Meta open-sources speech AI covering 1,600+ languages

Omnilingual ASR is Meta's new open-source speech recognition family designed to transcribe spoken audio across more than 1,600 languages right out of the box.

To put that in perspective, OpenAI's widely used Whisper model only supports 99 languages. Meta's technical release covers over 500 languages that have never been supported by any AI speech recognition system until now.

Expanding to 5,400 languages on the fly

The real technical leap is how the model handles unseen dialects. Borrowing in-context learning from large language models, Omnilingual ASR includes a zero-shot capability Meta calls Bring Your Own Language.

Instead of launching a massive fine-tuning pipeline, you simply supply a few audio clips paired with their written text at inference time. The model picks up the phonetic and lexical patterns immediately, allowing it to transcribe new utterances without retraining.

Meta estimates this mechanism expands practical speech-to-text capabilities to over 5,400 languages, covering virtually every spoken tongue with an established writing system.

Specs, performance, and true open source

The suite was trained on over 4.3 million hours of audio across 1,600 languages, built on top of FAIR's PyTorch-based fairseq2 framework.

Model sizes scale from a 300-million parameter version suited for low-power edge devices up to a 7-billion parameter powerhouse for high accuracy. The lineup spans wav2vec 2.0 self-supervised encoders, CTC-based models for rapid decoding, and hybrid LLM-ASR architectures.

In benchmark evaluations, Omnilingual ASR clocked a character error rate under 10 percent across 78 percent of the 1,600 evaluated languages. For languages backed by at least ten hours of training audio, 95 percent hit that quality bar. Even among ultra-low-resource languages with under ten hours of data, 36 percent stayed below that 10 percent error ceiling.

Crucially, Meta avoided its controversial, enterprise-restricted Llama license this time. Omnilingual ASR is released under a clean Apache 2.0 license, making it completely free for unrestricted enterprise and commercial implementation.

Meta also published the Omnilingual ASR Corpus covering 350 previously underserved languages under CC-BY. The code, weights, and technical paper are available on GitHub and Meta's research portal, alongside a test space hosted on Hugging Face.

Why it matters

If you build voice interfaces for emerging markets, run media localization pipelines, or preserve oral histories, this release is worth testing immediately. It turns thousands of underserved dialects into actionable digital text overnight.

We think shifting to a permissive Apache 2.0 license is Meta's smartest move here. The restrictive terms on past releases drove corporate builders away, but an unencumbered foundation model with lightweight in-context scaling has the potential to displace Whisper as the default open-source transcription standard. Keep in mind that recognition quality on minimally represented languages will still require real-world validation before you deploy it to critical workflows.