Dual Task Vector Merging Improves Alzheimer's Detection from Speech

Dual Task Vector Merging adapts a frozen wav2vec 2.0 backbone for Alzheimer's detection by injecting weight deltas from emotion and ASR tasks. This parameter-efficient method updates less than 0.1% of parameters and outperforms LoRA and full fine-tuning on ADReSS datasets.