Meta’s 1,600-Language ASR Joins 2026 Best Inventions List

Meta’s Omnilingual Automatic Speech Recognition (ASR) was named to a 2026 Best Inventions list on Oct. 15, drawing attention to a speech-to-text system designed to serve languages often unsupported by current recognition tools.

Of the 1,600-plus languages supported, 500 are low-resource languages with no prior AI transcription. More than 7,000 languages are spoken worldwide, and most lack support in existing speech-recognition systems.

Released Nov. 10, 2025, the suite includes models ranging from 300 million to 7 billion parameters. Its largest model had character error rates below 10% for 78% of supported languages. The accompanying corpus contains transcribed speech in 350 underserved languages.

The code and models use the Apache 2.0 license, while the speech corpus uses a CC-BY license. The system can be extended to new languages with a handful of paired audio and text samples, though performance with this approach has not matched that of fully trained systems.

Omnilingual ASR can also be paired with Meta’s separate Omnilingual Machine Translation system, which translates text and supports more than 1,600 languages. “We’re very passionate about making sure technology supports as many languages as possible,” said Jean Maillard, a lead researcher on the project.