Meta's Fundamental AI Research (FAIR) team has recently made the automatic speech recognition system, Omnilingual ASR, open-source. This system boasts the remarkable ability to transcribe speech in over 1,600 different languages. Developed using the PyTorch-powered fairseq2 framework, Omnilingual ASR comes in versions with varying parameters. This initiative is a significant step toward closing the language coverage gap in AI tools and moving closer to the ambitious vision of a 'universal transcription system.' Notably, it brings 500 languages into the AI fold that were previously overlooked by any AI system.

Test outcomes reveal that the system achieves an error rate of fewer than 10 characters for 78% of the languages it supports. In terms of standard coverage accuracy, it attains a 95% accuracy rate for 'resource-abundant' languages and a 36% accuracy rate for 'resource-scarce' languages. The 'bring-your-own-language' feature of Omnilingual ASR empowers the system to acquire proficiency in new languages using just a small set of samples. This theoretically expands its reach to cover more than 5,400 languages. Furthermore, Meta has also introduced a comprehensive automatic speech recognition corpus encompassing a wide array of languages. This resource is designed to assist developers in fine-tuning models to cater to localization requirements.