@winbuzzer
“Unlike traditional ASR systems that require extensive fine-tuning, Omnilingual ASR introduces a novel “Bring Your Own Language” feature. This capability, inspired by large language models, allows users to add support for entirely new languages by providing just a handful of paired audio and text samples.”
Magical thinking, so I'm suspecting hype. A handful of paired samples is not enough data to correctly work out the phonological system of any language. (I used to teach how to do that, with a new language every year.) So I assume it's just transcribing into some generic alphabet (which is certainly possible), without regard to the actual pattern of the language in question. Like using Italian spelling to write down !Xhosa?
Their contribution to data on under-represented languages is welcome, though.
