What is Meta’s Omnilingual ASR?

On November 10, 2025, Meta AI unveiled Omnilingual ASR, a new suite of automatic speech recognition (ASR) models that can transcribe speech in more than 1,600 languages — including around 500 low-resource languages never previously supported by any ASR model.

The release includes models of varying sizes — from compact 300M-parameter variants for low-power devices to full-scale 7B models for maximum accuracy.

Built for “broad accessibility,” the model can be extended to new languages using only a few paired audio-text examples, without requiring large datasets or full retraining.

According to Meta, “Omnilingual ASR aims to make speech technology more inclusive and adaptable for communities and researchers worldwide.”

The company reported a character error rate below 10% for 78% of the 1,570 languages tested, outperforming other systems — including OpenAI’s Whisper — especially in low-resource conditions, and showing strong generalization to unseen languages.

Meta said the model’s impact is already tangible. In Nigeria, health practitioners are using Omnilingual ASR to facilitate Hausa transcriptions in community clinics, improving documentation and patient care. The company also highlighted potential applications in education, cultural preservation, and media accessibility — from making endangered language archives searchable to enabling local-language transcription of broadcasts.

Alongside the model, Meta released the Omnilingual ASR Corpus, a large-scale collection of transcribed speech in 350 underserved languages, gathered through partnerships with local communities. The company described it as “the largest ultra-low-resource spontaneous ASR dataset ever made available.”

Meta has open-sourced both the model weights and datasets under the Apache 2.0 license, aiming to accelerate multilingual speech research and commercial applications.

“By open sourcing these models and dataset, we aim to break down language barriers, expand digital access, and empower communities worldwide,” the company wrote on X.

Initial Reactions and Opportunities Identified

Reaction across X and LinkedIn was immediate and positive. Posts celebrating the model’s reach highlighted that “most voice AI systems ignore 90 % of the world’s languages,” due to data scarcity, and that Omnilingual ASR “breaks that cycle.”

“If we can build models that work across dialects, cultures, and scarce data, the future of voice AI in enterprise, customer service, and global markets changes fast,” said Armand Ruiz, Director of AI Engineering at IBM.

“If we can build models that work across dialects, cultures, and scarce data, the future of voice AI in enterprise, customer service, and global markets changes fast,” — Armand Ruiz, Director of AI Engineering, IBM

Practitioners working on low-resource ASR said they were eager to test the model, calling the release “incredible” for languages that had zero digital presence. Others focused on the accessibility impact, arguing that “low-resource languages have never been more accessible.”

Several highlighted the model’s accuracy and reach — achieving under 10% character error rate for 78% of languages — while others praised its ability to “learn new languages with just a few audio samples.” 

“In most existing systems, languages not included at release time can only be added through expert-driven fine-tuning — a path inaccessible to most communities. Omnilingual ASR instead introduces the first large-scale ASR framework capable of extending to entirely new languages with just a few in-context examples,” wrote Laurent Le Pen, Founder & CEO at Oxtak, an AI platform centered on voice technology, adding that it “shifts the paradigm for how new languages can be brought into the fold.”

Others emphasised the breadth and openness of the release, calling it “a huge step toward a more inclusive digital future,” “a monumental leap toward linguistic inclusion,” and “a major step toward making speech AI accessible to everyone, everywhere, regardless of language coverage.”

“In most existing systems, languages not included at release time can only be added through expert-driven fine-tuning — a path inaccessible to most communities. Omnilingual ASR instead introduces the first large-scale ASR framework capable of extending to entirely new languages with just a few in-context examples,” — Laurent Le Pen, Founder & CEO, Oxtak

Many also highlighted the possibilities the release opens — enabling researchers, developers, and organizations to build inclusive, multilingual applications ranging from speech-to-text tools and translation engines to accessibility features for global platforms. They also noted that ASR and text-to-speech are “two sides of the same coin — when one advances, the other accelerates too.” 

Additionally, optimized for both high-end and resource-constrained hardware, the models allow flexible deployment and help reduce reliance on limited commercial APIs, they noted.

Several observers also framed the launch as a “strategic reset” for Meta’s AI division, marking a return to its strengths in multilingual research and open-source infrastructure after a year with mixed reviews for its latest large language model, Llama 4.