Omnilingual 1B — Meta's massive multilingual speech recognition model covering over 1,600 languages — is now available in EveryScribe's Private Transcriber. If you've ever needed to transcribe audio in a language that no other ASR tool supports, this is your model.
What Is Omnilingual?
Omnilingual is a speech foundation model developed by Meta AI. It was built to address one of the most fundamental inequities in speech technology: most ASR models support 50–100 languages at best, leaving the vast majority of the world's 7,000+ languages completely unserved.
Omnilingual's goal is different. It supports over 1,600 languages, including more than 500 languages that have never before been supported by any AI speech recognition system. This includes numerous indigenous languages, regional varieties, and minority languages across Africa, the Americas, the Pacific, and Asia.
What Does 1,600 Languages Actually Mean?
To put this in context:
- Whisper Large-v3: ~100 languages
- SenseVoice Large: 50+ languages
- Parakeet v3: 25 European languages
- Omnilingual 1B: 1,600+ languages
The coverage spans every major language family: Indo-European, Sino-Tibetan, Afro-Asiatic, Niger-Congo, Austronesian, Dravidian, Turkic, and many smaller families. For researchers, linguists, journalists, and development workers operating in under-resourced linguistic environments, this kind of coverage has simply not existed before.
Accuracy: Below 10% CER for 78% of Languages
Raw language count is only meaningful if accuracy is real. Meta's benchmarks show character error rates (CER) below 10% for 78% of the supported languages — a remarkably high bar given the coverage breadth.
Omnilingual was also designed with zero-shot generalization: using a small number of in-context audio examples, it can extend to languages not explicitly present in its training data. This is particularly valuable for languages with extremely limited digital resources.
984 MB, Runs in Your Browser
The 1B variant we ship is a quantized INT8 ONNX model at 984 MB — the largest standard model in our lineup. It runs via WebAssembly in your browser. After the one-time download, all transcription is local. No audio leaves your device, regardless of what language you're transcribing.
This matters for rare language communities in particular. Recordings of indigenous elders, community ceremonies, oral histories — these are among the most sensitive recordings people might need to transcribe. Having them processed locally, with no cloud dependency, is the right way to handle them.
When to Choose Omnilingual 1B
Omnilingual 1B is the model for you if:
- Your audio is in a rare, regional, or indigenous language not covered by other models
- You work with multilingual recordings spanning many different language families
- You're a researcher or linguist working with under-documented languages
- You need the broadest possible language coverage in a single model
- You can accommodate the 984 MB download size for the accuracy and coverage trade-off
For most common languages (Chinese, English, Japanese, Spanish, French, German, etc.), our other models offer higher accuracy at smaller sizes. Omnilingual's strength is its breadth, not its depth on any single language.
If the 984 MB size is a concern, see Omnilingual 300M — a lighter variant at 348 MB that still covers 1,600+ languages with slightly reduced accuracy.
Get Started
Visit everyscribe.com/dashboard/offline-transcriber, select Omnilingual 1B from the ASR model dropdown, and download it once. All transcription runs locally, covering over 1,600 languages from a single model in your browser.
Omnilingual was developed by Meta AI based on the Massively Multilingual Speech (MMS) project. The source code is on GitHub and model variants are available on Hugging Face. We're grateful to Meta for releasing this model under an open license that makes global language access possible without cloud infrastructure.
