baroned/musicnote-ondevice-models

Musicnote on-device transcription models (ONNX)

ONNX exports of published music-transcription models, used by the audio → sheet music tool on musicnote.app. The browser downloads them once and runs them on the user's device with onnxruntime-web; recordings never leave the device.

These are conversions, not new models. Each file keeps the licence of the weights it was converted from. Every file's size and SHA-256 are in manifest.json.

Models, sources and licences

Files	Original model	Weights licence	Changes made here
guitar_crnn.onnx, guitar_crnn.onnx.data	Guitar transcription CRNN (GAPS), guitar-gaps-paper-version-12200_iterations.pth from xavriley/midi-transcription-models — Riley et al., arXiv:2402.15258	MIT	exported to ONNX (opset 18), outputs as positional tensors
piano_kong.onnx, piano_kong.onnx.data	High-resolution piano transcription, Kong, Li, Song, Wan, Wang, Ma — bytedance/piano_transcription, checkpoint note_F1=0.9677_pedal_F1=0.9186.pth, Zenodo 4034264	CC BY 4.0	exported to ONNX (opset 18); note model only — the pedal sub-network is left out
tsumugi_other_v1_5.*	Tsumugi instrument-agnostic AMT, best_model_other_v1_5.pth — anime-song/tsumugi, anime-song/instrument_agnostic_amt	MIT	scoring network exported to ONNX with compact outputs (positive-score interval candidates only, int32 indices); boundary/instrument heads exported separately as heads.json + heads.bin
beat_this.onnx	Beat This! beat and downbeat tracker, final0 checkpoint — Foscarin, Schlüter, Widmer (ISMIR 2024), CPJKU/beat_this	MIT	exported to ONNX
basic_pitch_nmp.onnx	Basic Pitch — spotify/basic-pitch, the package's own nmp.onnx	Apache-2.0	none (copied as shipped)

Attribution (CC BY 4.0, piano model): "High-resolution Piano Transcription with Pedals by Regressing Onset and Offset Times", Qiuqiang Kong, Bochen Li, Xuchen Song, Yuan Wan, Yuxuan Wang, 2020 — weights from Zenodo record 4034264, licensed under CC BY 4.0. Changes: converted to ONNX, pedal sub-network removed.

MIT and Apache-2.0 notices of the original projects apply to the corresponding files; see the linked repositories for the full texts.

Training data

The original authors trained these models on public research datasets (for example MAESTRO, MusicNet, GuitarSet/GAPS). Some of those datasets carry non-commercial terms. The weights themselves are released under the licences above; whether a dataset's terms extend to weights trained on it is not settled. If you reuse these files, check the original projects and datasets for your own use.

Not affiliated

These conversions are not made or endorsed by the original authors.