Meta has open-sourced an AI project that it says expands text-to-speech and speech-to-text capabilities from 100 languages to more than 1,000. The Massively Multilingual Speech (MMS) models could make it easier to access information and use devices in a preferred language, including less common languages, according to the company. - Meta says its MMS models can identify over 4,000 spoken languages, a 40x improvement over previous technologies.
- The models are capable of generating and understanding speech in more than 1,100 languages.
- Meta said it expanded the model's language coverage by using unlabeled audio recordings from different translated religious texts, including the Bible.
- The company envisions a future where "a single model can solve several speech tasks for all languages."
- Meta said it shared the models and code as part of its effort to preserve language diversity and promote further research in the space.
|