Meta’s new Omnilingual ASR redefines open source speech recognition, offering native support for 1600 languages and extendable coverage to over 5000 more free for all under the Apache 2.0 license.
Meta has made a decisive return to open source with the launch of Omnilingual ASR, a multilingual automatic speech recognition (ASR) system that supports 1600+ languages, the largest coverage ever achieved in open source speech technology. Through zero-shot in-context learning, the model can extend recognition to more than 5400 languages, effectively covering nearly every spoken language with a known script.
The suite includes multiple ASR model families, wav2vec 2.0, CTC-based, LLM-ASR, and LLM-ZeroShot ASR, alongside a 7-billion parameter multilingual model and a vast corpus spanning 350+ underserved languages. Crucially, all resources are open-sourced under the permissive Apache 2.0 license, allowing free commercial and enterprise use.
“By open sourcing these models and dataset, we aim to break down language barriers, expand digital access, and empower communities worldwide,” Meta stated via its @AIatMeta handle on X.
Trained on 4.3 million hours of multilingual audio, Omnilingual ASR achieves character error rates below 10% in 78% of supported languages, including 500+ languages never before covered by any ASR model. Its zero-shot capability enables transcription of new or endangered languages with minimal data, making it both adaptable and inclusive.
The release marks Meta’s first major open-source model since the underwhelming Llama 4, representing a strategic shift under Chief AI Officer Alexandr Wang toward community-driven innovation. Built collaboratively with organisations such as Mozilla Common Voice, African Next Voices, and Lanfrica, Omnilingual ASR transforms ASR from a closed corporate service into an open, extensible framework for global linguistic inclusion and digital accessibility.
