GPU-based training of autoencoders for bird sound data processing

Bird sounds have been studied in recent years due to their significance in helping ornithologists, and ecologists to monitor birds activities, which reflect climate changes, biodiversity, and reserves local protection status. Within the increasingly collected large amount of bird sound data from experts and amateurs, how to handle, and employ the state-of-the-art deep learning methods to mining such large amount of data, is bringing a huge challenge, and opportunity for the research community. In this work, we propose a framework using the GPU to accelerate autoencoders training for a large amount of bird sound data. Experimental results show that the GPU can considerably speed up the training process of bird sounds when fed within different scales of data, or feature numbers, compared with CPU-based learning.