Alzheimer's disease detection from speech typically relies on both paralinguistic cues, such as emotion, and linguistic content. However, training models for this task is often hampered by limited data, and fine-tuning a large pretrained speech encoder can easily lead to overfitting.
To address this, the arXiv preprint introduces a method called Dual Task Vector Merging. It leverages knowledge from two auxiliary tasks—emotion recognition and automatic speech recognition (ASR)—by merging their task vectors into the model. This approach aims to make use of both types of cues without requiring large amounts of Alzheimer's-specific data.
The paper argues that this merging strategy can help low-resource Alzheimer's detection by transferring relevant knowledge from emotion and ASR tasks. The abstract does not provide experimental results, so the effectiveness is not yet quantified in this source.