Omnilingual ASR

Omnilingual ASR (Omnilingual Automatic Speech Recognition) is an open-source family of speech recognition models from Meta's Fundamental AI Research team (FAIR) that transcribes speech in more than 1,600 languages. Meta announced it on 10 November 2025 and says more than 500 of those languages had never been served by any ASR system before.^[1][2] The release combined a scaled-up wav2vec 2.0 speech encoder of up to 7 billion parameters, two families of ASR decoders (a CTC head and an LLM-style Transformer decoder), a zero-shot model that can be pointed at an unsupported language using a few paired audio and text examples, and the Omnilingual ASR Corpus, a CC-BY 4.0 dataset of spontaneous speech in several hundred under-served languages.^[1][3][4] The code and models are released under the Apache 2.0 license.^[1][3] In October 2026 TIME named Omnilingual ASR one of its Best Inventions of 2026.^[5][6]

Background

The system is a successor to Meta's Massively Multilingual Speech (MMS) project. The Omnilingual ASR paper describes MMS as a 1B-parameter wav2vec 2.0 encoder pre-trained on around 500,000 hours of unlabeled speech covering about 1,400 languages, then fine-tuned with CTC on around 45,000 hours of labeled data to cover roughly 1,100 languages.^[2] The authors state that, when choosing which languages to target, they deprioritized languages already covered by that earlier work, "on which Omnilingual ASR builds."^[2]

The paper frames the project around the long tail of the world's roughly 7,000 languages. It argues that existing ASR systems treat language coverage as fixed at release, so a language left out of training can only be added through expert-driven fine-tuning that most communities cannot carry out. It also discusses the ethical concerns that have kept researchers away from low-resource languages, including questions of language ownership, extractive data collection and compensation.^[2] Omnilingual ASR's response was to commission paid recordings through local partners and to build a model that can be extended with a handful of examples at inference time.^[1][2]

Release history

Architecture

All Omnilingual ASR models use an encoder-decoder design. A large Transformer speech encoder extracts cross-lingual representations from audio, and a decoder (either a single linear layer or a Transformer decoder) maps them to character tokens.^[2]

Speech encoder (Omnilingual wav2vec 2.0)

The encoder is trained with wav2vec 2.0 self-supervised learning: spans of the latent audio representation are masked, and the model learns to pick the correct quantized latent for each masked step from a set of distractors.^[2] Meta scaled the method from the original 300M-parameter model to 1B, 3B and 7B parameters. The paper notes that the largest previously reported speech self-supervised models, Google's Universal Speech Model (USM) and Meta's XLS-R, were each about 2B parameters.^[2] The encoders were pre-trained with the fairseq2 framework on 4.3 million hours of speech; the paper's data section breaks this down as 3.84 million hours across 1,239 identified languages plus another 460,000 hours with no language identification.^[2] Meta released the encoders separately as Omnilingual wav2vec 2.0 for use in other speech tasks.^[1]

Source: Table 4 of the paper. The repository README lists the 3B encoder checkpoint at 3,064,124,672 parameters, slightly more than the paper's 3,046M.^[2][3]

CTC and LLM-ASR decoders

The CTC models add a single linear layer on top of the encoder and fine-tune the whole network with the connectionist temporal classification loss. The second family, which Meta calls LLM-ASR, places an autoregressive Transformer decoder of the kind used in large language models on top of the encoder and trains it with next-token prediction.^[2] Every LLM-ASR variant uses the same decoder: 12 layers, model dimension 4,096 and eight attention heads, about 1.2B parameters. Variants are named by encoder size.^[2] Both families output characters from a single tokenizer built from the union of characters across the training data, which after cleaning contained 9,812 symbols.^[2]

LLM-ASR models can optionally be conditioned on a language and script code (for example eng_Latn or cmn_Hans). The authors added this because some languages, such as Urdu, can be written in more than one script and closely related languages can confuse the model. The language input is randomly dropped during training so that the model still works without it.^[2][3] The paper reports that CTC models often fail by predicting the wrong script, and that language conditioning in LLM-ASR "largely resolves" this problem.^[2]

Zero-shot "bring your own language"

The zero-shot model extends LLM-ASR with in-context learning. During training the decoder is shown N speech-text pairs from a language as context, followed by a target utterance whose transcript it must predict. At inference time a user supplies a few recorded utterances with transcripts in a language the model has never seen, and the model transcribes new audio in that language without retraining.^[2] Meta's blog calls this "Bring Your Own Language" and acknowledges that "zero-shot performance cannot yet match that of fully trained systems."^[1]

The paper studied how to choose the context examples. Retrieving examples similar to the target utterance using Meta's SONAR multilingual embeddings lowered error rates compared with random selection, by up to 11.2% relative.^[2]

Model family

The repository lists the following checkpoints. Parameter counts for LLM-ASR models include the 1.2B decoder.^[3]

The README reports inference measurements on an A100 GPU with 30-second audio in BF16. The 300M CTC model needs about 2 GiB of VRAM and is listed as 96 times faster than the 7B LLM-ASR model, which needs about 17 GiB.^[3] The README states that the standard CTC and LLM models accept only audio files shorter than 40 seconds; the Unlimited variants were added to handle longer recordings.^[3] Meta describes the 300M variants as designed for low-power devices and the 7B models as offering the highest accuracy.^[1][2]

Training data

Sources

The labeled training set, called AllASR in the paper, combines existing public datasets (among them Common Voice v22, FLEURS, Multilingual LibriSpeech, VoxPopuli, VoxLingua-107 and NaijaVoices), an internal dataset of publicly available recordings with transcripts, licensed data including the 17 IARPA Babel language packs, and data from partner projects.^[2] The partner projects named in the paper and blog are:^[1][2]

Omnilingual ASR Corpus

The commissioned part of the training data was released as the Omnilingual ASR Corpus on Hugging Face (facebook/omnilingual-asr-corpus) under CC-BY 4.0.^[3][4] The dataset card describes it as spontaneous speech recordings and transcriptions for 348 under-served languages; Meta's blog rounds this to 350.^[1][4] The original plan was 10 hours from 10 native speakers (one hour each) for roughly 350 to 400 languages. Vendors recruited and paid speakers, often in remote regions, and gave them a pool of more than 1,500 open-ended prompts (for example, "Is it better to have a few close friends or many casual acquaintances? Why?") so that they could choose topics and speak naturally.^[2] Transcripts mark laughter, hesitations, unintelligible words and noise with tags such as <laugh> and <hesitation>, and record false starts and repetitions as spoken.^[2][4]

Quality control included automated screening, manual inspection of every language and in-depth review of 279 languages. A cross-vendor validation step on 206 languages found misattributed language codes in 20 of them.^[2] Languages are identified by ISO 639-3 codes plus ISO 15924 script codes, with Glottolog codes where finer distinctions were needed, such as separating Vallader and Sutsilvan Romansh.^[2][4] Meta says the corpus is the largest ultra-low-resource spontaneous ASR dataset made available to date.^[1]

Evaluation

All results below are reported by Meta in its own paper and have not been independently replicated in the sources cited here. The main metric is character error rate (CER), averaged across languages; lower is better.^[2]

Coverage-wide results

Meta's headline figure is that the 7B LLM-ASR model reaches a CER below 10 for 78% of the languages it evaluated.^[1][3] The paper's language-grouping analysis gives 1,237 of 1,570 languages (78%) at CER 10 or below, with an average CER of 7.1. Every one of 14 language groupings averaged below 10 except the Afroasiatic group, at 11.8.^[2] Results depend heavily on how much labeled data a language has:^[2]

Comparison with other systems

Sources: Tables 5, 6 and 7 of the paper.^[2]

Against Whisper large-v3, the 7B LLM-ASR model had a lower CER on 65 of the 81 FLEURS languages (80%) and on 24 of the 34 FLEURS languages that are among the world's 50 most spoken. Whisper large-v3 remained ahead on the eight high-resource languages of MLS, where its average CER was 2.0 against 2.4.^[2] The authors attribute their edge over USM mainly to encoder scaling, noting that their encoders saw less unlabeled speech than USM (4.3 million versus 12 million hours).^[2] They also caution that error rates on different benchmarks are not directly comparable, because read-speech sets such as MMS-Lab are much easier than spontaneous speech from many speakers.^[2]

Zero-shot results

To test the zero-shot model, the team held out 32 languages from training. A CTC baseline trained without them averaged a CER of 26.3 on those languages; the zero-shot LLM-ASR model with 10 context examples brought this down to 14.4.^[2] The paper notes that zero-shot models are somewhat less accurate on some datasets of languages they were trained on, which is why separate supervised models are released for those.^[2]

Speech translation

With source and target language tokens added to the input, the paper also trains LLM-ASR models for speech-to-text translation into English. On CoVoST 2 (21 languages into English), the 7B variant scored 37.1 BLEU, compared with 29.1 for Whisper large-v2 and 34.1 for SeamlessM4T v1 Large; AudioPaLM-2-8B-AST, also listed in the paper's comparison, scored higher at 37.8. On FLEURS it trailed SeamlessM4T v1 Large by 0.5 to 0.6 BLEU.^[2] These translation models are a research result in the paper; the repository's model list covers ASR checkpoints.^[2][3]

Availability

The models were developed with fairseq2, Meta's PyTorch-based sequence modeling toolkit, and are installed as the omnilingual-asr Python package. Checkpoints download automatically on first use.^[1][3][9] The repository includes an inference pipeline, data preparation guides and fine-tuning recipes for the CTC and LLM models, and a CSV of per-language CER and training hours for the 7B LLM-ASR model.^[3] Meta also published a transcription demo on Hugging Face Spaces and a language exploration demo.^[1][3]

Applications and recognition

The paper reports that health practitioners in Nigeria were using the system to transcribe Hausa in community clinics, with the aim of improving documentation and patient care.^[2] TIME's Best Inventions of 2026 entry, written by Richard Eisenberg, repeated that example and quoted Jean Maillard, described as a lead researcher on the project: "We're very passionate about making sure technology supports as many languages as possible."^[5] AI at Meta announced the recognition on X on 15 October 2026.^[6] Contemporary coverage at launch, such as The Decoder's, focused on the 1,600-language figure, the in-context extension mechanism and the Apache 2.0 license.^[7]

The TIME entry also mentions Meta's Omnilingual Machine Translation system, which it says can translate the resulting transcripts across more than 1,600 languages.^[5] That system is described in a separate paper, "Omnilingual MT: Machine Translation for 1,600 Languages," first submitted to arXiv on 17 March 2026. It explores two designs, a decoder-only model (OMT-LLaMA) and an encoder-decoder model (OMT-NLLB), with models from 1B to 8B parameters, and the authors report that these match or exceed a 70B LLM baseline on their translation evaluations.^[8]

Limitations and ethical considerations

The authors list several limits. Zero-shot quality is below that of supervised models. Accuracy for languages with under 10 hours of training data is much lower than for better-resourced ones, and the paper offers a recipe for fine-tuning per-language models to close some of that gap.^[2] Corpus quality checks found transcripts with mismatches, summarized content and wrong writing systems, and the language-code validation covered only a small sample of the data.^[2] The paper also warns that the same capabilities could be used for purposes that conflict with community priorities, such as surveillance, and calls for participatory governance. It notes that paying speakers is "not a panacea," since some communities may prefer voluntary participation and others may feel financial pressure to contribute.^[2]

References

1. Meta AI. "Omnilingual ASR: Advancing Automatic Speech Recognition for 1,600+ Languages." 10 November 2025. https://ai.meta.com/blog/omnilingual-asr-advancing-automatic-speech-recognition/
2. Omnilingual ASR team, Keren, G., Kozhevnikov, A., et al. "Omnilingual ASR: Open-Source Multilingual Speech Recognition for 1600+ Languages." arXiv:2511.09690, 12 November 2025. https://arxiv.org/abs/2511.09690
3. facebookresearch/omnilingual-asr. GitHub repository (README, model table, December 2025 update, license). https://github.com/facebookresearch/omnilingual-asr
4. Meta. "Meta Omnilingual ASR Corpus" (facebook/omnilingual-asr-corpus). Hugging Face dataset card. https://huggingface.co/datasets/facebook/omnilingual-asr-corpus
5. Eisenberg, Richard. "Meta Omnilingual ASR." TIME, The Best Inventions of 2026, 15 October 2026. https://time.com/collection/best-inventions/2026/meta-omnilingual-asr/
6. AI at Meta (@AIatMeta). Post on X announcing TIME Best Inventions recognition, 15 October 2026. https://x.com/AIatMeta/status/2110861420321759269
7. Bastian, Matthias. "Meta's Omnilingual ASR brings speech recognition to 1,600 languages." The Decoder, 10 November 2025. https://the-decoder.com/metas-omnilingual-asr-brings-speech-recognition-to-1600-languages/
8. Omnilingual MT Team, Alastruey, B., et al. "Omnilingual MT: Machine Translation for 1,600 Languages." arXiv:2603.16309, 17 March 2026. https://arxiv.org/abs/2603.16309
9. "omnilingual-asr." Python Package Index (PyPI). https://pypi.org/project/omnilingual-asr/