[Paper Review] FMA: A Dataset For Music Analysis

Michaël Defferrard, Kirell Benzi+2 more|arXiv (Cornell University)|Dec 6, 2016

Music and Audio Processing38 references161 citations

TL;DR

Introduces the Free Music Archive (FMA), a large CC-licensed audio dataset with full-length high-quality tracks, rich metadata, and pre-computed features to enable reproducible MIR benchmarking and end-to-end learning. It provides subsets, splits, and baselines for genre recognition and related tasks.

ABSTRACT

We introduce the Free Music Archive (FMA), an open and easily accessible dataset suitable for evaluating several tasks in MIR, a field concerned with browsing, searching, and organizing large music collections. The community's growing interest in feature and end-to-end learning is however restrained by the limited availability of large audio datasets. The FMA aims to overcome this hurdle by providing 917 GiB and 343 days of Creative Commons-licensed audio from 106,574 tracks from 16,341 artists and 14,854 albums, arranged in a hierarchical taxonomy of 161 genres. It provides full-length and high-quality audio, pre-computed features, together with track- and user-level metadata, tags, and free-form text such as biographies. We here describe the dataset and how it was created, propose a train/validation/test split and three subsets, discuss some suitable MIR tasks, and evaluate some baselines for genre recognition. Code, data, and usage examples are available at https://github.com/mdeff/fma

Motivation & Objective

• Provide a large, openly licensed music dataset to enable reproducible MIR benchmarking and support end-to-end learning research.
• Offer full-length high-quality audio with rich per-track, per-album, and per-artist metadata in a hierarchical 161-genre taxonomy.
• Propose standardized train/validation/test splits and data subsets to suit different resource levels.
• Demonstrate baseline genre recognition performance to establish lower-bound benchmarks.

Proposed method

• Assemble 106,574 tracks from 16,341 artists across 14,854 albums licensed under Creative Commons.
• Construct a 161-genre taxonomy with a multi-level hierarchy and compute and store genres_all and genres_top for each track.
• Provide 518 pre-computed audio features using librosa 0.5.0 on 2048-sample windows with 512-sample hops, summarized by seven statistics.
• Create four dataset subsets (Full, Large, Medium, Small) with tailored properties to balance realism and accessibility.
• Define an 80/10/10 stratified split with artist-level separation to avoid train-test leakage and ensure representation of root genres.
• Offer baselines for genre recognition using classifiers (LR, kNN, SVM, MLP) across multiple feature sets.

Experimental results

Research questions

• RQ1 How can a large-scale CC-licensed music dataset support reproducible MIR research and end-to-end learning?
• RQ2 What metadata richness and genre hierarchy features enhance music genre recognition and related tasks?
• RQ3 What are suitable train/validation/test splits and subsets to balance realism, accessibility, and computational constraints?
• RQ4 What baseline performance do common classifiers achieve on top-genre recognition using pre-computed features on the FMA?

Key findings

• The dataset provides 106,574 tracks with full audio (44.1 kHz, ~320 kbps) and rich metadata across per-track, per-album, and per-artist levels.
• A 161-genre hierarchy with multiple sub-genres and a genres_all annotation is implemented to support fine-grained MGR tasks.
• Pre-computed 518 Librosa features per track enable rapid baseline experiments, with seven statistical summaries over windows.
• Baseline genre recognition on the Medium subset shows various accuracies across feature/classifier combinations, illustrating task difficulty and providing lower-bound references.
• The authors release code, data, and reproducible workflows to foster transparency and reproducibility in MIR benchmarking.