GAVDNet Automated Animal Call Detector - Trained Models & Validation Test Results

This dataset contains the validation testing results for the GAVDNet automated animal call detector framework. GAVDNet was tested on two baleen whale calls, the Antarctic blue whale Z-call, and the Chagos pygmy blue whale. For each target call, we trained and tested four different models, each with training samples synthesised at a different range of random signal-to-noise ratios. Additionally, in order to evaluate precision-recall trade-off, we also tested each of these eight (8) models with a range of post-processing activation threshold values. For each of the 2 target animal calls, this yielded 4 models, and 8 activation thresholds, for a total of 64 experimental conditions. Due to unknown reliability of the annotation logs used as ground truth, the best performing model was then subjected to human adjudication, which resolved disagreements between the detector and the ground truth. The results for all tests these tests are included in this repository. Additionally, the trained model weights are also provided. This repository does not contain the original audio data on which the detectors were tested. The original audio datasets are not the property of the authors and cannot be redistributed.