piano-visualizer / README.md
RaveVR's picture
Add files using upload-large-folder tool
079f438 verified
|
Raw History Blame Contribute Delete
1.82 kB
metadata
license: other
license_name: mit-and-cc-by-3.0
license_link: LICENSE.md
library_name: onnx
tags:
  - audio
  - music-transcription
  - piano
  - midi
  - onnx
  - onnxruntime-web

Piano Visualizer assets

Static files used by the Piano Visualizer on rave-tools.com: a free tool that turns a piano recording or a MIDI file into a falling-notes video, entirely in the browser. Nothing here runs on a server; the browser downloads these files and does all the work on the device.

transkun/: piano transcription model

Transkun V2 by Yujia Yan (MIT License), exported to ONNX in two graphs so it runs in ONNX Runtime Web:

File What it does
transkun_score.onnx audio segment → interval scores (which notes sound, from when to when)
transkun_attr.onnx detected intervals → velocity, refined onset and offset

The weights are the published Transkun V2 checkpoint, unchanged; only the graph format differs. Decoding (the semi-CRF search) runs in JavaScript.

salamander/: piano samples

Salamander Grand Piano V3 by Alexander Holm, licensed under CC BY 3.0. A Yamaha C5 sampled every minor third from A0 to C8 in 16 velocity layers, plus the string resonance played on key release.

Changes from the original: each sample is trimmed after its audible decay (-55 dB below its peak, with a 0.25 s fade) and re-encoded twice, as Opus in Ogg (ogg/) and as AAC in M4A (m4a/, for Safari). # in note names is written s (Ds1v8 = D#1, velocity layer 8).

License

See LICENSE.md: MIT for the Transkun model, CC BY 3.0 for the Salamander samples.