Meta Omnilingual Automatic Speech Recognition (ASR) Officially Introduced: Offers ASR Capabilities for 1,600+ Languages

By Sidharth Joseph Published on November 11, 2025, 10:20 IST

Meta has now officially introduced its Omnilingual Automatic Speech Recognition (ASR) system, bringing in ASR capabilities for more than 1600 languages. Furthermore, it also includes 500 low-coverage languages too that have never been covered previously by any ASR systems.

Check out more about it below.

Meta Omnilingual Automatic Speech Recognition (ASR) – Officially Introduced

The Meta Omnilingual Automatic Speech Recognition (ASR) system has been introduced as a suite of AI models which is capable of providing automatic speech recognition for 1,600+ languages and as noted above, these include 500 low-coverage/resource languages too. In comparison to other ASR systems, most of them are focused on serving a limited number of languages that are already well-represented on the internet, whereas Meta’s Omnilingual Automatic Speech Recognition (ASR) system by including low-coverage/resource languages has taken a major step in achieving and offering a truly universal transcription system.

Speaking more, as the Omnilingual ASR system of Meta has adopted a community-driven framework, users and people across the world will be able to increase its language capabilities by just adding a few samples from their own languages. Along with the Omnilingual ASR, the brand also released the Omnilingual ASR Corpus and the Omnilingual wav2vec 2.0.

The latter is a multilingual speech recognition model that has been scaled up to 7B parameter, and claims character error rates (CER) of less than 10% in 78% of the covered languages. Meta’s Omnilingual ASR Corpus is a collection of transcribed speeches from 350 low-covered languages from around the world. Working with local organizations as well as native speakers, combining the efforts of linguists, researchers, and language communities via the Language Technology Partner Program, and also joining with Mozilla foundation’s Common Voice and Lanfrica (or NaijaVoices), Meta was able to form this collection of transcribed speech dataset.

To sum up, a full suite of AI models and a dataset are part of this release, and all these are built over the previous research done by the Meta Fundamental AI Research (FAIR) team. As noted by Meta, the main goal of all these is to bring the world together – more closer together.

Source X (@AIatMeta)