Clement11800/pianino-model

Pianino — piano transcription model (notes only, ONNX)

ONNX conversion of the note transcription part of the high-resolution piano transcription model by Qiuqiang Kong et al. (ByteDance), used on-device by the Pianino app to turn a solo piano recording (MP3, M4A, WAV…) into notes.

What it does

• Input audio: float32[1, 160000] — 10 seconds of mono audio at 16 kHz.
• Outputs, each float32[1, 1001, 88] — 100 frames per second × the 88 piano keys (A0–C8):
  • reg_onset_output — onset regression
  • reg_offset_output — offset regression
  • frame_output — key down
  • velocity_output — velocity

Longer recordings are cut into 10-second segments overlapping by half; only the middle of each segment is kept, as in the original inference code. Notes are then decoded from the four outputs with the original post-processing (onset/offset peak picking with sub-frame regression).

File	Size	SHA-256
piano-notes-v1.onnx	98,689,388 bytes	9e33ffc0382fb023e0b1aac2abd0c3cb26a5d56fbcc246d7f2c0fc33c569f2ba

Modifications from the original

• Only the note sub-model is kept; the sustain-pedal sub-model is removed.
• Converted from PyTorch to ONNX (opset 17) with a fixed 10-second input. Outputs match PyTorch within 1e-6.
• The weights themselves are unchanged.

Origin and licence

• Original model and weights: Qiuqiang Kong, Bochen Li, Xuchen Song, Yuan Wan, Yuxuan Wang, High-resolution Piano Transcription with Pedals by Regressing Onset and Offset Times, IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021.
• Weights: zenodo.org/record/4034264, released under CC BY 4.0.
• Original code: github.com/bytedance/piano_transcription and piano_transcription_inference.

This converted model is shared under the same CC BY 4.0 licence. If you use it, credit the original authors above.

Limits

Trained on MAESTRO (solo classical piano on a Disklavier). Works best on clean solo piano recordings; voices, other instruments, heavy reverb or phone recordings reduce accuracy.

Citation

@article{kong2021high,
  title   = {High-resolution Piano Transcription with Pedals by Regressing Onset and Offset Times},
  author  = {Kong, Qiuqiang and Li, Bochen and Song, Xuchen and Wan, Yuan and Wang, Yuxuan},
  journal = {IEEE/ACM Transactions on Audio, Speech, and Language Processing},
  volume  = {29},
  pages   = {3707--3717},
  year    = {2021}
}