Meta has launched the Omnilingual Automatic Speech Recognition (ASR) artificial intelligence (AI) model, which is capable of automatically recognizing over 1600 languages. It was developed using 7 billion parameters and is offered as open source using the Apache 2.0 license. Of these 1600 languages, 78% of the languages can be translated with an error of less than 10%.
To train Omnilingual ASR, Meta used 249 languages with high resource usage, 881 languages with medium resource usage, and 546 languages with low resource usage. According to Meta, Omnilingual ASR can theoretically be expanded to support up to 5400 world languages. This is a far cry from OpenAI's Whisper language translation model, which only supports 99 major languages.
Meta added that Omnilingual ASR also supports 500 new languages that were not supported by any previous automatic speech recognition (ASR) system. Omnilingual ASR can be used to automatically identify languages in audio and text, with transcripts being produced simultaneously.
