aiqbench analysis
This research introduces Orukeet, an automatic speech recognition model that integrates frozen Gabor kernels into a Parakeet encoder to significantly improve multilingual accuracy. By replacing half of the temporal filters with these fitted kernels, the approach achieves a 10.6% relative reduction in word error rate across 25 languages while maintaining standard inference operators. The study demonstrates that architectural modifications focused on specific frequency-domain features can yield substantial performance gains without altering the underlying model's deployment requirements.
Orukeet: Multilingual ASR with Frozen Gabor Kernels
arXiv:2609.10054v1 Announce Type: cross Abstract: Orukeet replaces half of an adapted Parakeet encoder's temporal filters with 12,288 fitted Gabor kernels, freezes these replacements, and trains the remaining parameters on multilingual and multi-accent data. Final adaptation and checkpoint selection use LibriSpeech test-other. Across 20,146 FLEURS recordings in 25 languages, pooled word error rate (WER) falls from Parakeet's 11.01% to Orukeet's 9.85%, a 10.6% relative reduction. Orukeet has lower WER on 23 of the 25 languages. Orukeet outperforms Parakeet on 61 out 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%). All comparisons decode the same audio with matched NeMo settings. The fitted kernels are stored as ordinary convolution weights, retaining Parakeet's architecture and inference operators.
Abstract:Orukeet replaces half of an adapted Parakeet encoder's temporal filters with 12,288 fitted Gabor kernels, freezes these replacements, and trains the remaining parameters on multilingual and multi-accent data. Final adaptation and checkpoint selection use LibriSpeech test-other. Across 20,146 FLEURS recordings in 25 languages, pooled word error rate (WER) falls from Parakeet's 11.01% to Orukeet's 9.85%, a 10.6% relative reduction. Orukeet has lower WER on 23 of the 25 languages. Orukeet outperforms Parakeet on 61 out 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%). All comparisons decode the same audio with matched NeMo settings. The fitted kernels are stored as ordinary convolution weights, retaining Parakeet's architecture and inference operators.
5 pages, 2 figures. Code and model: this https URL
Sound (cs.SD); Machine Learning (cs.LG); Audio and Speech Processing (eess.AS)
arXiv:2609.10054 [cs.SD]
(or arXiv:2609.10054v1 [cs.SD] for this version)
https://doi.org/10.48550/arXiv.2609.10054 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 232 words
Related coverage
The Cross-Architecture Substrate: A Domain-Transcendent, Calibration-Surviving Geometric Invariant of Modern Vision Encoders
arXiv · Jun 9, 2026
Four-Qubit Kernel Preserves Geometry on IBM Quantum Hardware to 98.9%
Quantum Zeitgeist · Aug 5, 2026
Representation Learning with Quantum Signal Processing
arXiv · Sep 1, 2026
MaRK: Markov-adapted Recurrent Kernels for Dynamic Operator Conditioning in State Space Models
arXiv · Oct 8, 2026