Abstract
We describe our method for automatic bird species classification, which uses raw audio without segmentation and without using any auxiliary metadata. It successfully classifies among 501 bird categories, and was by far the highest scoring audio-only bird recognition algorithm submitted to BirdCLEF 2014. Our method uses unsupervised feature learning, a technique which learns regularities in spectro-temporal content without reference to the training labels, which helps a classifier to generalise to further content of the same type. Our strongest submission uses two layers of feature learning to capture regularities at two different time scales.