pdf pdf |
|---|
Piano Score Generation Sample
A continuation produced by a compact Transformer (16.6M parameters) that generates piano scores at the notation level: the representation covers staves, voices, ties, slurs, dynamics, pedal and ornaments, rather than MIDI note events. The sample is provided as playable audio and as the underlying MusicXML score, next to the prompt it was given.
This is a demo of research output, not a model release. It accompanies a research internship at UC Berkeley (2026).
The model, briefly
A 16.6M-parameter causal Transformer over a compound-token representation of **kern notation, in the spirit of Compound Word Transformer but applied to score notation rather than performance events. Trained on ~206k piano scores from public symbolic-music corpora (details to follow with the associated publication). The sample comes from the final internship variant.
The research question was whether a representation designed for notation (one token per musical event, its attributes predicted by typed output heads) lets a small model spend its capacity on musical structure rather than on the surface syntax of the format, and so reach acceptable generation quality at a fraction of the size of published notation-level systems. The decoder itself is deliberately plain: a standard GPT-style Transformer with no architectural novelty beyond the factorised output layer. The contribution lies in the representation, not in the network.
The sample
One 76-measure score in three formats: sample.mp3 to listen, sample.pdf to read, and sample.mxl to open in a score editor.
Measures 1–5 are the prompt: a short piano improvisation, recorded and transcribed for this purpose, not in the training corpus, so nothing here can be memorised. Everything from measure 6 onward is generated. At the file's tempo (2/4, ♩ = 120) the prompt is the first 5 seconds of the audio; the full sample lasts about 80 seconds.
Sampling. Temperature 0.9, single pass. The model produced 112 measures; the file is truncated at measure 76. Nothing else was edited: no note, rhythm, tie or beam was changed. The metronome marking in the file header is not generated by the model and is set at conversion time.
Selection. One of the first generations run with this prompt, kept after listening to a handful at three temperatures. This particular generation contains no dynamics, pedal or slur markings, which is consistent with a prompt that has none. On the Chopin prompt mentioned below, the same model produces slurs, dynamics, pedal, ornaments and articulations throughout.
A note on the prompt. The prompt is harmonically static: four measures of a single A-minor triad in repeated block quarter notes, then a bass shift in measure 5. There is no chord progression for the model to carry forward, and the continuation should be read in that light. On harmonically richer prompts, such as the opening of Chopin's Nocturne op. 9 no. 2, the model does keep a coherent harmonic accompaniment going under a melodic line over several measures.
Known limitations, please read before judging
- Repetitive loops. The model regularly falls into repeated motifs, sometimes repeated tokens, and struggles to exit them. This is the dominant defect of current outputs.
- Qualitative evaluation only. No validated out-of-sample metric is available yet; the assessment behind this sample is listening and score reading by one person, without a blind protocol.
- Two orders of magnitude fewer parameters than published notation-level systems (NotaGen, MuPT). This sample is not meant to compete with them.
What this is not
Not a model release: no weights, code or inference script are provided for now.
Licence
Prompt (own recording), generated score and audio: CC BY-NC-SA 4.0.
- Downloads last month
- 29