This tool annotates sound files using neural networks. It uses a 1D architecture based on U-Net with additional post-processing heuristics including a Hidden Markov Model.
DISCO is ideal for long streams of sound that need to be classified over time, producing output fully compatible with The Cornell Lab of Ornithology's sound tool RAVEN. Work is currently underway to annotate short samples of data with a single label. DISCO began jointly with the University of Montana's Emlen Lab as an annotator for Japanese and Taiwanese Rhinoceros Beetle courtship songs, but it now generalizes to any kind of recording.
Install requires python version >=3.8. Install directly from git with pip:
pip install git+https://github.com/TravisWheelerLab/disco.git
DISCO contains subcommands useful for training and evaluating models on sound data. Deep learning projects typically follow a series of steps, and DISCO tries to emulate each of these steps:
label, extract, shuffle, train, infer.
NOTE
Models were updated in the most recent version. Remove them with rm ~/.cache/disco/* before running any disco commands.
Learn more about how to use the tools provided in this package in the wiki.