What is Omnilingual ASR?
Omnilingual ASR is a family of open speech recognition models from Meta's FAIR research group, released in November 2025, that transcribes more than 1,600 languages, including over 500 that Meta says no speech recognition system covered before. It is Meta's current multilingual speech-to-text work and follows its 2023 Massively Multilingual Speech (MMS) models, which covered about 1,100 languages.
All the models share a wav2vec 2.0 speech encoder that Meta scaled to 7 billion parameters and pre-trained on 4.3 million hours of unlabelled audio. Two decoders sit on top: a fast CTC decoder, and an LLM-ASR decoder in the style of a language model, which can be told the language and script to expect. Each comes at 300M, 1B, 3B and 7B encoder sizes. A separate 7B zero-shot model transcribes a language it was never trained on from a handful of example recordings with their transcripts.
In December 2025 Meta released v2 checkpoints with lower error rates and Unlimited variants of LLM-ASR that handle audio of any length; the other models accept clips under 40 seconds. Code and models are under Apache 2.0, and the Omnilingual ASR Corpus of transcribed speech in 350 under-served languages is under CC-BY 4.0. The repository has had no commits since 30 December 2025.
Is Omnilingual ASR worth it?
Omnilingual ASR is the widest-reaching open speech recognition system, and it is not close. Meta's 7B LLM-ASR model reaches a character error rate below 10 for 78% of the more than 1,600 languages it supports. On the 81 FLEURS languages that Whisper large-v3 also covers, it averages a 5.6 character error rate on the test set against Whisper's 22.6, and wins on 65 of the 81 languages. All figures come from Meta's own paper.
The headline average flatters it on the languages most buyers care about first. Whisper's average is dragged down by languages it barely supports; on MLS, eight high-resource European languages, Whisper large-v3 scores 2.0 against 2.4 for Omnilingual's best model, and among the world's most spoken languages the 7B model wins 24 of 34. Meta reports character error rate, not the word error rate most commercial vendors quote, so its numbers cannot be laid next to a vendor's spec sheet.
The long tail is where it earns its place, and where it is weakest. For 546 languages with under 10 hours of training audio, the 7B model's mean character error rate is 18.0 and only 36% of them fall below 10. The zero-shot model offers a route for a language not on the list, from a few paired recordings, but Meta says it does not yet match a fully trained model. Against Google's USM, which is not public, the 7B model scores 6.2 on FLEURS-102 to USM's best 6.5.
The engineering is research-grade. The models run on Meta's fairseq2 toolkit rather than Hugging Face Transformers, standard checkpoints refuse audio over 40 seconds, and the documentation describes no word timestamps; a request for confidence scores is among 38 open issues. The repository has been quiet since late December 2025, so fixes are slow to arrive. A team adding a language like Ligurian or Faroese gets a strong base under a licence with no strings; a team transcribing English sales calls gets nothing it lacked before.
Where it earns its place
A data team at an NGO running radio programmes in East and West African languages
Make hundreds of hours of broadcasts in local languages searchable
Run the 7B LLM-ASR Unlimited model on an A100-class GPU with the language code set for each programme. At Meta's real-time factor of 0.097 for long audio, 1,000 hours takes roughly 97 GPU-hours at batch size 1; check error rates per language in Meta's published table first, since low-resource languages vary widely.
A computational linguist documenting an endangered language
Get a first transcription model for a language with no speech recognition at all
Feed the 7B zero-shot model a handful of recordings with their transcripts as context, then fine-tune a CTC or LLM-ASR checkpoint with Meta's recipes once more data exists. Zero-shot quality is below a trained model, so treat its output as a draft for correction.
An engineer building an offline voice feature for a rugged field device
Transcribe short commands in a regional language without a network connection
The 300M CTC model needs about 2 GiB of memory and processes audio at a real-time factor of 0.001 on an A100, so smaller hardware has headroom. CTC models can pick the wrong script in low-resource languages, which the language-conditioned LLM-ASR models largely avoid.
Before you commit
- Setup
- pip install omnilingual-asr (version 0.2.0) plus the libsndfile system library, then three lines of Python with ASRInferencePipeline; checkpoints download to the fairseq2 cache on first use. Language codes follow the eng_Latn style.
- Integrations
- Built on fairseq2 and PyTorch; Meta provides data-preparation and fine-tuning recipes and a Hugging Face demo Space. There is no Transformers integration and no hosted API from Meta.
- Data and privacy
- Self-hosted, so audio never leaves your hardware. Apache 2.0 places no commercial or field-of-use limits on the models; the corpus is CC-BY 4.0 and needs attribution.
- Cost at scale
- On an A100 at batch size 1, the 7B LLM-ASR uses about 17 GiB of VRAM at a real-time factor of 0.092, about 5.5 GPU-minutes per hour of audio; the 300M CTC uses about 2 GiB at 0.001, a few seconds per hour. By Meta's BF16 figures a 24 GB GPU runs every model, the 7B zero-shot at about 20 GiB included.
- Limits
- Standard CTC and LLM-ASR models accept clips under 40 seconds; the Unlimited variants do not yet support fine-tuning. Meta reports character error rates only, from its own evaluations, and documents no timestamps or confidence scores.
What’s good
- More than 1,600 languages, over 500 never covered by speech recognition before
- Apache 2.0 for code and weights; commercial use allowed
- 7B LLM-ASR: CER under 10 for 78% of supported languages (Meta)
- 300M CTC needs about 2 GiB of VRAM and runs 96 times faster than the 7B LLM-ASR
What’s not
- Whisper large-v3 still wins on high-resource European languages (MLS 2.0 to 2.4)
- Low-resource languages average 18.0 CER; only 36% fall below 10
- Standard models cap clips at 40 seconds; no documented word timestamps
- fairseq2 stack, no Transformers support; no commits since December 2025
Best for
Teams transcribing languages that commercial speech APIs do not cover, who can run their own GPUs and accept research-grade tooling.
Omnilingual ASR alternatives
Omnilingual ASR: common questions
- Is Omnilingual ASR worth it?
- No other open model transcribes 1,600 languages under Apache 2.0, and for the long tail Omnilingual ASR is the default; for English and the big European languages Whisper still holds its ground. We rate it 4.2 out of 5.
- How much does Omnilingual ASR cost?
- Free under Apache 2.0, commercial use included; the training corpus is CC-BY 4.0. Meta offers no hosted API, so the cost is your own GPU time.
- What are the downsides of Omnilingual ASR?
- Whisper large-v3 still wins on high-resource European languages (MLS 2.0 to 2.4); low-resource languages average 18.0 CER; only 36% fall below 10; standard models cap clips at 40 seconds; no documented word timestamps; fairseq2 stack, no Transformers support; no commits since December 2025.
- Who is Omnilingual ASR best for?
- Teams transcribing languages that commercial speech APIs do not cover, who can run their own GPUs and accept research-grade tooling.
- How much does Omnilingual ASR cost for a team or in production?
- On an A100 at batch size 1, the 7B LLM-ASR uses about 17 GiB of VRAM at a real-time factor of 0.092, about 5.5 GPU-minutes per hour of audio; the 300M CTC uses about 2 GiB at 0.001, a few seconds per hour. By Meta's BF16 figures a 24 GB GPU runs every model, the 7B zero-shot at about 20 GiB included.
- Is Omnilingual ASR safe for company data?
- Self-hosted, so audio never leaves your hardware. Apache 2.0 places no commercial or field-of-use limits on the models; the corpus is CC-BY 4.0 and needs attribution.
- What are the best Omnilingual ASR alternatives?
- The closest alternatives we have reviewed are Whisper large-v3-turbo, Qwen3-ASR (1.7B and 0.6B), Parakeet TDT 0.6B v3 and Voxtral 3 (Transcribe and TTS).
Answers checked Nov 9, 2026.
Make Omnilingual ASR? Show this rating on your site
