AI BriefingParakeet-based ASR Model 'Orukeet' Released
Key point
Orukeet, a 25-language ASR model based on NVIDIA Parakeet TDT, has been released, improving average WER by 10.6%.
Details
A new speech recognition model, Orukeet, developed based on NVIDIA Parakeet TDT 0.6B v3, has been released. The model supports 25 languages, replacing half of the temporal depthwise filters in the encoder with 12,288 fixed Gabor kernels and training the remaining parameters on multilingual and multi-accent data.
Performance Improvements
Orukeet outperformed the existing Parakeet model on 61 of the 74 tested datasets. Key benchmark results are as follows:
- LibriSpeech test-clean: WER 1.46% (previous 1.53%)
- LibriSpeech test-other: WER 2.86% (previous 3.14%)
- FLEURS English: WER 3.82% (previous 4.28%)
- FLEURS 25-language average: Recorded 3.82%, a 10.6% relative reduction from 11.01%
The LibriSpeech test-other dataset was used for final adaptation and checkpoint selection. The model is available on Hugging Face.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.