Meta releases Omnilingual ASR: a speech recognition system covering 1,600+ languages

Meta’s AI team has introduced Omnilingual ASR, the world’s first automatic speech recognition system supporting more than 1,600 languages, including 500 languages receiving AI speech recognition support for the first time. Powered by large-scale pretraining and in-context learning techniques, the system can adapt to new languages with only a small amount of audio–text data. It achieves under 10% character error rate for 78% of the supported languages.