Spatial Bioacoustics: Soundscape Analysis with a Co-located Microphone Array

The integration of passive acoustic sensors with machine learning enables large-scale, low-cost, non-invasive monitoring of vocal animals. A remaining challenge to the successful deployment of such monitoring systems is in maintaining high classification accuracy across complex real-world soundscapes composed of overlapping calls and variable noise patterns. In this talk, we present an algorithmic pipeline for using a terrestrial co-located microphone array to (1) estimate acoustic direction-of-arrival (DoA), (2) distinguish individual sound sources in the environment, and (3) approximate their spectrograms and time-domain signals. First, we evaluate the performance of multiple DoA estimation algorithms — including an active intensity method, white noise gain constraint beamforming, and multiple signal classification — as well as multiple approaches for source separation — including angular thresholding, a Gaussian mixture model, and non-negative matrix factorization. Next, we demonstrate the analysis pipeline for recordings collected at wildlife refuges during the dawn chorus in late spring, when birds are most vocally active. We significantly improve species-level performance metrics by applying source separation to the recordings prior to classification with the BirdNET network. This approach opens possibilities for additional spatiotemporal analysis of soundscapes, including the ability to visualize movement and perform directional filtering.

Speaker(s): Irina Tolkova,

Bldg: Coastal Institute (bldg. 26), 215 South Ferry Rd., Narragansett, Rhode Island, United States, 02882, Virtual: https://events.vtools.ieee.org/m/314823