MERT v1 · 95M / 330M
Extract music representations for downstream tasks such as tagging, key or emotion classification.
UnderstandingOpen-source code95M / 330M
- Start here
- Choose a concrete downstream task and compare a small supervised head against a simple feature baseline. Embeddings alone do not produce reliable labels.
- Training data
- The card reports 20,000 hours for v1-95M and 160,000 for v1-330M. The paper describes private music pretraining.
- Training rights
- The v1 corpus is not released as a cleared public audio collection. The separate v0-public experiment is not the v1 training set.
- License & output terms
- v1 weights: CC BY-NC 4.0.
- Limitations
- Non-commercial weights. Features need task-specific evaluation and usually a trained downstream head.
- Technique
- Self-supervised masked audio modeling
- Runtime
- Local feature extraction; preserve the model's 24 kHz preprocessing.
- Parameters
- 95M / 330M
Sources reviewed . Training and rights notes summarize provider disclosures. A model license does not clear every source recording or generated output. Parameter counts refer to the named variant or component, not total runtime memory.