Matej Martinc(Jozef Stefan Institute) and Senja Pollak(Jožef Stefan Institute)
The paper describes a multimodal approach to the automated recognition of Alzheimer’s dementia in order to solve the ADReSS (Alzheimer’s Dementia Recognition through Spontaneous Speech) challenge at INTERSPEECH 2020. The proposed method exploits available audio and textual data from the benchmark speech dataset to address challenge's two subtasks, a classification task that deals with classifying speech as dementia or healthy control speech and the regression task of determining the mini-mental state examination scores (MMSE) for each speech segment. Our approach is based on evaluating the predictive power of different types of features and on an exhaustive grid search across several feature combinations and different classification algorithms. Results suggest that even though TF-IDF based textual features generally lead to better classification and regression results, specific types of audio and readability features can boost the overall performance of the classification and regression models.