2020
Bird song classification
Sound-based bird ID with convolutional networks and mel‑cepstrum spectrograms — a real‑world WiMLDS collaboration.
A volunteer group—data scientists, PhD students, ornithologists, analysts, and engineers—ran two‑week sprints on Python, audio features, modelling, and presentation. The resulting pipeline reached about 87% accuracy on the held‑out test set for predicting the correct species from a recording.
What we built
End‑to‑end scope: data discovery, preparation, model iterations, error analysis, and a public write‑up. The team combined domain knowledge from ornithology with deep learning on spectrograms so the system stays interpretable enough to discuss with stakeholders—not only a leaderboard score.
Resources hub
Along the way we collected datasets, papers, competitions, open‑source tools, and articles on bird audio in ML— maintained as an open list for the community.
Bird recognition review (GitHub) →