cstr / orukeet-GGUF

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

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:

= 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 ):

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.