As the result of aging society, we face an increasing number of people being affected by Alzheimer's disease (AD). Early prediction of AD has a major importance to not only to prevent the disease become worse but also further to make the patients to be fully recovered. We propose a language-independent approach of detecting AD patients by leveraging paralinguistic features from the audio data. We achieved the best result of 71.35% by aggregating the utterance-level prediction from gated convolutional neural network (GCNN).