fma

Why we included this project

FMA gives music information retrieval researchers a large, ready-made audio dataset without the usual scramble to assemble and label one yourself. It holds 917 GiB of Creative Commons-licensed audio from over 106,000 tracks, arranged in a 161-genre hierarchy, with metadata tables, pre-computed features, and code that loads the data and reproduces baseline genre recognition models. You can start with a small eight-genre subset and work up to the full untrimmed archive, so the data volume fits whatever compute you have. The included notebooks walk through exploration, feature extraction, and baseline evaluation, which makes the project useful for teaching as well as for benchmarking. For reproducible experiments, the documented train/validation/test split and working example code take care of a lot of the setup friction.