Text Fusion for Dementia-Related Speech Analysis

Abstract

This study examines the cross-dataset generalizability of a foundation-model-based acoustic–text framework for dementia-related speech analysis across three datasets: ADReSSo2021, Ivanova, and Delaware. The proposed pipeline combines WavLM-Large acoustic embeddings with Qwen3-ASR-1.7B transcription and Qwen3-Embedding-0.6B text representations, integrated through validation-guided weighted late fusion. Results demonstrate a dataset- and task-dependent contribution of multimodal fusion: fusion yielded the highest F1-scores on ADReSSo2021 (90.18%) and Delaware (62.12%), whereas acoustic representations showed greater prominence across the Ivanova tasks. Label-composition analyses further revealed clearer HC–AD than HC–MCI separability, highlighting the potential role of MCI heterogeneity in classification difficulty. Chunk-length ablation across 5, 10, 15, and 30 s revealed task-specific preferences for temporal context, with the best performance occurring at different segment durations across tasks. These findings emphasize the importance of speech protocol, diagnostic composition, dataset characteristics, and temporal context when evaluating acoustic–text integration for dementia-related speech analysis.