Meta's MMS halved Whisper's error rate while covering 11x more languages

When Meta released its Massively Multilingual Speech (MMS) models in May 2023, it compared them directly with OpenAI’s Whisper speech-recognition system. On the languages the two systems shared, Meta reported that MMS achieved half the word error rate of Whisper while covering 11 times as many languages overall - a rare case of a model being both more accurate and far broader at the same time. MMS reached this breadth partly by training on New Testament audio recordings, which exist in more than 1,100 languages.