drewbie/jazz-pianist-style-classifier

Jazz Pianist Style — Pianist Classifier

Pianist identification model from the ISMIR 2026 paper Learning Jazz Pianist Style with Cross-Attention Conditioning (Edwards, Maezawa, Dixon).

Aria-medium contrastive backbone fine-tuned to classify 12 PiJAMA pianists from 1024-token windows. On the held-out test split it reaches 95.8% per chunk and 98.8% per track; on the external Deep Pianist Identification benchmark the same recipe reaches 96.9% track accuracy, above the published ResNet-50 baseline of 94.4%. The paper uses this model as its measuring instrument: it scores generated music and locates the most characteristic moments within a performance.

Files

File	Contents
best.pt	Classifier weights (state dict)
config.json	Architecture summary

Usage

from llama_pijama.evaluation import load_model, run_inference

model, _ = load_model("best.pt", model_name="medium", num_classes=12, device="cpu")

Track-level scores aggregate chunk predictions by majority vote, with exact vote ties broken by mean logit.

Code and evaluation scripts: https://github.com/almostimplemented/jazz-pianist-style

Intended use and limitations

Trained on 12 pianists chosen for stylistic separability; it cannot recognize anyone else, and confident predictions outside that set are meaningless. Performer identification carries obvious privacy and attribution implications — this is a research artifact, not an attribution service.

Credits

Built on Aria (Apache-2.0) and the PiJAMA dataset.

@inproceedings{edwards2026jazzstyle,
  title     = {Learning Jazz Pianist Style with Cross-Attention Conditioning},
  author    = {Edwards, Drew and Maezawa, Akira and Dixon, Simon},
  booktitle = {Proc. Int. Society for Music Information Retrieval Conf. (ISMIR)},
  year      = {2026},
}