Learning Jazz Pianist Style with Cross-Attention Conditioning
Summary
ISMIR 2026 presents Learning Jazz Pianist Style with Cross-Attention Conditioning, fine-tuning a transformer on piano MIDI with a cross-attention block conditioned on pianist embeddings. Evaluation shows the conditioned model achieves about 70% agreement with the intended pianist on generated continuations, while a synthetic-music classifier trained on generated data can identify real recordings with about 95% accuracy. The work includes open-source code and discusses measuring style beyond perplexity.