fma

FMA is a comprehensive dataset designed for music information retrieval (MIR) tasks, featuring 917 GiB of Creative Commons-licensed audio encompassing 106,574 tracks across 161 genres. It includes high-quality audio, pre-computed features, and extensive metadata, facilitating tasks such as genre recognition and user modeling. Key features include a hierarchical genre taxonomy, train/validation/test splits, and tools for analyzing and leveraging the dataset in machine learning applications.