Meta launched the Massively Multilingual Speech (MMS) project, which provides advanced language-to-text and text-to-speech services for more than 1100 languages, and supports nearly 4000 kinds of untagged spoken languages.
Meta integrated self-supervised learning work wav2vec 2.0 and a new dataset for the MMS project and found that MMS outperformed existing models and supported 10 times as many languages through real-world testing.
The key to Meta’s achievement was the introduction of a large audio dataset containing more than 1,100 readings from the New Testament, and Meta says that the integration of this audio dataset produced speech models that performed equally well in male and female voices.
The original Meta article is attached here, and interested users can read it here.
