Piano transcription is an open-source high-resolution piano transcription system by ByteDance that converts raw audio recordings of piano performance into symbolic MIDI files — detecting note onsets, offsets, pitch, velocity, and even pedal usage. The system is implemented in Python (PyTorch) and is capable of accurate transcription of polyphonic piano recordings, even with complex passages and pedal techniques, making it suitable for classical piano music. By using this transcription tool, users can transform live performance audio (or recordings) into editable, machine-readable MIDI — enabling tasks such as analysis, editing, remixing, or generation of piano music. The authors used this system to build a large-scale classical piano MIDI dataset (see next project), but as a standalone tool it enables researchers, musicians, or hobbyists to transcribe their own piano recordings automatically.
Features
- High-resolution polyphonic piano transcription: detects note onset, offset, pitch, velocity from audio recordings
- Pedal detection included — captures sustain pedal usage for expressive performance details
- Outputs standard MIDI format — enabling editing, playback, symbolic analysis or further music processing
- Implemented in Python / PyTorch — accessible and modifiable by researchers, musicians, or hobbyists without specialized environment
- Useful for both music research / analysis and creative music workflows (remixing, composition, symbolic generation) thanks to expressive transcription
- Open-source (Apache/MIT) — free to use, adapt, and integrate into pipelines or projects