Home / Projects / The Genre Detective
DATA SCIENCE & AI · INTERNATIONAL

Pick a song. See what the model hears.

Real data — 114,000 tracks

Problem Statement

Research question: Given a track's audio characteristics, can a model accurately predict its genre?

Music platforms host thousands of genres, and manually tagging every track doesn't scale. This project predicts a track's most likely genre directly from its audio profile, trained and evaluated on 89,741 real, deduplicated tracks across 113 genres.

42.9%
Top-1 Accuracy
63.1%
Top-3 Accuracy
71.8%
Top-5 Accuracy

Try it live

Runs on Hugging Face Spaces
Waking up the model — this can take up to a minute on the first visit today. It'll be instant after that.

A real bug caught and fixed

24,259 tracks appeared more than once under different genre labels. Left unhandled, this would let the same track leak into training and test data under different labels. Fixed by deduplicating before splitting.

Why 42.9%, not 90%

114-way genre classification is genuinely hard — the taxonomy includes near-duplicate labels (metal vs. black-metal vs. death-metal). Top-3/Top-5 accuracy is reported for that reason.