Meta’s Omnilingual Automatic Speech Recognition system has landed on TIME’s list of the Best Inventions of 2026. The recognition came on October 15, 2026, almost a year after Meta’s research team first released the tool to the public.
The pitch is simple: speech-to-text for more than 1,600 languages. That includes approximately 500 low-resource languages that, until this project, no automatic speech recognition system had touched.
What Meta actually built
Automatic speech recognition, or ASR, is the technology that listens to spoken words and turns them into written text. It powers dictation on your phone, captions on videos, and the transcripts that pile up after every work call.
The world has roughly 7,000 languages. Meta’s Fundamental AI Research team, known as FAIR, set out to close a large slice of that gap and released Omnilingual ASR as open source on November 10, 2025.
The system supports over 1,600 languages. More than 500 of those had never been supported by any ASR solution before.
On accuracy, Meta reports a character error rate below 10% for 78% of the supported languages. Character error rate measures how many individual letters the system gets wrong. A score under 10% means fewer than one character in ten needs fixing.
The technical choices that matter
Meta did not ship a single model. It released a family of models ranging from 300 million to 7 billion parameters.
Alongside the models, Meta published the Omnilingual ASR Corpus. This dataset covers 350 underserved languages and gives other researchers raw material to build on.
The system can adapt to a new language using only a small number of audio samples, an approach known as zero-shot adaptation.
Open source, with real-world users
Meta released Omnilingual ASR under the Apache 2.0 license, allowing both commercial and research use.
The project was developed with global partners, and Meta emphasized open data collection and licensing throughout.
One of the first real-world uses comes from Nigeria. Health practitioners there have deployed the system to improve patient care.
Background: the long tail of language tech
Previous ASR systems, including OpenAI’s Whisper, have only been capable of transcribing approximately 99 languages, leaving much of the world’s linguistic diversity without adequate voice technology support.
What this means
For developers and researchers, the Apache 2.0 license is the headline feature. Open weights, a public corpus, and models sized from 300 million to 7 billion parameters mean the barrier to building local-language products is lower than it has ever been for many of these communities.
There are limits worth watching. A character error rate below 10% for 78% of languages still leaves 22% of supported languages above that line.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.