Versioned, browser-ready piano transcription model artifacts used by Una Corda.
- Asset:
note_pedal.fp16.onnx - SHA-256:
8f2533cc2b466e0b5cb41be22747ed3e3a0ab0dc80b3d11f9b4e7fccda043779 - Size: 86,110,082 bytes
- Format: ONNX with float32 input/output and a mixed fp32/fp16 graph
The graph is derived from the public high-resolution piano transcription checkpoint
CRNN_note_F1=0.9677_pedal_F1=0.9186.pth published with Qiuqiang Kong et al.'s
High-resolution Piano Transcription with Pedals by Regressing Onsets and Offsets Times.
The original checkpoint is distributed through
Zenodo record 4034264, and the reference implementation
is archived at bytedance/piano_transcription.
Release assets are immutable. A replacement graph is published under a new release tag rather
than overwriting an existing model. See LICENSE for this repository's terms; upstream components
retain their original licenses and attribution requirements.
- Asset:
note_pedal.onnx - SHA-256:
192bd2dbdf06e0a8f49da10c459542b9e1809ab332f6afdceaa9bf6108a20d5f - Size: 154,214,196 bytes
- Format: ONNX with float32 input/output and a full fp32 graph
This is the exact fp32 reference graph. The smaller mixed-precision model-v1 artifact remains
available for evaluation, but it is not a production candidate because its measured note and
note-off regressions on amateur phone recordings are not acceptable.
- Asset:
note_pedal.int8.onnx - SHA-256:
c9bc6ecbf945f752206f1dcf614472bda9dff48b7481078f93480f6ab4fa2b31 - Size: 52,395,634 bytes
- Format: ONNX with float32 input/output and a weight-only, per-channel int8 neural backbone
This is the production browser default. Standard DequantizeLinear nodes restore the weights for
the original float32 Conv, MatMul, and GRU operators, keeping the graph portable to ONNX Runtime
Web's WASM fallback while reducing delivery by 66% versus model-v2. The precision-sensitive
STFT/log-mel front end stays fp32: quantizing it reduced the file to 39 MB but introduced false
notes on noisy phone audio.
Validation compared the complete event post-processor against model-v2 on four real 10-second
piano passages (clean and amateur-phone audio): three passages had identical note sets; the fourth
had one missed and one extra threshold event out of 154 fp32 notes. Matched onset drift was at most
0.24 ms mean, velocity drift at most 0.27 mean, and note-off drift at most 70.59 ms mean. Three
passages had identical pedal sets; the fourth had one missed and one extra threshold event out of
eight, with 7.14 ms matched mean onset drift.
- Asset:
yamnet-v1-waveform.fp16weights-a487d4bdb7f1.onnx - Model identity:
yamnet-v1-waveform-fp16weights - SHA-256:
a487d4bdb7f1f877b284c24902d5f02b85399961c5cd7a3bce0a20eea03d7e2a - Size: 8,636,805 bytes
- Format: ONNX opset 17 with float32 waveform I/O and compute; learned Conv/classifier weights stored as fp16 behind fp32 casts
This experimental piano-activity model is reproducibly derived from official
Google YAMNet v1 and pinned TensorFlow Models source commit
4d7bdd8c170ee90850f2f9ccef0f6d19b817de35. The graph retains the official
16 kHz raw-waveform frontend and emits the native 521 independent sigmoid
AudioSet scores, embeddings, and log-mel frames. Una Corda changed 28 learned
Conv/classifier initializers to fp16 storage with explicit fp32 casts; the
frontend, model I/O, biases, activations, and compute remain fp32.
The official model-version metadata labels the weights Apache 2.0. The Release
therefore includes APACHE-2.0.txt and THIRD_PARTY_NOTICES.md alongside the
model. They preserve the license and attribution and identify the ONNX/fp16
storage modifications. No upstream NOTICE file was present at the pinned
TensorFlow Models repository root.
This is not the transcription model and does not change ordinary recording or upload flows. It serves an explicit, foreground-only ambient experiment. Its frozen preliminary detector failed the local onset-p95 acceptance gate (1.16 s observed versus 1.0 s required), and YAMNet's weakly labeled AudioSet training provenance does not establish household or device fitness.
- Immutable source: GitHub Release
model-v1 - Immutable fp32 source: GitHub Release
model-v2 - Immutable int8 source: GitHub Release
model-v3 - Immutable YAMNet source: GitHub Release
yamnet-v1 - Browser fp16 URL: https://user198743.github.io/unacorda-models/note_pedal.fp16.onnx
- Browser fp32 URL: https://user198743.github.io/unacorda-models/note_pedal.onnx
- Browser int8 URL: https://user198743.github.io/unacorda-models/note_pedal.int8.onnx
- Browser YAMNet URL: https://user198743.github.io/unacorda-models/yamnet-v1-waveform.fp16weights-a487d4bdb7f1.onnx
- Browser YAMNet license: https://user198743.github.io/unacorda-models/APACHE-2.0.txt
- Browser YAMNet notices: https://user198743.github.io/unacorda-models/THIRD_PARTY_NOTICES.md
The GitHub Pages deployment downloads all Release assets and verifies their SHA-256 checksums before publishing them. This keeps the binaries out of Git history while providing stable, cross-origin URLs that browser applications can fetch.
Release tags and assets are immutable. Any replacement requires a new tag, checksum-named filename, and pinned Pages workflow entry. Model binaries remain absent from Git history.