Feature design for multilabel bird song classification in noise (NIPS4B challenge)

Our submission to the challenge therefore focusses on feature design for two goals: noise robustness, and the representation of temporal structure. We first analyse each sound file into basic features, either MFCCs (13 MFCCs plus delta features) or our peak-chirplet representation [3]. Importantly, both of these feature algorithms are modified to apply noise reduction in their spectral analysis step, simply by median-filtering: taking a spectral median profile across time, subtracting this median from the values, and keeping only the positive values.