Dementia is a significant threat to our society, both in terms of personal life and public health. Early diagnosis of dementia is critical to provide timely treatment. As an easily accessible and non-invasive technique, we study voice analysis techniques to identify Mild Cognitive Impairment. We build a deep neural network model using acoustic features to represent a voice signal. A model using x-vectors per 5-second audio segments outperformed other feature models. The model prediction sensitivity is 0.58 and the specificity is 0.83.
