YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Orukeet v0.1.0, combined INT8 ONNX
Local, offline transcription with Parakeet-compatible Rust engines. The encoder is unchanged from the canonical Orukeet INT8 export; the decoder and joiner are combined into one graph, without requantization.
Files and use
Download encoder-model.int8.onnx, decoder_joint-model.int8.onnx, nemo128.onnx, vocab.txt, and config.json into one directory, together with the license and notice files. Input audio is mono, 16 kHz. The configuration declares the NeMo TDT model type, 128 acoustic features and 8x subsampling. This is a local model; it does not require the Oruk API.
The joined graph was checked against the original decoder/joiner over eight recurrent steps: logits, recurrent states, greedy tokens and durations matched exactly. Real inference and timestamp bounds passed in Meetily's Rust core (ort 2.0.0-rc.10 / ONNX Runtime 1.22) and AutoSubs' transcribe-rs 0.3.11 engine (ort 2.0.0-rc.12 / ONNX Runtime 1.24.4).
Measured comparison
AMD EPYC 9B45 CPU, default engine thread settings, Linux; first five FLEURS validation clips in each of en_us, de_de, es_419, fr_fr, ru_ru and uk_ua. Thirty clips, 629 normalized reference words, two timed passes per model in baseline/candidate/candidate/baseline order, 60 calls per model. Warm inference includes preprocessing and decoding, excluding WAV reads, model loading, VAD, GUI, network and editor integration. These are sequential processing rates, not server concurrency limits.
| Engine | Model | Median | p95 | Audio seconds / processing second | WER |
|---|---|---|---|---|---|
| Meetily | Parakeet v3 | 261.6 ms | 425.2 ms | 36.62 | 10.02% |
| Meetily | Orukeet | 195.0 ms | 324.3 ms | 47.30 | 9.70% |
| AutoSubs | Parakeet v3 | 299.5 ms | 501.2 ms | 30.63 | 9.70% |
| AutoSubs | Orukeet | 230.0 ms | 378.2 ms | 40.13 | 9.70% |
The baseline is istupakov/parakeet-tdt-0.6b-v3-onnx at 8f23f0c03c8761650bdb5b40aaf3e40d2c15f1ce; its four runtime files were also byte-identical to Meetily's shipped CDN files. This small, fixed sample does not establish general accuracy superiority. German and Russian WER regressed in both engines; other language results were mixed. Training overlap was not excluded. No claim is made about all 25 languages or other hardware. Both models use the same application preprocessor. Its short-input valid-frame convention differs by one frame from the NeMo Python preprocessor; exact NeMo feature parity is not claimed.
Sources and attribution
- Orukeet weights and original INT8 graphs: oruk/orukeet, revision 55a984d46f68323301837194ce647c702f55facc. Weight terms: CC BY-SA 4.0; see LICENSE-WEIGHTS and NOTICE.md, including upstream NVIDIA Parakeet attribution.
- Decoder composition: Derek Zeng (derekszen), Voxtype PR #752, script import-orukeet-onnx.py at a20aa75371191f3d73332051ef7aff5cb8b59c51. MIT license reproduced in LICENSE-CONVERTER.txt. Converter source is included for reproducibility.
- Unmodified nemo128.onnx and config.json: Ivan Stupakov, parakeet-tdt-0.6b-v3-onnx, revision 8f23f0c03c8761650bdb5b40aaf3e40d2c15f1ce, CC BY 4.0. See LICENSE-PREPROCESSOR.txt. No endorsement is implied.
SHA-256 digests and sizes are in manifest.json. Benchmark raw outputs and per-language summaries are included in validation/.
- Downloads last month
- 13