Meta's new Omnilingual AI can understand 1,600+ languages: How it works

Meta's new Omnilingual ASR can transcribe speech in over 1,600 languages, including several regional Indian ones such as Awadhi, Maithili, Chhattisgarhi, and Tulu

Meta has introduced Omnilingual ASR (Automatic Speech Recognition), a new AI framework designed to transcribe speech into text across more than 1,600 languages. The system, developed by Meta’s Fundamental AI Research (FAIR) team, includes support for around 500 lesser-known or low-resource languages, some of which are being recognised by AI systems for the first time.

 

The company has also released the Omnilingual ASR Corpus, a dataset of transcribed audio in 350 “underserved” languages, along with a new multilingual speech model called Omnilingual wav2vec 2.0, which powers the system.

What is Meta’s Omnilingual ASR?

Omnilingual ASR is Meta’s new suite of speech recognition models aimed at understanding spoken language and converting it into text. It expands on existing voice transcription systems that usually cover a few dozen major languages by extending coverage to over 1,600.

 

 

For users, this means that speech-to-text tools, translation systems, and voice-enabled AI assistants could soon work with languages that are currently unsupported — including many regional Indian languages such as Awadhi, Maithili, Chhattisgarhi, and Tulu.

 

  Meta said that the system can also learn new languages without requiring massive training datasets. Instead, speakers can provide a small number of paired audio and text samples to help the system adapt. This could allow communities with limited digital resources to make their languages usable in speech technologies.

How is it different from existing speech recognition systems?

Traditional speech recognition systems, such as those used in popular assistants or transcription tools, depend on large amounts of labelled audio data. This has limited high-quality transcription to a small number of widely spoken languages.

 

Omnilingual ASR takes a different approach. It builds on Meta’s wav2vec 2.0 speech representation model, scaled up to 7 billion parameters, and uses AI methods similar to large language models (LLMs). This allows it to generate transcriptions with fewer examples and adapt more easily to new languages.

 

Another key difference is its flexibility. The Omnilingual ASR framework supports both lightweight models for low-power devices and larger, high-performance versions for more complex applications.

When could it appear in consumer products?

At present, Omnilingual ASR is available as an open research model under the Apache 2.0 license, meaning developers and researchers can use or modify it freely.

 

While Meta has not confirmed when this technology will appear in its consumer products, the ASR system could eventually support voice features across Meta’s platforms, including WhatsApp voice message transcription, Instagram Reels captions, or even Meta AI assistant.