LAUSR

This paper considers computer-assisted learning of sound spectra in environmental recordings to facilitate manual bird species identification. Today, a variety of automated methods have been successfully applied for acoustic recognition of specific bird species. These methods are more effective for single targeted species detection. For in-field recordings, however, simultaneous vocalisations and unknown species usually make such methods less effective. In this study, we propose a non-negative matrix factorisation based method to facilitate manual bird species identification from environmental recordings. First, distinct sound spectra are extracted from each audio clip by applying non-negative matrix factorisation and clustering techniques. Based on these distinct sound spectra, a greedy algorithm is then designed to sample audio clips. Each sampled audio clip maximises the number of new spectra. People who follow this sampled sequence of audio clips should be able to identify the most species given a fixed number of audio clips. The efficiency is validated with annotated bird species per minute provided by experienced ornithologists.