Orukeet — GGUF
GGUF conversions of oruk/orukeet (checkpoint r3) for
CrispASR.
Orukeet is Oruk AI's fine-tune of nvidia/parakeet-tdt-0.6b-v3.
Half of the encoder's temporal depthwise convolution kernels (12,288 of 24,576) were replaced with fitted
Gabor functions and frozen, and the model was then re-adapted. Tensor shapes and operators are unchanged, so it
runs on CrispASR's existing Parakeet runtime: 25 European languages, automatic language detection, TDT word
timestamps. See the technical report.
About the benchmark numbers. Orukeet's final adaptation step trained on all 2,939 LibriSpeech test-other recordings (report: "Test-other is used for training, checkpoint selection and re-evaluation"). Its LibriSpeech test-other score is therefore not a held-out result. Judge it on other test sets.
Files
| File | Size | Notes |
|---|---|---|
orukeet-f16.gguf |
1.26 GB | F16 |
orukeet-q8_0.gguf |
674 MB | Q8_0 |
orukeet-q5_0.gguf |
470 MB | Q5_0 |
orukeet-q4_k.gguf |
402 MB | Q4_K, the default for -m orukeet |
Converted from orukeet-v0.1.0.nemo at revision 555136b5 (SHA-256 031c8dda…73b56, matching the report)
with models/convert-parakeet-to-gguf.py, and quantised with crispasr-quantize.
Verification against NeMo
The same WAVs were run through the original checkpoint in NeMo (greedy TDT, CPU) and through CrispASR:
| clip | NeMo | F16 | Q8_0 | Q4_K |
|---|---|---|---|---|
en, samples/jfk.wav |
And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country. | = | = | = |
| de | Guten Morgen. Die Sitzung beginnt heute um neun Uhr im großen Saal. | = | = | punct. |
| fr | Bonjour à tous, la réunion commence demain matin à huit heures. | = | = | = |
| es | Buenas tardes. El tren para Madrid sale a las cinco y media. | = | = | punct. |
= means identical text. punct. means identical words with one sentence-final period rendered as a comma.
Per-stage comparison with crispasr-diff against a NeMo reference dump of the same checkpoint
(tools/dump_reference.py --backend parakeet), F16 GGUF, worst frame (cos_min):
| stage | en (jfk, 138 frames) | de (61 frames) |
|---|---|---|
| mel spectrogram | 1.000000 | 1.000000 |
| pre-encode | 0.999998 | 0.999986 |
| encoder layers 0–23 (worst layer) | ≥ 0.999981 | ≥ 0.999942 |
| encoder output | 0.999883 | 0.999988 |
The quantised files differ from NeMo by ordinary quantisation error (Q8_0 encoder-output cos_mean 0.9996 / 0.9961),
and the transcripts above are the end-to-end check for them.
Usage
# auto-download the Q4_K file and transcribe
crispasr -m orukeet -f audio.wav
# or point at a downloaded file; the Parakeet backend is detected from the GGUF
crispasr -m orukeet-q8_0.gguf -f audio.wav -osrt
Every Parakeet option applies, including VAD, chunking, word timestamps and subtitle output.
Licence and attribution
The weights are CC BY-SA 4.0, inherited from Orukeet, so these GGUF files are distributed under the same licence: you must credit the authors and share derivatives under the same terms.
- Base model:
nvidia/parakeet-tdt-0.6b-v3, CC BY 4.0, NVIDIA. - Fine-tune:
oruk/orukeet, CC BY-SA 4.0, Oruk AI (Roll, Yi, Marşan, Grenez, Stein, Mrkaic, Padjin, Zeljkovic, Graham). - Conversion and runtime: CrispASR. The only change made here is the format conversion and quantisation.
- Downloads last month
- 42
5-bit
8-bit
16-bit