Research
Papers, datasets, and a thesis in progress.
Learning Jazz Pianist Style with Cross-Attention Conditioning
Drew Edwards, Akira Maezawa, Simon Dixon. ISMIR 2026, Abu Dhabi.
Can a model learn not just to recognize who is playing, but to play in their manner? We fine-tune a transformer pretrained on piano MIDI on solo recordings of twelve jazz pianists, adding cross-attention over learned pianist embeddings. Conditioned continuations are attributed to the intended pianist 70% of the time, against 37% without conditioning, and the same classifier can point to the most characteristic moments in a performance.
Audio demos · arXiv · Code · Thread on X
Publications
MIDI-to-Tab: Guitar Tablature Inference via Masked Language Modeling
Edwards, Riley, Sarmento, Dixon. ISMIR 2024.
An encoder–decoder Transformer that assigns a string and fret to every note of a guitar performance, trained as masked language modeling: pretrained on 25,000 DadaGP tablatures, then fine-tuned on professionally transcribed performances. In a user study, guitarists rated its tablature above competing systems.
GAPS: A Large and Diverse Classical Guitar Dataset and Benchmark Transcription Model
Riley, Guo, Edwards, Dixon. ISMIR 2024.
Guitar-Aligned Performance Scores: 14 hours of classical guitar recordings from over 200 performers, aligned to scores with note-level MIDI, plus a benchmark transcription model that generalizes to unseen real-world audio.
A Data-Driven Analysis of Robust Automatic Piano Transcription
Edwards, Dixon, Benetos, Maezawa, Kusaka. IEEE Signal Processing Letters, 2024.
Piano transcription models overfit to the acoustics of their training data. We re-recorded the MAESTRO dataset on a Yamaha Disklavier in a studio and trained with data augmentation, reaching state-of-the-art note-onset accuracy on MAPS without seeing any of its training data.
High Resolution Guitar Transcription via Domain Adaptation
Riley, Edwards, Dixon. ICASSP 2024.
Aligning scores to commercially available guitar recordings yields training data for an instrument that has little of it. Fine-tuning a high-resolution piano transcription model on that data set a new state of the art on GuitarSet, zero-shot.
PiJAMA: Piano Jazz with Automatic MIDI Annotations
Edwards, Dixon, Benetos. Transactions of the ISMIR, 2023.
200+ hours of solo jazz piano, automatically transcribed to MIDI from thousands of recordings, with an analysis of how reliable the transcriptions are and what the data show about the pianists.
Thesis in progress: Modeling Jazz Piano: Symbolic Music Generation via Large-scale Automatic Transcription, supervised by Simon Dixon and Emmanouil Benetos at the Centre for Digital Music, Queen Mary University of London. Full list on Google Scholar.