Dementia Detection by Analyzing Spontaneous Mandarin Speech

Abstract

Ahstract-The Chinese population has been aging rapidly resulting in the largest population of people with dementia. Unfortunately, current screening and diagnosis of dementia rely on the evidences from cognitive tests, which are usually expensive and time consuming. Therefore, this paper studies the methods of detecting dementia by analyzing the spontaneous speech produced by Mandarin speakers in a picture description task. First, a Mandarin speech dataset contains speech from both healthy controls and patients with mild cognitive impairment (MCI) or dementia is built. Then, three categories of features, including duration features, acoustic features and linguistic features, are extracted from speech recordings and are compared by building logistic regression classifiers for dementia detection. The best performance of identifying dementia from healthy controls is obtained by fusing all features and the accuracy is 81.9% in a 10-fold cross-validation. The importance of different features is further analyzed by experiments, which indicate that the difference of perplexities derived from language models is the most effective one.