GGUF conversion of TabCNN for CrispASR's
--tab surface. 833,982 parameters.
Per frame, the model emits six independent distributions over 21 fret classes
— one per guitar string. It contains no decoder: no inter-string coupling, no
temporal model, no search.
⚠️ These are emission SCORES, not a decided tablature. Turning them into a
playable fingering needs a constrained decoder (one note per string, fret
range, capo, hand span). Taking the argmax of this grid ignores every
playability constraint. CrispASR ships the scorer and leaves the decoder to
the caller.
Provenance and licence — verifiable, not asserted
Source artifact (fetch it and check the hashes yourself):
file
best_TabCNN_tablature_trancription_model (sic — typo is upstream's)
Every GGUF here carries these as metadata (general.source.url,
general.source.record_url, general.source.md5_upstream,
general.source.sha256), so provenance travels with the artifact.
⚠️ There is no Zenodo DOI for this record.10.5281/zenodo.11406378 looks
plausible and 404s — it does not exist. The record's actual DOI is the arXiv
one, 10.48550/arXiv.2405.14679 (resolves). Cite what resolves.
Licence evidence. The Zenodo API reports, for record 11406378:
Zenodo has no per-file licensing (a file object exposes only checksum, id,
key, links, size), so the record licence governs every deposited file. And
the record description states explicitly that the weights are part of the
deposit:
"The weights for the best performing model (TabCNN trained with
"GuitarProFX") in the paper are also provided."
That sentence is also what confirms this is the GuitarProFX-augmented
variant rather than the baseline — the distinction that matters, since the
baseline collapses from tablature F1 0.748 to 0.447 on real electric guitar
while the augmented one recovers to 0.585 (DAFx-24).
Required citation (the record asks for this explicitly):
Pedroza HE, Abreu W, Corey R, Roman IR. "Leveraging real electric guitar tones
and effects to improve robustness in guitar tablature transcription modeling."
In 27th International Conference on Digital Audio Effects (DAFx), 2024.
Upstream chain:
Model
TabCNN — Wiggins & Kim, Guitar Tablature Estimation with a Convolutional Neural Network, ISMIR 2019
dense0 Q4_0, head f32 — smallest, no measurable loss
tabcnn-f32.gguf
3.34 MB
0.7732
—
full precision, for parity work
(F1 on EGSet12 track 01 against its JAMS ground truth; the +0.0017 is noise on a
single 10 s clip — read it as "no loss", not "better".)
The output layer must stay at full precision
Only two tensors are quantizable at all: the conv stack is 3×3, so ne0=3,
far below any block size, and the biases are 32–126 wide. dense0.weight
(761 k of 834 k params) is effectively the whole model.
Quantizing dense0 costs nothing — but quantizing head.weight alongside it is
catastrophic:
head quantized
head preserved
q8_0
0.7676 (−0.0057)
0.7749
Q4_0
0.7153 (−0.0579)
0.7749
head.weight is 16 k params, 1.6 % of the file, and directly determines the
21-way per-string softmax. Preserving it makes Q4_0 as accurate as q8_0.
crispasr-quantize encodes this as an arch rule; anyone converting these
weights by another route should do the same.
⚠️ q4_k here is really Q4_0. No tensor has ne0 % 256 == 0
(dense0.weight is 5952; 5952 % 256 = 64), so k-quants cannot apply and the
quantizer falls back to Q4_0. The filename keeps the requested name; the content
is Q4_0.
Front end — read it from the file, do not assume it
CQT: sr 22050, hop 512, 192 bins, 24 per octave, fmin C1 (32.70 Hz)
-> amplitude_to_db(ref = max of the WHOLE clip) -> [-80, 0]
-> /80 + 1 -> [0, 1]
-> 9-frame centred context window
⚠️ fmin is C1, not the guitar's low E. Assuming E2 is the obvious guess and
it is wrong — and every wrong value still runs, producing plausible tensors
that pass shape and cosine checks while the model emits garbage. Measured on
EGSet12 track 01: fmin C1 → tablature F1 0.771, E1 → 0.040, E2 at 44.1 kHz →
0.001. All of these constants are stored as GGUF metadata
(tabcnn.sample_rate, tabcnn.fmin_hz, …) precisely so a consumer never has to
guess.
⚠️ ref = max of the whole clip is a per-clip normalisation, so features
cannot be computed streaming or chunked without changing them. This model is
two-pass by construction.
For real use, take the log-probabilities through the C ABI
(crispasr_session_tab → crispasr_session_tab_emissions) and run your own
constrained decoder. crispasr_session_tab_silent_class() tells you which class
means "not played" — read it rather than assuming it is the highest index.
Validation
crispasr-diff tabcnn against a reference dump from the amt-tools model, run
from the waveform (not replayed features, so the CQT is covered):
stage
cos
cqt_db
0.998925
conv0_relu
0.991401
conv1_relu
0.987366
conv2_relu
0.972109
pool
0.975496
dense0_relu
0.991824
logits
0.999661
End to end on EGSet12 track 01 against its JAMS ground truth:
P
R
F1
torch reference
0.7183
0.8315
0.7708
CrispASR (this GGUF)
0.7179
0.8378
0.7732
ΔF1 +0.0024, argmax agreement 98.57 %. The residual difference is the
front end — CrispASR uses direct Brown-kernel CQT against librosa's recursive
downsampling — and it costs nothing on the task metric.
Limitations
Trained on solo guitar. Mixed music is out of domain.
GuitarSet numbers overstate real-world performance; EGSet12 is the honest
reference point.
The scorer is per-frame and per-string independent, so it can emit physically
impossible combinations. That is expected — the decoder is where playability
is enforced.
