The population of people who suffer from dementia is increasing year by year. Early detection of dementia is very important for the patient. However, traditional testing tools rely on experienced doctor to conduct and take long time. Therefore, this thesis presents an approach to detect dementia automatically through acoustic features. The proposed method employs LSTM recurrent neural network on MFCC features extracted from spoken utterances to build a predictive model. We train the model on utterances which come from two speech corpora (Mandarin_Lu and Aishell) and deal with imbalanced data. Our model achieves an accuracy of 98% on test set.