Meta Open-Sources Omnilingual ASR: State-of-the-Art Speech Recognition for 1,600+ Languages

Meta’s FAIR team has unveiled Omnilingual ASR, a suite of models claiming to provide speech recognition for over 1,600 languages and setting a new benchmark for scale and quality. The release includes an Omnilingual ASR Model Family available in configurations from a 300-million-parameter, low-power variant to a 7-billion-parameter version delivering state-of-the-art accuracy, and an Omnilingual wav2vec 2.0 Base Model expanded to 7 billion parameters to serve as a foundation for ASR and other tasks.

The Omnilingual ASR Corpus, released under CC-BY, comprises transcribed speech from 350 under-served languages and accompanies public access to related models, datasets, demos, and tools via GitHub, Hugging Face and the Meta AI website. According to Meta’s published research, the 7-billion-parameter LLM-ASR system achieved state-of-the-art performance across more than 1,600 languages, with 78% of them showing a Character Error Rate below 10%.

The project also introduces a Bring Your Own Language paradigm, enabling rapid adaptation to new languages with minimal data, and it partners with the Mozilla Foundation, Lanfrica and NaijaVoices to gather speech samples.