Preprint

Omnilingual ASR: Open-Source Multilingual Speech Recognition for 1600+ Languages

Omnilingual ASR team, arXiv, 2025

Read on October 17, 2026

What it is: Meta's full technical report on Omnilingual ASR, a speech recognition model family covering more than 1,600 languages. It reports character error rates by language group and by how much training audio each language has.
How much weight it carries: Primary technical report from the developers, with per-group tables. It is a company's own evaluation on its own test sets, so the figures show what the authors measured rather than independent results.

Facts taken from this source

Meta's Omnilingual ASR paper reports that its 7B model meets the character error rate threshold for 78% of the 1,570 languages it was evaluated on.
“overall our model meets the CER quality threshold for 78% of the 1570 languages we evaluate on” (Section 5.3)
In Omnilingual ASR, only 36 percent of the languages with less than 10 hours of training audio meet the character error rate threshold of 10 percent.
“On the low-resource bucket, where we have less than 10 hours of training data per language, the percentages of languages that meet the CER quality threshold fall to 34% and 36%” (Section 5.3.1)