papersSEP 10 04:00 UTC
Orukeet paper proposes multilingual ASR using frozen Gabor kernels in Parakeet encoder
A new research paper introduces Orukeet, a speech recognition model that replaces half of an adapted Parakeet encoder's temporal filtering layers with 12,288 fitted Gabor kernels, which are then kept frozen. The remaining parameters are trained on multilingual and multi-accent speech data, followed by a final adaptation and checkpoint selection stage. The approach aims to extend strong transcription models to more languages and accents.