Bird Song Classification in Field Recordings: Winning Solution for NIPS4B 2013 Competition *

The challenge of the NIPS4B competition is to identify 87 sound classes of birds and other animals present in 1000 audio recordings, collected in the field. The difficulty of this task lies in the large number of species and sounds that have to be identified in various contexts dealing with different levels of background noise and simultaneously vocalizing animals. The solution presented here ranks first place on the kaggle private leaderboard and achieves an Area Under the Curve of 91.7% (AUC). 1 In trod u cti on The audio data was recorded at different places in Provence France and is provided by the BIOTOPE society, having one of the largest collections of wildlife recordings of birds in Europe. The nearly 2 hours of recordings are split into smaller clips ranging from 0.25 to 5.75 seconds. The recordings were done with Wildlife Acoustics SM2 and are presented in uncompressed WAV format with a sample rate of 44.1 kHz. The 87 individual sound classes within these recordings represent different bird species and their songs, calls and drumming. Other animal species living in the same environment like insects and one amphibian are also included. The training set consists of 687 audio files. Each file is paired with the subset of sound classes present in that recording. Some recordings are empty, containing only background noise, others contain up to 6 different simultaneously vocalizing birds or insects. Each species is represented by nearly 10 training files within various contexts, different background noises and an arbitrary number of other species. The goal of the competition is to identify which of the 87 sound classes of birds and amphibians are present in 1000 continuous wildlife recordings, using only the provided audio files and machine learning algorithms for automatic pattern recognition.