Discogs-EffNet
EfficientNet music embeddings and 400 style tags learned from Discogs editorial metadata.
About
Discogs-EffNet is a family of music representation models from Barcelona's Music Technology Group. Trained on an in-house collection annotated with Discogs metadata, the EfficientNet-B0 backbone produces 1280-dimensional embeddings; the classification variant predicts 400 Discogs style labels, while contrastive variants learn similarity by pulling together tracks from the same artist, label, release, or the same track itself. The work behind it was published at ISMIR 2022.
The models are distributed through the Essentia model zoo as TensorFlow graphs (plus ONNX exports) with matching classification heads for genre, mood, danceability, voice/instrumental, and a 40-class MTG-Jamendo instrument classifier. The embeddings are 18 MB and run comfortably on a CPU, which is why several open-source music-library projects use Discogs-EffNet as their default embedder. The catch is the license: the weights are CC BY-NC-SA 4.0, so commercial use is off the table.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported: the ISMIR 2022 paper evaluates representation quality on tagging and similarity-style downstream tasks; no independent head-to-head benchmark table published.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified, metrics-v1). Evidence: ISMIR 2022 paper ("Music Representation Learning Based on Editorial Metadata from Discogs"); models documented in the Essentia model zoo with downloadable weights and metadata JSONs; used as the default embedder by independent open-source projects. Strengths: strong music-specific embeddings, tiny footprint, real downstream adoption. Docked for: (1) CC BY-NC-SA 4.0 weights exclude commercial use; (2) no independent published benchmark. License: CC BY-NC-SA 4.0 (weights); Essentia library itself is AGPL-3.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 5.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · ISMIR 2022 paper 'Music Representation Learning Based on Editorial Metadata from Discogs'; MTG-UPF model zoo at essentia.upf.edu
- · 1280-dim embeddings + 400 Discogs style labels; ONNX exports ~18MB run on CPU
- · Used as default embedder by independent open-source music-library projects
- · NC-SA weights exclude commercial use; no independent published benchmark