Meta's Omnilingual ASR Brings Speech Recognition to 1,600 Languages
Meta's open-source speech system covering 1,600+ languages, including 500 never before transcribed by AI, lands on TIME's Best Inventions list.
- Meta's Omnilingual ASR named one of TIME's Best Inventions of 2026.
- Covers 1,600+ languages, including 500 never before served by any ASR system.
- Zero-shot in-context learning extends coverage to a theoretical 5,400+ languages with a few samples.
- Character error rate under 10% on 78% of supported languages, trained on 4.3M hours of audio.
- Full suite released under Apache 2.0: wav2vec 2.0, CTC, LLM-ASR, and zero-shot variants (300M-7B params).
- Beats Google USM on FLEURS 102 benchmarks despite using one-third the unlabeled training hours.
Meta’s Omnilingual ASR transcribes 1,600 languages and can extend to more
Meta’s Omnilingual Automatic Speech Recognition suite was included on TIME’s 2026 list for its coverage of more than 1,600 languages. Meta says roughly one-third of those languages had no working speech-to-text model before the release.
The suite gives developers several open model families for conventional transcription, language-model-based decoding, and adaptation to unseen languages from a small set of examples. Meta’s release notes describe the models, corpus, and access points.
A language list that can grow
Conventional multilingual ASR systems usually fix their output tokens and supported scripts during training. Adding a language can require new labeled audio, vocabulary changes, decoder updates, and another training run, which concentrates development on languages with large digital datasets.
More than 7,000 languages are spoken worldwide, while reliable speech-recognition tooling covers only a fraction of them. Omnilingual ASR narrows that gap through shared multilingual encoders and an LLM-style decoder that can adapt from paired audio and text examples supplied at inference time.
Five pieces, from encoders to data
Meta reports pretraining its speech encoders on about 4.3 million hours of unlabeled audio spanning more than 1,600 languages. Self-supervised training lets an encoder learn acoustic patterns from recordings that lack transcripts.
| Component | Scale | Role |
|---|---|---|
| wav2vec 2.0 encoders | 300 million to 7 billion parameters | Learn reusable speech representations from unlabeled audio. |
| CTC ASR models | Multiple sizes | Use Connectionist Temporal Classification for efficient supervised transcription. |
| LLM-ASR | Includes a 7-billion-parameter model | Applies a language-model-style decoder to speech representations. |
| LLM-ZeroShot ASR | Inference-time adaptation | Transcribes an unseen language from a small prompt of paired speech and text. |
| Omnilingual ASR Corpus | More than 350 underserved languages | Provides spontaneous speech data for training and evaluation. |
Meta released the suite under the Apache License 2.0. The license permits commercial use, modification, and redistribution, subject to its notice, attribution, and other compliance requirements.
A few examples can define a language
In Meta’s “language-in-the-box” workflow, a user supplies a small set of audio clips and matching transcripts in the target language. The LLM-ZeroShot ASR model treats those demonstrations as an inference-time prompt and transcribes new audio without gradient updates or a dedicated training run.
Meta extrapolates that this method could extend coverage to more than 5,400 languages. The published evaluation covers 1,600 languages, and additional languages will require testing for transcription quality, script handling, prompt sensitivity, and compatibility with the model’s learned speech representations.
Accuracy follows the available data
Meta evaluates transcription using character error rate, or CER, which divides character insertions, deletions, and substitutions by the number of characters in the reference transcript. A CER below 10% corresponds to fewer than 10 character-level errors per 100 reference characters on average.
| Evaluation group | Languages below 10% CER |
|---|---|
| All 1,600 evaluated languages | 78% |
| Languages with at least 10 hours of training audio | 95% |
| Languages with less than 10 hours of training audio | 36% |
On the 102-language FLEURS benchmark, Meta reports that its 7-billion-parameter LLM-ASR model achieved a lower average CER than its 7-billion-parameter CTC models and Google Universal Speech Model variants. Meta’s model used about 4.3 million hours of unlabeled pretraining audio, compared with 12 million hours reported for Google’s system.
Low-resource performance remains uneven, with 64% of languages in the under-10-hour group missing the 10% CER threshold. Languages with limited paired data, no widely adopted writing system, or phonetic and script patterns far from the training set present the hardest cases. Meta also reports stronger in-context adaptation when the target language shares relevant features with languages represented during training.
Deployment depends on the edges
The suite can support voice assistants, transcription services, subtitles, oral-archive digitization, and accessibility tools. A shared model family can reduce the number of language-specific vendors and pipelines, although each target language still needs evaluation under the product’s actual recording conditions.
Product teams should compare the available models against concrete deployment requirements:
- Latency and cost: Measure real-time factor, memory use, batch throughput, and accelerator requirements across model sizes.
- Language quality: Test native speakers, regional accents, code-switching, background noise, and domain-specific vocabulary.
- Prompt robustness: Vary the number, quality, speakers, and ordering of examples used for inference-time adaptation.
- Text normalization: Define expected scripts, punctuation, numerals, capitalization, and spelling variants before scoring output.
- Operational controls: Decide where audio is processed, logged, retained, and reviewed, especially for sensitive recordings.
- License compliance: Preserve required Apache 2.0 notices and review the terms attached to datasets used for further training.
Local partners shaped the corpus
Meta says it worked with local organizations to recruit and compensate native speakers, including people in remote and under-documented regions. The commissioned recordings form the Omnilingual ASR Corpus, which covers more than 350 underserved languages and includes spontaneous speech.
Compensation and local recruitment provide clearer provenance and consent for the commissioned portion of the dataset than indiscriminate collection from the web. Teams using the corpus should still review its documentation for consent scope, permitted uses, demographic coverage, and known sampling gaps.
The research paper provides architecture, training, and benchmark details for reproducing Meta’s comparisons. Production decisions will depend on per-language tests covering target accents, acoustic environments, scripts, vocabulary, latency, and hardware.
Comments
