We fine-tune Aria, a transformer pretrained on piano MIDI, on solo performances by twelve jazz pianists from the PiJAMA dataset, adding a gated cross-attention layer that reads a learned embedding for each pianist. To check whether the style comes through, we slide a pianist classifier along the generated music: conditioned continuations are attributed to the intended pianist 70% of the time, against 37% without conditioning. A second classifier trained only on generated music then identifies real recordings with 95% accuracy.

Listen first; how it works is further down.

The opening bars of “Ain’t Misbehavin’” are played by one of us (Drew). Everything after the dashed line is generated: twelve takes of the same opening, each conditioned on a different pianist. Pick a pianist to hear their take from the top.

Every take generates the same number of notes, so they end at different times: Erroll Garner packs them into 1:26, Cedar Walton spreads them across 2:45. That difference in density is itself part of a pianist’s signature.

A shared prompt pulls every pianist toward the same tune. Here each pianist instead continues a few bars of their own playing. The strip under each take shows what our classifier heard as it slid along the continuation, one cell per window of about 300 notes: gold where it named the intended pianist, mauve where it named someone else. The two takes per pianist are the best of eight we scored; the line under them says how the rest did. Or switch on the blindfold and guess for yourself.

The classifier can also point at moments. On a real performance it is near-certain almost everywhere, so instead we ask where it is *even more sure than usual*: its margin for the true pianist over the runner-up, compared with its own average across that performance. Below, for one held-out recording per pianist, that curve over the whole piece and fifteen seconds from its highest and lowest points.

We start from Aria (Bradshaw et al., ISMIR 2025; code), a 16-layer transformer pretrained on a large corpus of piano MIDI. Into each of its last eight layers we insert a cross-attention block: the music attends to a small learned embedding for the chosen pianist, four vectors per pianist. A learned gate scales what the block adds, starting at 0.1, so fine-tuning begins from Aria’s own behaviour and learns how much to listen. Because the embedding is attended to at every step, the conditioning does not fade as generation goes on, the way a prompt prefix does.

How can we tell whether the model has learned a pianist’s style? The standard yardstick for a generative model, perplexity on held-out music, turns out to be nearly blind to it: given the real preceding notes, the next one is predictable whoever is playing, so conditioning barely moves the score. Style shows up when the model generates freely and has to stay in character on its own output. So instead we let it play, and ask a pianist classifier who it sounds like. Agreement is how often the classifier names the intended pianist, in windows slid along each continuation.

Perplexity (lower is better) barely separates the two fine-tuned models; agreement nearly doubles. Continuations are 4096 tokens from 256-token prompts; chance agreement is 8%.

• Sliding-window agreement, above. The classifier identifies 98.8%

  of held-out songs, and the conditioned model’s lead holds from the start of a continuation to its end. The strips under the scored takes are this measurement.

• Synthetic transfer. A fresh classifier trained only on

  generated music identifies real recordings: 87% of 1024-token chunks and 95% of songs, within nine points of one trained on real data. The scored takes are samples of that training data.

• Characteristic regions. Turned on real performances, the

  classifier points to where a pianist’s style is most concentrated — where the style lives.

The paper has the details: per-pianist results, the mismatch experiment (prompting with one pianist and conditioning on another), memorization checks, and a from-scratch classifier that confirms the transfer result.

Everything here is MIDI, rendered in your browser on a sampled piano. The model was trained on automatic transcriptions of commercial recordings, so dynamics and pedalling are approximate, and the rendering is plainer than the records.

The twelve pianists were chosen for separability: they are the twelve of PiJAMA’s thirty whose recordings a pretrained model already tells apart most easily. Within them the model imitates some far better than others — across the paper’s evaluation, continuations were attributed to the intended pianist 96% of the time for Hank Jones and Dick Hyman, but only 29% for Cedar Walton.

The scored takes are selected, not copied. They are samples from the corpus of generated music that the paper’s synthetic-only classifier learned from; for each pianist we show the two highest-scoring of eight candidates. Their prompts come from the training recordings, and the paper checks that the continuations do not copy them: they resemble their closest training performance less than real held-out performances do.

The scores come from a classifier, not from listeners. It is a strong one (98.8% of held-out songs), but it has habits: many of its mistakes on generated music land on Dick Hyman — fitting, perhaps, for a pianist who made a career of playing in everyone else’s style. A listening study is the natural next step.