E Etude
arXiv:2509.16522 · Preprint Piano cover generation

Expressive piano covers, steerable by design.

A three-stage deep learning system that lifts a source track into a symbolic form, imposes musical structure, and decodes a stylized performance — conditioned on high-level attributes like polyphony, rhythm intensity, and note sustain.

Extract strucTUralize DEcode
etude • piano roll
Polyphony Rhythm Sustain
realtime
etude/overview.mp4

Pipeline

Three stages, end-to-end

Each stage has a focused responsibility, and together they turn a raw audio track into a stylized, expressive piano cover.

01 · Extract

Audio → Symbolic

Lift musical content out of the source recording into a symbolic representation we can reason about.

02 · strucTUralize

Impose structure

Organize content into a structured form that captures phrasing, density, and rhythmic hierarchy.

03 · DEcode

Render a performance

Decode a stylized piano performance, steerable via high-level attributes like polyphony and rhythm intensity.

Demo 01

Model Comparison

Listen side-by-side: Etude variants against baselines (Music2Midi, PiCoGen2, AMT-APC) across 100 songs in four genres.

Open demo

Demo 02

Interactive Style Control

Move the sliders — polyphony, rhythm intensity, note sustain — and hear how Etude's output bends to your taste in real time.

Try it live

BibTeX

@article{chen2025etude,
  title   = {Etude: Piano Cover Generation with a Three-Stage Approach-Extract, strucTUralize, and DEcode},
  author  = {Chen, Tse-Yang and Joung, Yuh-Jzer},
  journal = {arXiv preprint arXiv:2509.16522},
  year    = {2025}
}