Meta AI Transcribes 500 Niche Languages for the First Time, Admits It's Not Perfect

Meta AI Transcribes 500 Niche Languages for the First Time, Admits It’s Not Perfect

November 11, 2025

Don’t miss out on our latest stories. Add PCMag as a preferred source on Google.

Meta’s newly formed Superintelligence Labs has debuted a suite of AI-powered transcription models, which it says is the first to transcribe 500 “low-resourced” languages.

An open-source transcription tool, dubbed Omnilingual Automatic Speech Recognition (ASR), is now available on Hugging Face. Meta says it can recognize over 1,600 languages, including Hwana (Nigeria), Rotokas (Papua New Guinea), and Güilá Zapotec (Mexico), according to a Meta research paper.

These “low-resourced” languages are not “well represented on the internet,” Meta says. “This means high-quality transcriptions are often unavailable for speakers of less widely represented or low-resource languages, furthering the digital divide.”

Meta hopes that advances in AI technology will make it possible to build a universal transcription service, though native speakers of those languages will be the true test of the technology’s effectiveness. It also remains to be seen if Meta can convince people in Papua New Guinea, or those who speak the indigenous Zapotec language in Mexico, to use its AI products.

The base model, Omnilingual wav2vec 2.0, is also now available. Meta trained it to understand many languages and speech patterns without needing a lot of specific examples for each one. Meta also released its “unique collection” of transcribed speech in 350 underserved languages, called the Omnilingual ASR Corpus.

If the user finds their language of choice is not represented, they can upload “only a handful” of audio snippets, plus accompanying text, and “obtain usable transcription quality,” Meta says. The company admits the performance is likely not yet as good as systems that are fully trained, but it might get the user at least most of the way there. Meta pitches this approach as a more scalable way to bring new languages into the digital fold.

“This is experimental software,” Meta says. “While we strive for accuracy, transcriptions are not perfect. You should always double-check the outputs and make edits accordingly to ensure accuracy for your specific use case.”

Recommended by Our Editors

Live Translation with Apple Intelligence on AirPods Pro 3

Live Translation Doesn’t Require AirPods Pro 3, But You Do Need This

WhatsApp Finally Adds Built-In Message Translation for These Languages

Thanks to Apple Live Translation, My Wife Can Speak Spanish With Her In-Laws

Google Translate added zero-shot machine translation in 2022, which allows it to translate texts without ever seeing an example. The company has a more in-depth process of working directly with native speakers and translators before adding new languages to the site, Isaac Caswell, a software engineer on Google Translate, told us last year. Meta says it also worked with “local organizations that recruited and compensated native speakers, often in remote or under-documented regions,” for languages with little to no digital presence.

Many tech companies have leveraged the recent surge in large language models for language translation. ChatGPT has been able to produce high-quality translations since its release, and Apple’s AirPods software can now translate speech in real-time on newer models.

Meta is also working on expanding its AI technologies globally through infrastructure. In February, it announced the world’s longest underwater cable, Project Waterworth, which aims to “facilitate digital inclusion” and bring AI to more regions around the world.