Source-linked AI summary
FMA: A Dataset For Music Analysis
Michaël Defferrard, Kirell Benzi, Pierre Vandergheynst, Xavier Bresson
TL;DR
MIR research has been constrained by the limited availability of large, open audio datasets for reproducible evaluation and feature learning. The paper introduces FMA, a large Creative Commons-licensed collection with full-length audio, metadata, features, subsets, splits, and baseline evaluations. FMA provides a reference-scale benchmark while remaining smaller than commercial music services and has documented coverage limitations and genre bias.
Problem
MIR lacks large, complete, and easily available audio datasets, limiting reproducible benchmarking and data-intensive learning.
Method
The paper constructs and documents FMA, providing distributable audio, metadata, pre-computed features, genre-based subsets and splits, usage code, and genre-recognition baselines.
Results
FMA provides a large-scale collection that enables benchmarking beyond pre-computed features, while remaining two orders of magnitude behind commercial services.
Takeaways & Limitations
FMA supports reproducible MIR evaluation on audio and metadata in a collection closer to real-world scale than most openly available quality-audio datasets.
Takeaways & Limitations
FMA lacks metadata for some tasks and is biased toward experimental, electronic, and rock music, excluding mainstream music and few commercially successful artists.
Abstract
from arXiv · showhide
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
1. INTRODUCTION
The FMA addresses MIR’s shortage of large, open, and easily available audio benchmarks by providing a reference dataset with distributable audio, rich metadata, and reproducible access. Its scale, licensing, audio quality, and supporting resources are designed to support comparable evaluation and data-intensive learning.
- Motivation: MIR lacks a large, complete, and easily available dataset comparable to established benchmarks in other fields.This shortage has hindered research on data-heavy models such as deep learning.
- Large scale: FMA provides a large-scale collection intended to reduce over-training, represent musical ambiguity, average annotation noise, and limit confounding characteristics.The authors note that other datasets with available quality audio are two orders of magnitude smaller.
- Permissive licensing: FMA uses tracks whose licenses permit redistribution, and its data and code are released under CC BY 4.0 and MIT licenses.This licensing approach responds to copyright constraints that have historically limited publicly available MIR benchmarks.
- Available and quality audio: FMA distributes audio rather than only derived features or unstable external links, enabling feature learning and end-to-end learning.Unlike typical 10-to-30-second or low-quality clips, FMA includes full-length and high-quality audio.
- Metadata rich: The dataset includes rich track-, album-, artist-, user-, tag-, and free-form-text metadata, comparing favorably with datasets offering less metadata.Coverage is not complete, but the metadata includes fields such as favorites, play counts, comments, tags, descriptions, and biographies.
- Future proof and reproducible: FMA is designed for reproducible and durable benchmarking through downloadable archives, checksums, long-term hosting, shared code, and usage examples.The released code covers collection, analysis, subset and split generation, feature computation, and baseline testing.
2. DATASET
The FMA dataset is a dump of the Free Music Archive combining legally downloadable audio with extensive metadata and a hierarchical genre taxonomy. It provides multiple dataset subsets and a reproducible split for MIR research, while retaining real-world outliers and artist-provided labeling noise.
- 2.1 The Free Music Archive: The FMA dump combines track audio and metadata from a free, open archive directed by WFMU.The archive contains artist-released works under permissive licenses and combines user-generated content with curation.
- 2.2 Creation: The collection process gathered archive metadata, a 161-genre hierarchy, album and artist information, and MP3 audio, yielding 106,574 tracks after exclusions.Excluded items included unavailable downloads, failed trimming or feature extraction, and tracks whose licenses prohibited redistribution.
- 2.3 Content: The dataset preserves rich track-, album-, artist-, and user-level metadata, including genres, favorites, play counts, comments, tags, descriptions, and biographies.Metadata coverage varies by field, and the metadata was produced by artists while the musical content was curated.
- 2.4 Genres: FMA supports genre recognition through multiple artist-assigned sub-genres, a 161-genre hierarchy, and derived all-genre and root-genre columns.The hierarchy contains 16 root genres; labels may be inconsistent and are recommended as one of several complementary ground truths.
- 2.5 Features: The authors provide 518 pre-computed audio features, enabling experiments without feature extraction.Features are computed from windowed audio using librosa-derived descriptors and seven summary statistics, excluding zero-crossing rate from windowing treatment.
- 2.6 Subsets: Four subsets—Full, Large, Medium, and Small—support different resource and prediction settings, from the complete multi-label collection to 8,000 balanced 30-second clips.The proposed split is 80/10/10% for training, validation, and test, with stratification and an artist filter to avoid artist and album effects.
3. USAGE
The FMA supports diverse MIR tasks, especially music classification and genre recognition, while its full-length audio enables feature learning and direct audio analysis. Its metadata and complete tracks also support musicological studies and analyses of higher-level musical properties.
- 3.1 Music Classification and Annotation: The FMA supports genre recognition, artist identification, year prediction, and automatic tagging, with tags covering genre, mood, instrumentation, and style.
- 3.2 Genre Recognition: The proposed subsets and baseline evaluations provide straightforward settings for top-genre recognition, including single-label and multi-label tasks.
- 3.1 Music Classification and Annotation: Genre recognition is nontrivial because genre boundaries remain fuzzy despite genres being useful for organizing music collections.
- 3.2 Genre Recognition: The baselines provide a lower bound and indicate task difficulty, while the released code can be modified to test other features and classifiers.
- 3.2 Genre Recognition: Audio availability enables deep-learning architectures to operate on waveforms without feature engineering, although higher-level representations can perform better.
- 3.3 Data Analysis: Complete tracks support music-structure analysis, while metadata and audio can support musicological studies of relationships with higher-level representations.
4. DISCUSSION
The discussion identifies missing metadata and sampling bias as important boundaries on FMA-based research. It also notes that evaluating algorithms across multiple datasets and expanding annotations would strengthen conclusions.
- FMA lacks metadata for mood classification and instrument recognition, although existing tags may support further investigation.
- Public favorites, comments, and user mixes could support recommender-system evaluation, while additional activity data might become available after anonymization.
- The collection is biased toward experimental, electronic, and rock music, with little mainstream music and few commercially successful artists.
- Whether mainstream and non-mainstream datasets share similar properties, or whether algorithms transfer between them, remains unknown.
- The authors recommend evaluating algorithms on multiple datasets, especially because dataset composition may matter for recommendation and other tasks.
5. CONCLUSION AND PERSPECTIVES
The conclusion presents FMA as a large-scale, open benchmark that supports reproducible comparison and audio-based learning across MIR tasks. Future development should validate labels and add metadata and labels.
- FMA enables large-scale benchmarking with shared audio, defined tasks, and splits, making results from different research groups more comparable.
- By providing audio rather than only pre-computed features, FMA supports new feature sets, learned features, and mappings directly from audio.
- The 161-genre hierarchy makes FMA especially useful for music classification and offers possibilities rarely found in alternative collections.
- Future work should measure agreement among independent annotators and obtain additional metadata and labels.
- The dataset is grounded in Creative Commons sharing, with artists making music available for public, educational, and research use.