LanOss/mobimml-piano-transcription

MobiMML Piano Transcription (ONNX)

ONNX conversion of ByteDance's high-resolution piano transcription model, packaged for in-browser inference (onnxruntime-web / WebAssembly) in MobiMML — a web app that converts piano audio into 마비노기 모바일 MML scores. No server-side inference is involved; this repository only serves the model file.

• piano_transcription.onnx — the full note + pedal model (opset 17), input waveform (float32, [1, 160000] = 10 s @ 16 kHz), outputs: reg_onset_output, reg_offset_output, frame_output, velocity_output, reg_pedal_onset_output, reg_pedal_offset_output, pedal_frame_output.

Attribution (원저작자)

This model is a format conversion of the checkpoint released by the authors of:

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. arXiv:2010.01815

• Original code: bytedance/piano_transcription (Apache 2.0)
• Original checkpoint: Zenodo record 4034264 (CRNN_note_F1=0.9677_pedal_F1=0.9186.pth, CC BY 4.0)

Changes made: the PyTorch checkpoint was exported to ONNX (opset 17, constant folding, dict outputs flattened to a fixed-order tuple) using tools/export_onnx.py. No weights were modified or fine-tuned.

License

The model weights are distributed under CC BY 4.0, the license of the original checkpoint. If you redistribute or build upon this file, credit the original authors above.

Usage (MobiMML)

The MobiMML app downloads this file automatically on first use of the 고정밀(High-res) tab and caches it in the browser (IndexedDB). Manual download works too — grab piano_transcription.onnx and select it in the app.

https://huggingface.co/LanOss/mobimml-piano-transcription/resolve/main/piano_transcription.onnx