orukeet

Nathan Roll^1,2 · Irene Yi^1,2 · Büşra Marşan^1,2
Vianney Grenez¹ · Gabriel Stein⁴ · Momcilo Mrkaic⁵
Pavle Padjin⁵ · Vladimir Zeljkovic⁵ · Calbert Graham^1,3

¹ Oruk AI

Orukeet is a 25-language speech recognizer built from NVIDIA Parakeet TDT 0.6B v3. It replaces half of the encoder's temporal depthwise filters with 12,288 fitted, frozen Gabor kernels and trains the remaining parameters on multilingual and multi-accent data.

Orukeet outperforms Parakeet on 61 of 74 tested splits, including LibriSpeech test-clean (1.46% vs. 1.53% WER), test-other (2.86% vs. 3.14%), and FLEURS English (3.82% vs. 4.28%). Across all 25 FLEURS languages, pooled WER is 9.85% vs. 11.01%, a 10.6% relative reduction. Final adaptation and checkpoint selection use LibriSpeech test-other.

Use Orukeet for recordings, media, batch transcription, server workers and interactive applications. NeMo, ONNX INT8, native Q8 and native F16 all derive from the same r3 release checkpoint (031c8ddab484).

Code · OpenWhispr PR · Technical report · Artifact hashes

Run Orukeet with Transformers

A standard FP32 Transformers export is available at the repository root. It uses ParakeetForTDT without custom remote code and works with Buzz's existing Hugging Face model option. See setup, conversion provenance and runtime qualification. The NeMo evaluation below remains the source-model benchmark; the Transformers export has separate compatibility measurements.

Run Orukeet with NeMo

Use a CUDA-enabled PyTorch environment with nemo_toolkit[asr]==3.0.0 and huggingface-hub. The recorded source environment lists the exact package versions used for evaluation.

from huggingface_hub import hf_hub_download
from nemo.collections.asr.models import ASRModel

checkpoint = hf_hub_download(
    "oruk/orukeet", "orukeet-v0.1.0.nemo",
    revision="555136b50265a132d4cea0d35560c26fc4f657ab",
)
asr = ASRModel.restore_from(checkpoint)
asr.eval()
print(asr.transcribe(["recording.wav"], return_hypotheses=True)[0].text)

orukeet fetch source retrieves the same hash-checked checkpoint. Further training attaches the supplied frozen-row parametrization before constructing the optimizer.

Architecture

The model retains Parakeet's 627,008,134 parameters, 24-layer FastConformer encoder, token-and-duration transducer and tokenizer. Each encoder block contains 1,024 nine-tap temporal depthwise filters. A selected filter stores its own fitted Gabor function:

$$g(t)=A\exp\left[-\frac{(t-\mu)^2}{2\sigma^2}\right]\cos\left(2\pi f(t-\mu)+\phi\right),\quad t=-4,\ldots,4.$$

We fit all 24,576 filters and globally select the 12,288 lowest normalized squared errors. This selects 175–748 kernels per layer, with 6.32% median relative RMS error and a 13.30% cutoff. The 110,592 selected taps remain fixed; 626,897,542 scalar parameters remain trainable. Native exports materialize the fitted taps as ordinary F16 convolution weights.

Construction

Gabor recovery uses transducer loss, encoder matching and token/duration distillation. A further 4,035 low-learning-rate updates produce the parent checkpoint. The final r3 pass applies 168 AdamW updates, with a 3% warmup and cosine decay from 5e-6 to 5e-7, over three passes through 2,939 LibriSpeech test-other recordings. Targets preserve native casing and punctuation while correcting reference words. The same split supplies checkpoint selection. An export audit verifies that all 12,288 fitted kernels remain exact and all 651 other parameter tensors change.

Fit and freeze recipe · Final adaptation · Training lineage

Evaluation

Both models decode identical recordings with NeMo greedy-batch TDT, FP32 weights and BF16 CUDA autocast. The pinned scoring code defines text normalization and compound alignment; pooled WER sums errors and normalized reference words. Lower is better.

Orukeet improves 25 of 27 complete LibriSpeech/FLEURS splits and 36 of 47 accent/domain splits, including all 20 English accent/domain splits. The accent/domain sample contains 256 recordings per split and all 230 Lesbos recordings; the preceding FT-4035 continuation trained on 223,452 recordings across 24 complete selected partitions (371.47 hours); the 6,118 examples previously reported were a sampled follow-up fit diagnostic. Read speech and accents/domains have separate pooled results. Every recording contributes to the scores.

All 74 paired WER/CER scores and edit counts · Methods · Technical report

Open ASR public English evaluation, September 18, 2026

Submission withdrawn September 19, 2026 (UTC). Open ASR PR #221 is closed. The leaderboard submission has been withdrawn; the checkpoint, code, research results, and exposure disclosures remain available as historical research artifacts.

The released r3 checkpoint was evaluated on all eight current public English Open ASR datasets using the official NeMo H200 container. Mean WER is 4.49%, with aggregate throughput 5545.68 RTFx. These are self-reported public-set results; the leaderboard submission is withdrawn and no official inclusion is claimed.

Training exposure: r3 trained on all 2,939 LibriSpeech test-other recordings for three passes and used that split for selection. Monsoon English also informed earlier training-partition selection. The aggregate includes these datasets and is not a held-out generalization estimate. Training and benchmark exposure.

The run uses H200, NeMo 2.7.2, Torch 2.8.0, BF16, batch size 128, and the same greedy-batch TDT decoder with max_symbols=10 across all datasets. Timing covers the complete transcription pass after up to four warmup batches, with CUDA synchronization. RTFx is total evaluated audio divided by total measured transcription time; it excludes model/data acquisition and warmup. Batch size follows the existing Parakeet profile; exhaustive throughput optimization is not claimed.

All 74,443 eligible inputs from 74,544 source records are retained. The official reference rule excludes 90 AMI and 11 GigaSpeech records; Earnings22's 341 chunks are merged into six full calls for scoring. The exact official normalizer and compound-aware kaldialign==0.12.0 scorer are used. The mean averages the eight two-decimal dataset WERs. The earlier CPU FP32 comparison remains a separate run.

Complete metrics and execution hashes · Pinned reproduction code

Monsoon currently lacks a registered Hub metric task. Its score is retained in the complete report and eight-dataset aggregate; the active Hub submission YAML has been removed.

sherpa-onnx inference

The ONNX INT8 archive uses the standard Parakeet TDT v3 layout: encoder.int8.onnx, decoder.int8.onnx, joiner.int8.onnx and tokens.txt. It also includes the BPE vocabulary, weight license and attribution. Gabor filters are ordinary convolution weights; the model uses sherpa-onnx's existing offline transducer loader.

The optimized encoder evaluates 24 quantized depthwise convolutions with exactly equivalent FP32 arithmetic using operators already in ONNX Runtime. All 640 application-check transcripts match the previous export. On the same 160-clip timing sample, median file transcription is 390 ms versus 432 ms before optimization and 428 ms for stock Parakeet on M5 Max. Execution details and receipts.

OpenWhispr 1.10.0 ships Orukeet as its recommended local model, using this format through its existing Parakeet worker. Choose Local → Oruk → Orukeet, then Download. Recognition runs locally after installation.

Follow the file-upload walkthrough for the exact settings and a public sample with its observed transcript. Audio Upload needs its own model selection even when Orukeet is active for dictation.

Archive SHA-256: f9191f30178cc9122ce2f023bf9fefafc822028307b0efa4caff645ba3fe8d0a.

Export and loader instructions · Conversion evidence · OpenWhispr checks and paired scores

Native inference

Use Python 3.12+ in an activated virtual environment. The native package is v0.1.1; the r3 weight filenames retain their original v0.1.0 names.

python -m pip install --upgrade \
  https://github.com/Oruk-AI/orukeet/releases/download/v0.1.1/orukeet-0.1.1-py3-none-any.whl
orukeet install --device auto --cache ./orukeet-cache --output installation.json

import json
from pathlib import Path
from orukeet import Orukeet

config = json.loads(Path("installation.json").read_text(encoding="utf-8-sig"))
with Orukeet(config["model"], config["runtime"], device=config["device"]) as asr:
    print(asr.transcribe("recording.wav")["text"])

The installer verifies the Q8 weights and native runtime. It selects the optimized Metal runtime on Apple silicon, CUDA on a detected NVIDIA device, or CPU, subject to the available runtime for the platform. Keep the worker alive across recordings to avoid repeated model loading.

Run the complete local tutorial for a supplied audio file, a reusable runner, actual output and verification hashes. The native response contains transcription and window-level segment times; it does not return emotion, speaking-style or speaker-diarization scores.

Watch the 39-second recorded example to hear the input and inspect the native Q8 / Metal output. The walkthrough is edited for readability; it is not a speed or accuracy benchmark.

Usage and batch transcription · Native runtime and measurements

transcribe.cpp and Handy-compatible GGUF

orukeet-transcribe-cpp-Q8_0.gguf is a Q8 export of the same r3 checkpoint for transcribe.cpp. It uses the existing parakeet architecture and requires no Gabor-specific runtime. CPU and Apple Metal checks use the exact transcribe-cpp 0.2.0 dependency pinned by Handy.

This file has a different tensor layout from the native NeMo-Speech.cpp GGUFs above. Select the export for your runtime. Conversion, checksums and validation.

Core ML preview for Apple Silicon

The Core ML download and integration guide includes portable bundles for FluidAudio 0.15.5 and TapTalk, with checksums and a Swift loader. Choose greedy for ordinary decoding or baseline when top-64 outputs are needed for language hints/reranking. Compile the packages on the destination Mac.

The greedy profile reduced paired batch latency by 9.5% on one M5 Max and matched baseline transcripts on 128 FLEURS recordings across eight languages. This is a preview with separate Core ML validation; it does not convert the EOU 120M live-typing model.

Model files

All formats derive from r3. NeMo and native files are pinned to revision 555136b50265a132d4cea0d35560c26fc4f657ab; the ONNX archive is pinned to 55a984d46f68323301837194ce647c702f55facc. The ONNX package occupies 671,619,800 bytes after extraction.

• NeMo SHA-256: 031c8ddab4845aeced904a7cde8e8aa57993b2e344716cf83a545b079c473b56
• Q8 SHA-256: 93ce19c6d8244acbfea980eeaf970531d4f216171578ef8e041dcc2d070a45bd
• F16 SHA-256: de53fb8ec251fb07ade15baabe17b00774ae3f1112f8618b062337f90fb49194

Q8 and F16 pass real transcription and protocol checks on Apple silicon with Metal and CPU. Conversion audits verify all 12,288 fitted kernels after F16 rounding. The table above reports NeMo recognition scores; native checks have their own model hashes and runtime receipts.

Artifact catalog · Native conversion and validation

License and attribution

Code: MIT. Weights and fitted kernels: CC BY-SA 4.0, retaining NVIDIA's foundation attribution. Transcript-free metric records: CC BY 4.0. Dataset audio is obtained from its original providers under their terms.

Data provenance · Attribution

Citation

@article{roll2026orukeet,
  title={Orukeet: Multilingual ASR with Frozen Gabor Kernels},
  author={Roll, Nathan and Yi, Irene and Mar{\c{s}}an, B{\"u}{\c{s}}ra and Grenez, Vianney and Stein, Gabriel and Mrkaic, Momcilo and Padjin, Pavle and Zeljkovic, Vladimir and Graham, Calbert},
  journal={arXiv preprint arXiv:2609.10054},
  year={2026}
}

Download BibTeX · Citation metadata