Meta Expands AI’s Vocabulary to 1,600 Languages
With its Omnilingual speech model trained on 1,600 tongues, Meta takes a bold step toward making voice AI global, and giving low-resource languages a digital voice.
News
- Meta Expands AI’s Vocabulary to 1,600 Languages
- Google Bridges Cloud Power and Privacy with Private AI Compute
- California’s Genspark Enters AI Unicorn Club with $200 Million Round
- OpenAI Taps Intel’s Sachin Katti to Build Compute Backbone for AGI
- Flipkart Brings Voice-Led Wholesale Ordering to WhatsApp with Sarvam AI
- Varun Berry Steps Down at Britannia, Rakshit Hargave to Take Over
Meta’s Fundamental AI Research (FAIR) team has unveiled an open-source suite of “Omnilingual” automatic speech recognition (ASR) models capable of understanding and transcribing more than 1,600 spoken languages, including 500 low-resource languages never before supported by AI transcription tools.
The announcement was made in a LinkedIn post about AI at Meta, alongside a detailed company blog that said, “These high-quality speech-to-text systems are accessible to even the most underrepresented language communities.”
The launch marks a major advance for linguistically diverse regions such as India, where many low-resource languages have lacked AI transcription support.
Meta acknowledged the challenge of scaling ASR technology, noting that “expanding language coverage has been prohibitively resource intensive as current AI architectures are too data demanding to scale universally.”
To overcome this, the company expanded its earlier wav2vec 2.0 speech encoder to seven billion parameters, its largest yet, enabling more detailed multilingual representations from raw speech data.
Two decoding approaches were used to convert speech into text. The first, based on Connectionist Temporal Classification (CTC), aligns sounds to words without requiring exact timing.
The second employs a transformer model, similar to those powering large language models, that interprets speech contextually, producing smoother and more natural transcriptions.
Founded in 2013, Meta’s FAIR team advances artificial intelligence through open and responsible research. It noted that most existing ASR systems cater to dominant languages like English, Mandarin, and Spanish, which have extensive digital data. In contrast, “low-resource” languages have long lacked reliable speech transcription tools.
The Omnilingual ASR system aims to close that gap by supporting more than 1,600 languages. Adding 500 low-resource ones, many with minimal digital documentation or training data, represents a significant step for language preservation and accessibility.
Described as one of the largest speech recognition projects to date, Omnilingual ASR was trained using both public datasets and community-sourced recordings.
Meta partnered with the Mozilla Foundation’s Common Voice, Lanfrica, and NaijaVoices to collect real-world samples across accents, dialects, and speech styles.
“These partnerships have been instrumental in infusing Omnilingual ASR with deep linguistic knowledge and cultural understanding,” the company said.
To capture voices with little or no digital presence, Meta collaborated with local organizations that located and compensated native speakers, often in remote or under-documented regions.
Omnilingual ASR can also perform in-context learning, enabling it to incorporate new languages using only a few audio samples. This allows rapid improvement even when data is scarce, a crucial factor for under-represented languages.
Meta said performance varies by language data availability: more than 95% of high- and medium-resource languages achieved a character error rate below 10%, while only 36% of low-resource ones reached that benchmark.
The company acknowledged these gaps reflect the broader complexity of global linguistic diversity and underscore the need for continued data collection and model refinement.
By open-sourcing the Omnilingual ASR models and the accompanying ASR Corpus, covering over 350 underserved languages, Meta aims to enable researchers, developers, and organizations worldwide to adapt and enhance speech recognition systems.
The company said this move will spur innovation in speech-to-text, translation, and accessibility applications, while enabling transparency and faster progress in multilingual AI.
Meta said the initiative will help accelerate the adoption of multilingual AI across education, industry, government, and consumer technology worldwide.
