Meta releases open-source speech recognition for more than 1,600 languages

Meta released Omnilingual ASR, a suite of open-source automatic speech recognition models that transcribe spoken language in more than 1,600 languages, including over 500 that the company says had never before been served by any ASR system. The models and tools were published in November 2025 and made freely available through a public code repository, lowering the barrier for researchers and communities to build speech technology for languages the commercial market has long ignored.

For comparison, widely used systems typically support roughly 100 to 125 languages. Omnilingual ASR's reach is far broader, and the team built it to be extensible: through a technique called zero-shot in-context learning, the system can be pointed at a new, unseen language using only a few paired audio and text examples, without retraining the full model. The suite ranges from compact models for low-power devices to a 7-billion-parameter model for higher accuracy.

“The models and tools were published in November 2025 and made freely available through a public code repository, lowering the barrier for researchers and communities to build speech technology for languages the commercial market has long ignored.”

Crucially, the underserved languages were not scraped from the web. Meta gathered community-sourced recordings through compensated local partnerships, working with organizations and speakers in regions including parts of Africa and South Asia, and released a corpus of speech for hundreds of low-resource languages alongside the models. That approach treats language communities as collaborators rather than data sources.

Caveats remain. Accuracy varies widely by language and depends heavily on how much quality data exists, so performance on the rarest languages will be weaker than on well-resourced ones. Speech recognition also raises privacy and consent questions that responsible deployment must address. Even so, open-sourcing tools that work across 1,600 languages is a meaningful step toward digital inclusion, giving speakers of marginalized languages a path to voice search, transcription, education tools and accessibility features that English speakers have enjoyed for years.