Audio Localisation

This research presents a complete computational pipeline for acoustic source localisation using unsynchronised Passive Acoustic Monitoring (PAM) arrays. By deploying arrays of microphones with overlapping recording ranges (e.g., <150m spacing), we can triangulate individual vocalisation events, allowing researchers to track spatial dynamics and estimate abundance metrics, such as the Minimum Number of Individuals (MNI).

Methodology

Localising signals from independent, unsynchronised recorders in noisy environments presents significant challenges. We address these through a multi-stage automated pipeline:

1. Mask-Level Segmentation: Traditional bounding-box classification is insufficiently precise for Time Difference of Arrival (TDOA) calculations. We use synthetic data generation to train a single-class image segmentation model, which isolates vocalisations to precise time-frequency masks.
2. Species Identification: The segmented masks provide highly specific targets for a fine-tuned BirdNET classifier, significantly improving downstream classification accuracy (e.g., isolating korimako / bellbird vocalisations).
3. Signal Association & Cross-Correlation: Mask-isolated audio signals are cross-correlated across nearby microphone streams to determine precise pairwise time lags. Operating at the mask level prevents background noise and environmental echoes from corrupting the TDOA estimations.
4. Gaussian Maximum Likelihood Estimation: Source locations are estimated by performing a spatial grid search over the computed TDOAs for consistent microphone triplets. These individual estimates are then aggregated using Gaussian Maximum Likelihood Estimation (MLE) to produce a robust final spatial coordinate.

By mapping discrete vocalisation events in space, this approach bridges the gap between basic presence/absence PAM data and high-resolution spatial ecology, enabling scalable monitoring of ecosystem dynamics without the need for expensive, hardware-synchronised recording equipment.