Meta FAIR has unveiled Omnilingual ASR, a groundbreaking open-source automatic speech recognition (ASR) system that supports over 1,600 languages, with future potential to extend to more than 5,400 languages through zero-shot in-context learning. This innovative system is designed to address the needs of both high-resource and low-resource language communities, offering support for over 500 languages that previously lacked transcription resources.
Omnilingual ASR employs a 7 billion-parameter wav2vec 2.0 encoder coupled with two distinct decoder variants: a standard Connectionist Temporal Classification (CTC) decoder and a language model-based decoder. The latter enhances in-context learning by allowing users to achieve usable transcription quality with only a few audio-text samples, eliminating the necessity for extensive training data.
This system has demonstrated impressive performance, achieving character error rates below 10% for 78% of the supported languages. With a capacity to outperform OpenAI’s Whisper, which supports 99 languages, Omnilingual ASR stands as a formidable contender in the realm of speech recognition technology.
The comprehensive model suite released under Apache 2.0 and CC-BY licenses ranges in size from 300 million to 7 billion parameters. The distribution includes the Omnilingual ASR Corpus, consisting of 3,350 hours of speech data from 348 low-resource languages, compiled through collaborations with African and global language communities.
Meta’s implementation and deployment of Omnilingual ASR represent a significant leap forward in automatic speech recognition, offering expansive language coverage and fostering inclusivity in technology access.

Collections

Meta FAIR released Omnilingual ASR, an open-source speech recognition system supporting over 1,600 languages with potential expansion to 5,400+ through zero-shot learning. The system uses a 7B-parameter wav2vec 2.0 encoder with CTC and language model-based decoders, achieving sub-10% character error rates for 78% of supported languages. Released under Apache 2.0 and CC-BY licenses, it includes models ranging from 300M to 7B parameters and a corpus of 3,350 hours of speech data from 348 low-resource languages, outperforming OpenAI's Whisper which supports 99 languages.

Introducing Omnilingual ASR: A Pioneering Open-Source Speech Recognition System by Meta FAIR