All articles

Building Song2Notation: From Audio to Sheet Music

A working note on the product and engineering decisions behind an AI music-transcription workflow.

2026-06-18 · 1 min read

This is an evolving build note. It describes the product direction without claiming that every planned capability is already complete.

Transcription is a workflow, not a single model call

The visible output may be a score, but the product has to coordinate upload, audio validation, asynchronous processing, intermediate status, notation generation, playback mapping, and export.

Each layer can fail independently. A useful product therefore needs explicit states, recoverable errors, and durable results — not just a progress spinner around a black-box inference call.

Interoperable output matters

MusicXML and MIDI are important because they let a musician continue working outside the product. Staff notation, Jianpu, and piano roll serve different reading and learning habits; they should be different views over consistent musical data rather than disconnected exports.

The next hard problem

Piano transcription quality remains the central model challenge. Product work can make uncertainty legible, but it cannot pretend inaccurate notes are correct. The longer-term goal is to pair better piano-focused models with editing tools that make correction fast and obvious.