\n
![](\"https:\/\/images.ctfassets.net\/jdtwqhzvc2n1\/5WF8w75sB7wnAYEK5pYnPD\/16059c6ac3f6b853b3b301bc66f950d3\/cfr0z3n_graphic_novel_abstract_expressionist_outline_style_show_390da294-e50e-424d-ab36-20b02133d3d9.png?w=300&q=30\")
<\/p>\n
\n<\/p>\n
Meta has just released a new multilingual automatic speech recognition (ASR) system supporting 1,600+ languages \u2014 dwarfing OpenAI\u2019s open source Whisper model, which supports just 99. <\/p>\n
Is architecture also allows developers to extend that support to thousands more. Through a feature called zero-shot in-context learning, users can provide a few paired examples of audio and text in a new language at inference time, enabling the model to transcribe additional utterances in that language without any retraining.<\/p>\n
In practice, this expands potential coverage to more than 5,400 languages \u2014 roughly every spoken language with a known script.<\/p>\n
It\u2019s a shift from static model capabilities to a flexible framework that communities can adapt themselves. So while the 1,600 languages reflect official training coverage, the broader figure represents Omnilingual ASR\u2019s capacity to generalize on demand, making it the most extensible speech recognition system released to date.<\/p>\n
Best of all: it's been open sourced under a plain Apache 2.0 license \u2014 not a restrictive, quasi open-source Llama license like the company's prior releases, which limited use by larger enterprises unless they paid licensing fees \u2014 meaning researchers and developers are free to take and implement it right away, for free, without restrictions, even in commercial and enterprise-grade projects!<\/p>\n
Released on November 10 on Meta's website, Github, along with a demo space on Hugging Face and technical paper, Meta\u2019s Omnilingual ASR suite includes a family of speech recognition models, a 7-billion parameter multilingual audio representation model, and a massive speech corpus spanning over 350 previously underserved languages. <\/p>\n
All resources are freely available under open licenses, and the models support speech-to-text transcription out of the box.<\/p>\n
\u201cBy open sourcing these models and dataset, we aim to break down language barriers, expand digital access, and empower communities worldwide,\u201d Meta posted on its @AIatMeta account on X<\/p>\n
At its core, Omnilingual ASR is a speech-to-text system. <\/p>\n
The models are trained to convert spoken language into written text, supporting applications like voice assistants, transcription tools, subtitles, oral archive digitization, and accessibility features for low-resource languages.<\/p>\n
Unlike earlier ASR models that required extensive labeled training data, Omnilingual ASR includes a zero-shot variant. <\/p>\n
This version can transcribe languages it has never seen before\u2014using just a few paired examples of audio and corresponding text. <\/p>\n
This lowers the barrier for adding new or endangered languages dramatically, removing the need for large corpora or retraining.<\/p>\n
The Omnilingual ASR suite includes multiple model families trained on more than 4.3 million hours of audio from 1,600+ languages:<\/p>\n
wav2vec 2.0 models for self-supervised speech representation learning (300M\u20137B parameters)<\/p>\n<\/li>\n
CTC-based ASR models for efficient supervised transcription<\/p>\n<\/li>\n
LLM-ASR models combining a speech encoder with a Transformer-based text decoder for state-of-the-art transcription<\/p>\n<\/li>\n
LLM-ZeroShot ASR model, enabling inference-time adaptation to unseen languages<\/p>\n<\/li>\n<\/ul>\n
All models follow an encoder\u2013decoder design: raw audio is converted into a language-agnostic representation, then decoded into written text.<\/p>\n
While Whisper and similar models have advanced ASR capabilities for global languages, they fall short on the long tail of human linguistic diversity. Whisper supports 99 languages. Meta\u2019s system:<\/p>\n
Directly supports 1,600+ languages<\/p>\n<\/li>\n
Can generalize to 5,400+ languages using in-context learning<\/p>\n<\/li>\n
Achieves character error rates (CER) under 10% in 78% of supported languages<\/p>\n<\/li>\n<\/ul>\n
Among those supported are more than 500 languages never previously covered by any ASR model, according to Meta\u2019s research paper.<\/p>\n
This expansion opens new possibilities for communities whose languages are often excluded from digital tools<\/p>\n
Here\u2019s the revised and expanded background section, integrating the broader context of Meta\u2019s 2025 AI strategy, leadership changes, and Llama 4\u2019s reception, complete with in-text citations and links:<\/p>\n
The release of Omnilingual ASR arrives at a pivotal moment in Meta\u2019s AI strategy, following a year marked by organizational turbulence, leadership changes, and uneven product execution. <\/p>\n
Omnilingual ASR is the first major open-source model release since the rollout of Llama 4, Meta\u2019s latest large language model, which debuted in April 2025 to mixed and ultimately poor reviews, with scant enterprise adoption compared to Chinese open source model competitors.<\/p>\n
The failure led Meta founder and CEO Mark Zuckerberg to appoint Alexandr Wang, co-founder and prior CEO of AI data supplier Scale AI, as Chief AI Officer, and embark on an extensive and costly hiring spree that shocked the AI and business communities with eye-watering pay packages for top AI researchers.<\/p>\n
In contrast, Omnilingual ASR represents a strategic and reputational reset. It returns Meta to a domain where the company has historically led \u2014 multilingual AI \u2014 and offers a truly extensible, community-oriented stack with minimal barriers to entry. <\/p>\n
The system\u2019s support for 1,600+ languages and its extensibility to over 5,000 more via zero-shot in-context learning reassert Meta\u2019s engineering credibility in language technology. <\/p>\n
Importantly, it does so through a free and permissively licensed release, under Apache 2.0, with transparent dataset sourcing and reproducible training protocols.<\/p>\n
This shift aligns with broader themes in Meta\u2019s 2025 strategy. The company has refocused its narrative around a \u201cpersonal superintelligence\u201d vision, investing heavily in infrastructure (including a September release of custom AI accelerators and Arm-based inference stacks) source while downplaying the metaverse in favor of foundational AI capabilities. The return to public training data in Europe after a regulatory pause also underscores its intention to compete globally, despite privacy scrutiny source.<\/p>\n
Omnilingual ASR, then, is more than a model release \u2014 it\u2019s a calculated move to reassert control of the narrative: from the fragmented rollout of Llama 4 to a high-utility, research-grounded contribution that aligns with Meta\u2019s long-term AI platform strategy.<\/p>\n
To achieve this scale, Meta partnered with researchers and community organizations in Africa, Asia, and elsewhere to create the Omnilingual ASR Corpus, a 3,350-hour dataset across 348 low-resource languages. Contributors were compensated local speakers, and recordings were gathered in collaboration with groups like:<\/p>\n
African Next Voices<\/b>: A Gates Foundation\u2013supported consortium including Maseno University (Kenya), University of Pretoria, and Data Science Nigeria<\/p>\n<\/li>\n Mozilla Foundation\u2019s Common Voice<\/b>, supported through the Open Multilingual Speech Fund<\/p>\n<\/li>\n Lanfrica \/ NaijaVoices<\/b>, which created data for 11 African languages including Igala, Serer, and Urhobo<\/p>\n<\/li>\n<\/ul>\n The data collection focused on natural, unscripted speech. Prompts were designed to be culturally relevant and open-ended, such as \u201cIs it better to have a few close friends or many casual acquaintances? Why?\u201d Transcriptions used established writing systems, with quality assurance built into every step.<\/p>\n The largest model in the suite, the omniASR_LLM_7B, requires ~17GB of GPU memory for inference, making it suitable for deployment on high-end hardware. Smaller models (300M\u20131B) can run on lower-power devices and deliver real-time transcription speeds.<\/p>\n Performance benchmarks show strong results even in low-resource scenarios:<\/p>\n CER <10% in 95% of high-resource and mid-resource languages<\/p>\n<\/li>\n CER <10% in 36% of low-resource languages<\/p>\n<\/li>\n Robustness in noisy conditions and unseen domains, especially with fine-tuning<\/p>\n<\/li>\n<\/ul>\n The zero-shot system, omniASR_LLM_7B_ZS, can transcribe new languages with minimal setup. Users provide a few sample audio\u2013text pairs, and the model generates transcriptions for new utterances in the same language.<\/p>\n All models and the dataset are licensed under permissive terms:<\/p>\n Apache 2.0<\/b> for models and code<\/p>\n<\/li>\n CC-BY 4.0<\/b> for the Omnilingual ASR Corpus on HuggingFace<\/p>\n<\/li>\n<\/ul>\n Installation is supported via PyPI and uv:<\/p>\n Meta also provides:<\/p>\n A HuggingFace dataset integration<\/p>\n<\/li>\n Pre-built inference pipelines<\/p>\n<\/li>\n Language-code conditioning for improved accuracy<\/p>\n<\/li>\n<\/ul>\n Developers can view the full list of supported languages using the API:<\/p>\n Omnilingual ASR reframes language coverage in ASR from a fixed list to an extensible framework<\/b>. It enables:<\/p>\n Community-driven inclusion of underrepresented languages<\/p>\n<\/li>\n Digital access for oral and endangered languages<\/p>\n<\/li>\n Research on speech tech in linguistically diverse contexts<\/p>\n<\/li>\n<\/ul>\n Crucially, Meta emphasizes ethical considerations throughout\u2014advocating for open-source participation and collaboration with native-speaking communities.<\/p>\n \u201cNo model can ever anticipate and include all of the world\u2019s languages in advance,\u201d the Omnilingual ASR paper states, \u201cbut Omnilingual ASR makes it possible for communities to extend recognition with their own data.\u201d<\/p>\n All resources are now available at:<\/p>\n Code + Models<\/b>: github.com\/facebookresearch\/omnilingual-asr<\/p>\n<\/li>\n Dataset<\/b>: huggingface.co\/datasets\/facebook\/omnilingual-asr-corpus<\/p>\n<\/li>\n Blogpost<\/b>: ai.meta.com\/blog\/omnilingual-asr<\/p>\n<\/li>\n<\/ul>\n For enterprise developers, especially those operating in multilingual or international markets, Omnilingual ASR significantly lowers the barrier to deploying speech-to-text systems across a broader range of customers and geographies. <\/p>\n Instead of relying on commercial ASR APIs that support only a narrow set of high-resource languages, teams can now integrate an open-source pipeline that covers over 1,600 languages out of the box\u2014with the option to extend it to thousands more via zero-shot learning.<\/p>\n This flexibility is especially valuable for enterprises working in sectors like voice-based customer support, transcription services, accessibility, education, or civic technology, where local language coverage can be a competitive or regulatory necessity. Because the models are released under the permissive Apache 2.0 license, businesses can fine-tune, deploy, or integrate them into proprietary systems without restrictive terms.<\/p>\n It also represents a shift in the ASR landscape\u2014from centralized, cloud-gated offerings to community-extendable infrastructure. By making multilingual speech recognition more accessible, customizable, and cost-effective, Omnilingual ASR opens the door to a new generation of enterprise speech applications built around linguistic inclusion rather than linguistic limitation.<\/p>\nPerformance and Hardware Considerations<\/b><\/h3>\n
\n
Open Access and Developer Tooling<\/b><\/h3>\n
\n
pip install omnilingual-asr<\/code><\/p>\n\n
from omnilingual_asr.models.wav2vec2_llama.lang_ids import supported_langs<\/code><\/p>\nprint(len(supported_langs))
\nprint(supported_langs)<\/code><\/p>\nBroader Implications<\/b><\/h3>\n
\n
Access the Tools<\/b><\/h3>\n
\n
What This Means for Enterprises<\/b><\/h3>\n