Soundscape Viewer: An integrated toolbox of audio source separation for soundscape-based ecosystem assessment

Abstract

Soundscapes contain collective information on habitat quality and ecosystem dynamics. Current soundscape assessment methods, including clustering and ecoacoustic indices, may deliver unsatisfied performance when multiple sound sources are recorded simultaneously. We introduce an open-source Soundscape Viewer toolbox for separating soundscapes into sounds of biological, environmental, and anthropogenic sources. Based on non-negative matrix factorization (NMF), Soundscape Viewer enables audio source separation in both supervised and unsupervised manners. Conventional supervised NMF learns spectral features from labeled animal vocalizations, yet the performance is sensitive to noise levels. By assuming source-specific periodicity in unsupervised NMF, Soundscape Viewer can learn discriminative features between animal vocalizations and background noise. To cope with dynamic acoustic environments, Soundscape Viewer also integrates adaptive and semi-supervised learning approaches. Adaptive learning improves model generalization when sound characteristics differ from the training data, and semi-supervised learning allows the model to recognize new sound sources. We applied Soundscape Viewer in the evaluation of soundscape dynamics and the automatic detection of animal vocalizations. Our results reveal that the mutual interference between sound sources can be effectively reduced in various ecosystems. Source separation will facilitate biodiversity assessment and increase our understanding of biotic and abiotic sounds' interactions.