This model was trained to as part of collaboration between Mote Marine Laboratory & Aquarium, Southeast Coastal Ocean Observing Regional Association, and Axiom Data Science to develop a model capable of detecting and classifying fish vocalizations from audio files collected from hydrophones.

More information available at the project archive repo.

This model was trained on spectrograms

A reproducible Jupyter notebook describing the training of the model is available in the archive repo.

The model was intended to be a proof on concept to aid researchers identify fish vocalizations through vast amounts of audio data collected from hydrophones. Although the training data was collected using multiple devices in multiple locations, the model may not be generally applicable to other uses.

A training set of spectrograms of fish calls was created based on annotations of fish sounds in passive acoustic recordings by a hydrophone were provided by Jim Locascio, Max Fullmer, and volunteers from the Mote Marine Laboratory & Aquarium.

Due to severe imbalances in the number of samples per class, the training involved both under-sampling classes with many samples and over-sampling classes with few classes so that the model was trained on 50 samples per class. This number was derived in a completely ad-hoc fashion based on the distribution of class samples.

Some classes did not meet the training criteria, high signal-to-noise ratio and minimum call overlap, and were therefore excluded from the model training. As such, the number of classes represented in the trained model is few than the amount of labeled classes in the training set.