tabcnn-onnx

TabCNN (ONNX) — guitar tablature estimation

ONNX exports of TabCNN (Wiggins & Kim, ISMIR 2019) — a small CNN that estimates guitar tablature (which fret is played on each of the 6 strings) from audio, frame by frame. Both exports end in a per-string LogSoftmax head so a decoder can consume log-probs directly, and share the frozen output contract: [N, 6, 21], class 0 = string silent/closed, class k = fret k−1 (class 1 = open, class 20 = fret 19).

Built for the pure-Dart onnx_runtime_dart runtime (no native ORT / FFI, web-capable), but it's standard ONNX.

Two models — pick by input

Use tabcnn-gpfx.onnx unless you specifically want raw-magnitude features — it's the robust variant (the vanilla model collapses on distorted/electric tones).

• tabcnn-gpfx.onnx — sha256 8d9ce59157bdab37fb4816d32d7f29f3da0cdbf3c7876707c819af4d1f88e6b7, 833,982 params
• tabcnn.onnx — sha256 15c58000ed2d1deb3d3fc07581aa1823482dad91d913399dc0209ef240ad8a51
• tabcnn-cqt.bin — sha256 4e5dfa1f10f76545a30cbfd3224431503dbad943b1def78624632284e6df597a

Inputs / outputs (both models)

• Input input : float32[N, 192, 9, 1] — per frame, a 9-frame context window of the 192-bin CQT (bins × context × channel). N = a batch of windows.
• Output output : float32[N, 6, 21] — per string, a LogSoftmax over 21 classes; class 0 = closed, class k = fret k−1.
• Frame hop = 512 / 22050 = 0.023220 s (≈ 43 fps).

CQT front-end — the #1 correctness risk (differs by variant!)

Both use the same CQT geometry (sr 22050, hop 512, n_bins 192, bins_per_octave 24, fmin C1 = 32.703 Hz), and tabcnn-cqt.bin is that 192-bin filterbank precomputed (banded, n_fft 32768, boxcar STFT) for a librosa-free front-end — it matches librosa.cqt at cosine 0.999947 / median magnitude ratio 0.9999.

The normalization after the CQT differs:

import librosa, numpy as np
C = np.abs(librosa.cqt(y, sr=22050, hop_length=512, n_bins=192, bins_per_octave=24))  # [192, frames]

# tabcnn-gpfx.onnx  (GuitarProFX — recommended):  per-clip dB, then [0,1]
feats = librosa.amplitude_to_db(C, ref=np.max)          # [-80, 0], per clip
feats = (feats - feats.min()) / (feats.max() - feats.min() + 1e-9)

# tabcnn.onnx  (vanilla):  RAW magnitude, no log/norm
# (peak-normalize the *waveform* first: y = librosa.util.normalize(y))
feats = C

repr_ = np.swapaxes(feats, 0, 1)   # [frames, 192]; window = pad 4 each side, [f:f+9] -> [192,9,1]

With tabcnn-cqt.bin, magnitude = |Σ band·boxcarSTFT| / √length (the blob's mean/std header fields are 0/1 and unused). Apply the per-variant normalization above to that magnitude.

Performance & verification

• tabcnn-gpfx.onnx: EGSet12 (12-track, frame-level tab F1) = 0.59 micro / 0.55 macro — MEASURED here (per-track 0.30–0.80), matching the paper's reported 0.585 for this model; on clean GuitarSet it reaches ~0.77. A big lift over the vanilla model's ~0.45 EGSet12 zero-shot.
• tabcnn.onnx: held-out GuitarSet tab F1 0.745 (paper 0.748).
• Both run on onnx_runtime_dart faithfully vs the reference (per-string argmax agreement; standard ops only — Conv/MaxPool/Relu/MatMul/LogSoftmax + a class reorder for gpfx).

Class-layout note (gpfx)

The GuitarProFX (amt-tools) model natively uses class 20 = silence, class k = fret k. The export remaps that to the shared contract above (class 0 = silent, class k = fret k−1) with a roll, so both .onnx files present the identical [6,21] layout to the decoder.

Licence & attribution (CC BY 4.0)

• Vanilla: trained here on GuitarSet (CC BY 4.0). Attribution to GuitarSet required.
• GuitarProFX: weights from Zenodo 11406378 (best_TabCNN_tablature_trancription_model, CC BY 4.0), the DAFx-24 GuitarProFX model, built on Cwitkowitz's amt-tools (MIT). Attribution to GuitarSet + Pedroza et al. required.

@inproceedings{xi2018guitarset, title={GuitarSet: A Dataset for Guitar Transcription},
  author={Xi, Qingyang and Bittner, Rachel M. and Pauwels, Johan and Ye, Xuzhou and Bello, Juan Pablo}, booktitle={ISMIR}, year={2018}}
@inproceedings{wiggins2019tabcnn, title={Guitar Tablature Estimation with a Convolutional Neural Network},
  author={Wiggins, Andrew and Kim, Youngmoo}, booktitle={ISMIR}, year={2019}}
@inproceedings{pedroza2024guitarprofx, title={Leveraging Real Electric Guitar Tones and Effects to Improve Robustness in Guitar Tablature Transcription Modeling},
  author={Pedroza, Hegel and others}, booktitle={DAFx}, year={2024}}

Reproduction (vanilla) + export scripts: onnx_runtime_dart/tool/tabcnn/.

Provenance — mixed authorship, and outside the GPAI definition

Added 2026-08-02 during an account-wide provenance review.

This repository holds weights of two different origins, which is why the per-file table above matters more here than in a typical conversion repo: tabcnn.onnx (vanilla) was trained here on GuitarSet (CC BY 4.0), while the GuitarProFX weights were re-exported from Zenodo record 11406378 (Pedroza et al.). Attribution obligations differ between them and are set out above.

Most cstr/* repositories are conversions where the upstream research team remains the provider of the model. That description is only half true here, and a blanket "format conversion only" statement — of the kind added across the conversion repos in this review — would have been inaccurate for this one. It is deliberately not made.

EU AI Act Art. 53 does not apply. Art. 53 binds providers of general-purpose AI models, defined by Art. 3(63) as models displaying "significant generality" and capable of "competently performing a wide range of distinct tasks". A CNN transcribing guitar tablature from audio does not meet that definition, so the Art. 53(1)(c) copyright-policy and 53(1)(d) training-content duties are not engaged for either set of weights. The training-data and licence documentation above is published because attribution is required and because it is useful — not because Art. 53 compels it.