Researchers at ISMIR 2026 in Abu Dhabi present a model that learns to generate piano music in the style of twelve jazz masters, inspired by Dick Hyman's 1994 book of études mimicking pianists from Scott Joplin to Bill Evans. The team fine-tuned Aria, a 16-layer transformer pretrained on piano MIDI, on solo performances from the Pi JAMA dataset. They added a gated cross-attention layer that reads a learned embedding for each pianist — including Art Tatum, Erroll Garner, Oscar Peterson, and Hyman himself — across the last eight transformer layers. A learned gate scales the conditioning influence, starting at 0.1 so fine-tuning begins from Aria's base behavior.
To evaluate style transfer, they slid a pianist classifier along generated continuations. Conditioned outputs were attributed to the intended pianist 70% of the time, versus 37% without conditioning. A second classifier trained only on generated music identified real recordings with 95% accuracy. Perplexity on held-out data proved nearly blind to style differences, confirming that style emerges during free generation. The model generates twelve distinct continuations from the same prompt — such as "Ain't Misbehavin'" — with each pianist's characteristic note density: Garner packs notes into 1:26 while Cedar Walton spreads them across 2:45.
Source: Hacker News · Summarized by HeadlinesBriefing