MERIDIAN team members Bruno Padovese and Fabio Frazao developed an open-source North Atlantic Right Whale (NARW) call detector tool, enabling anyone with an acoustic data set to scan their data for NARW vocalizations and train their own models. The tool provides a high-level interface for developing deep learning-based detectors through a command-line interface (CLI), without requiring any programming skill and only a high level understanding of machine learning concepts.

The tool is written in Python and based on the Ketos deep learning package for underwater acoustics. It is organized into sub-modules, each comprehending a major task when developing a deep learning based detector. These tasks include:

  1. Create training and test datasets from raw acoustic data using different audio representations (i.e., spectrograms, cepstrogram, etc);
  2. Train a deep learning model to detect and classify NARW vocalizations;
  3. Evaluate the model on fully-annotated test datasets and report standard performance metrics;
  4. Run the model by simply pointing to a directory of audio files;
  5. Adapt an existing model to a new acoustic environment;

The tool comes with a pre-trained model using thousands of NARW upcalls from acoustic data offshore Eastern Canada and the US, that can be directly applied to a new acoustic dataset or adapted to a new set of NARW calls. The code repository and detailed documentation can be found online.