Learning Jazz Pianist Style with Cross-Attention Conditioning
5 hours ago
- Project inspired by Dick Hyman's book of études written in the style of jazz pianists.
- Fine-tunes Aria transformer on solo performances of 12 jazz pianists from PiJAMA dataset.
- Uses gated cross-attention to condition generation on learned pianist embeddings.
- Evaluates style transfer via a classifier, not perplexity; conditioned continuations show 70% attribution accuracy vs 37% without conditioning.
- Classifier trained on generated music can identify real recordings with 95% accuracy.
- Style manifests in density and timing differences (e.g., Erroll Garner vs Cedar Walton).
- Blindfold test and moment-by-moment classifier analysis provided.
- Perplexity is nearly blind to style; agreement metric doubles with conditioning.
- Results vary across pianists: high accuracy for Hank Jones and Dick Hyman, low for Cedar Walton.
- Paper includes per-pianist results, mismatch experiments, memorization checks, and synthetic-only classifier transfer.