Unsupervised dictionary extraction of bird vocalisations and new tools on assessing and visualising bird activity

Abstract

A broad range of organisations and individuals are collecting wildlife audio recordings. Huge amounts of audio data have been gathered in the past and since the popularisation of automatic recording units the data are piling up exponentially. The point in gathering them is to analyse them, evaluate insights and hypotheses, identify patterns of activity that are otherwise not apparent and finally design policies on biodiversity issues. For massive volumes of data even visual inspection of spectrograms is unfeasible and interesting cases that could provide valuable insight for concrete hypotheses on the biodiversity status can slip into bliss. In this paper we research a range of techniques that work with minor human supervision. These techniques will construct a dictionary of templates extracted in an unsupervised way from reference recordings and then crawl over a large number of recordings to examine the underlying bioacoustic activity. This work is general and we have applied it to many datasets of animal's vocalisations (e.g. cetaceans, mice, birds). To test our tools objectively and for the sake of reproducibility in this work we report on the MLSP 2013 bird dataset that recently has been publicly released along with all its annotations. We are not interested as to which is the best scoring approach for this dataset. Our aim is to describe novel machine learning tools that try to refine our understanding of biodiversity by answering questions such as: Is the recording under examination void of bird vocalisations or not? If there is bird activity, how many different species are in the recording? What are the most important characteristic spectral segments for recognizing a specific species? The database however is valuable to us to quantify our findings.