Meta's new AI can transcribe 1,600 languages
Meta just dropped Omnilingual ASR, a speech-to-text AI that can handle over 1,600 languages—including more than 500 rare or previously ignored ones.
It's a big step toward making voice tech accessible for everyone, especially communities whose languages have been left out until now.
Major and minor languages covered
This tool isn't just about major languages like Hindi and Telugu—it also understands lesser-known ones like Maithili and Awadhi.
Meta teamed up with local communities and linguists to build a massive dataset covering 350 underserved languages.
How it works
Omnilingual ASR uses some serious AI muscle—a 7-billion parameter speech encoder plus two types of decoders—to keep transcription errors low (under 10% for most languages).
Want to add your language? All you need are a few audio-text samples—no huge data sets or coding skills required.
Open access for all
Meta is making the models and dataset freely available under open licenses.
That means researchers, developers, or anyone curious can use or improve the tech—helping even more people get access to accurate speech recognition in their own language.