May 23, 2023 – In a groundbreaking move, Meta Corporation has introduced the Massively Multilingual Speech (MMS) project, revolutionizing language-to-text and text-to-speech services across more than 1100 languages. What sets MMS apart is its remarkable ability to process nearly 4000 unannotated spoken languages, surpassing existing models by tenfold.
At the core of Meta’s achievement lies the integration of their self-supervised learning work, wav2vec 2.0, with a novel dataset. Through rigorous testing, MMS has proven to outperform its predecessors while supporting an unprecedented number of languages.
Meta’s success in this endeavor can be attributed to the incorporation of an extensive audio dataset. This corpus comprises readings of the New Testament in over 1100 languages, ensuring that the resulting speech models excel in both male and female voices.