Orukeet: Multilingual ASR with Frozen Gabor Kernels
Nathan Roll, Irene Yi, Büşra Marşan, Vianney Grenez, Gabriel Stein, Momcilo Mrkaic, Pavle Padjin, Vladimir Zeljkovic, et al.
cs.SD · cs.LG · eess.AS
Quick verdict
Claims freezing fitted Gabor kernels cuts Parakeet's FLEURS WER 11.01%→9.85%; gains largely plausible but test-set tuning and train-eval overlap weaken it.
Summary
Orukeet adapts NVIDIA's Parakeet TDT 0.6B v3, a 25-language ASR model, by fitting a Gabor function to each of the encoder's 24,576 nine-tap temporal kernels, replacing and freezing the 12,288 best-fitting ones, and retraining everything else. Replaced kernels are stored as ordinary convolution weights, so architecture and inference cost are unchanged. After recovery training, 4,035 multilingual adaptation updates, and a final 168-update pass on LibriSpeech test-other — with checkpoint selection also on test-other — pooled FLEURS WER falls from 11.01% to 9.85% (20,146 recordings, 25 languages), winning on 23 of 25 languages, and LibriSpeech test-clean/test-other improve to 1.46%/2.86%. On a 12,006-recording accent/domain sample spanning 47 partitions, pooled WER falls from 16.72% to 15.25%. Key limitations: the final pass trains directly on test-other, making those gains in-sample; 6,118 of the 12,006 accent recordings were part of the adaptation data; Greek regresses sharply on FLEURS (21.07% to 30.81%) without analysis; no significance testing is reported; and the adaptation corpus composition is not enumerated. The paper ships checkpoint hashes, MIT code, and per-record metric records, so the FLEURS result is independently checkable.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 3
● low 0
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 13 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Replacing and freezing the 12,288 best-Gabor-fitting temporal kernels of an adapted Parakeet TDT 0.6B v3 and retraining the rest reduces pooled FLEURS WER from 11.01% to 9.85% across 25 languages, improving 23 of them.
- The fitted kernels are stored as ordinary convolution weights, so architecture, operators and inference cost are identical to Parakeet.
- Orukeet beats stock Parakeet on 61 of 74 splits, including LibriSpeech test-clean (1.46% vs 1.53%), test-other (2.86% vs 3.14%), and a 47-partition accent/domain sample (pooled 15.25% vs 16.72%).
- All comparisons decode identical audio under matched NeMo settings, and every reported value is reproducible from stored integer edit counts and a SHA-256-identified checkpoint.
Strengths
Exhaustive, matched-decode evaluation
Both models are restored independently and decode identical 16 kHz audio in the same duration-sorted batches under pinned NeMo settings (§3); all 27 read-speech partitions (Table 1) and all 47 accent/domain partitions (Table 2) are enumerated, with pooled WER defined precisely (Eq. 2) and language-macro rows included.
Zero inference-cost change by construction, with an audit
Fitted Gabor taps are stored as ordinary depthwise convolution weights with identical tensor shapes and operator counts (§4), and an independent export audit verifies the 12,288 frozen kernels are byte-identical while tokenizer and BN statistics stay unchanged — a genuine drop-in claim, not an approximation.
Strong breadth of improvement across languages and accents
Low Gabor fit error supports the design rationale
Unusual transparency of artifacts
Weaknesses
Final pass trains on and selects by LibriSpeech test-other
Half the accent/domain evaluation was seen in training
Unexplained large regression on Greek
No inferential statistics
Adaptation corpus composition not enumerated
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This five-page preprint (Oruk AI / Stanford / Cambridge / OpenWhispr / Hoid) introduces Orukeet, a 25-language ASR model derived from NVIDIA's Parakeet TDT 0.6B v3. The central claim: replacing half of the encoder's temporal depthwise kernels with frozen, analytically fitted Gabor functions and retraining the remainder yields better multilingual accuracy at identical inference cost — pooled FLEURS WER falls from 11.01% to 9.85% (10.6% relative, 20,146 recordings, 25 languages), with wins on 23/25 languages and 61/74 total splits. Secondary claims cover LibriSpeech (test-clean 1.46% vs 1.53%; test-other 2.86% vs 3.14%) and a 47-partition accent/domain sample (pooled 16.72% → 15.25%). The claims are scoped honestly and the protocol is unusually well documented, but two evaluation-design choices — training and checkpoint selection on LibriSpeech test-other, and 6,118 evaluation recordings inside the adaptation corpus — materially qualify how the LibriSpeech and accent numbers may be read.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Orukeet?
Orukeet is a 25-language speech recognition model built from NVIDIA's Parakeet TDT 0.6B v3: half of the encoder's temporal convolution kernels (12,288 of 24,576) are replaced by frozen, mathematically fitted Gabor functions, and the remaining parameters are retrained on multilingual and multi-accent data. It runs at the same speed and architecture as Parakeet.
How much better is Orukeet than Parakeet?
On FLEURS (20,146 recordings, 25 languages), pooled word error rate drops from 11.01% to 9.85% — a 10.6% relative reduction — with Orukeet better on 23 of 25 languages. On LibriSpeech it reports 1.46% vs 1.53% (test-clean) and 2.86% vs 3.14% (test-other), but the final training pass used test-other, so those specific numbers reflect tuning on that set.
Is the Orukeet code and model publicly available?
Yes. The paper states code is MIT-licensed, weights and fitted functions are CC BY-SA 4.0, and evaluation records are CC BY 4.0, released via the Orukeet repository together with a NeMo checkpoint identified by SHA-256 prefix 031c8ddab484.
Does replacing kernels with Gabor functions slow down inference?
No. The fitted Gabor kernels are stored as ordinary depthwise convolution weights with the same tensor shapes and operator counts as Parakeet, so inference is unchanged — the paper also reports an export audit confirming the frozen kernels are byte-identical to the fitted functions.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Text-only extraction: Figure 1/2 plots cannot be visually inspected, though their captions and key annotations (fit ranks, RMS error percentages, per-layer counts) are present in the text; no section of the paper text appears truncated.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.