A template-based automatic bird phrase classification in noisy environments using limited training data

Abstract

Bird Songs typically comprise a sequence of smaller units, termed phrases, separated from one another by longer pauses; songs are thought to assist in mate attraction and territory defense. Studies of bird song would often be helped by automated phrase classification. Past classification studies usually employed techniques from speech recognition, such as MFCC feature extraction and HMMs. Problems with these methods include degradation from background noise, and often require a large amount of training data. We present a novel approach to robust bird phrase classification using template-based techniques. One (or more) template is assigned to each phrase with its specific information, such as prominent time-frequency components. In our trials with 1022 phrases from Cassin’s Vireo (Vireo cassinii) that had been hand-identified into 32 distinct classes, far fewer few examples per class were required for training in some cases only 1 to 4 examples for 84.95%-90.27% accuracy. The choice of distance metrics was crucial for such systems. We found that weighted 2D convolution is a robust distance metric for our task.. We also studied phrase patterns using Multi-Dimensional Scaling, a discriminative feature for phrase patterns that are very similar