Skip to content
← Models
Open weightsMusic Technology Group, UPF (Pons, Serra)/10 trustVerified

MusiCNN

Musically motivated CNNs for music audio tagging, pre-trained on MagnaTagATune and the Million Song Dataset.

Visit official page →

About

MusiCNN (pronounced "musician") is the friendly classic of music tagging: convolutional networks whose filters are shaped by musical intuition (timbral and temporal filter shapes inspired by how music actually sounds), pre-trained on MagnaTagATune (MTT_musicnn) and the Million Song Dataset (MSD_musicnn, plus a bigger variant). Released as an ISMIR 2019 late-breaking demo with a pip-installable library, it gives you top-N tags, taggrams, and embeddings out of the box.

The training follow-up work reported 90.77 ROC-AUC / 38.61 PR-AUC on MagnaTagATune and 88.81 / 31.51 on MSD for the attention-based variant, provider-reported figures that were strong for their era. The repo (jordipons/musicnn) is ISC-licensed. The honest caveat is age: the training code is TensorFlow 1-era and the project is in maintenance mode, so treat it as a dependable, well-understood baseline and feature extractor rather than the current state of the art.

Specs

Context window

Audio input (3-second analysis windows in the tagger)

Pricing

Free — ISC licensed

Free quota

N/A (local inference)

API access

No hosted API — pip install musicnn; also ships as the mel frontend inside Essentia

Weights

Open weights

Released

Not disclosed

Benchmarks

Provider-reported: 90.77 ROC-AUC / 38.61 PR-AUC (MagnaTagATune) and 88.81 / 31.51 (MSD) for the attention variant; era-level, provider-claimed.

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

music-taggingtaggram-extractionmusic-embeddingsmir-baselinestransfer-learning

Trust score — 6/10

/10 trustVerified

Scored 6/10 (verified, metrics-v1). Evidence: ISMIR 2019 late-breaking paper (arXiv 1909.06654) from MTG-UPF; repo jordipons/musicnn (724 stars, ISC license) with pip package; mel frontend reused by Essentia. Strengths: genuinely usable out-of-the-box tagger, permissive license, wide name recognition. Docked for: (1) legacy TF1-era codebase in maintenance mode; (2) benchmarks are provider-reported and from 2019. License: ISC.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
6.0/10
Maintenance · recency, releases, issue load
3.0/10
Openness · license + docs
8.0/10
Manual review · sanity read + red-flag check
6.0/10
  • · ISMIR 2019 late-breaking demo (arXiv 1909.06654), Pons & Serra, MTG-UPF: musically-motivated CNNs pre-trained on MagnaTagATune and MSD
  • · Repo jordipons/musicnn (724 stars) with pip package; top-N tags, taggrams, embeddings; mel frontend reused by Essentia
  • · TF1-era codebase in maintenance mode; attention-variant benchmarks (90.77 ROC-AUC / 38.61 PR-AUC MTT) are provider-reported and from 2019

Sources