The research division of Meta is working on a universal translator that will simplify the interaction of users in the company's digital universe.
“This will be especially important when people start to travel through virtual worlds and communicate with people from different places,” says the head of the company, Mark Zuckerberg.
The Meta plan has two parts. First, the company is developing No Language Left Behind, a translation system capable of learning "any language, even if there isn't much material to learn," according to Zuckerberg. The system will know hundreds of languages - from Austrian to Ugandan and Urdu. The second project is called Universal Speech Translator; within its framework, the company is developing new approaches to translating speech from one language to another in real time. The instant translation model "will enable you to communicate with anyone in any language."
“This is a superpower that people have always dreamed of, and AI will provide it in our lifetime,” Zuckerberg promised.
The main obstacle in the development of the system is the lack of data.
“Machine translation systems typically rely on millions of annotated data,” writes Facebook AI Research in a blog post. “For this reason, systems capable of producing high-quality translations have only been developed for the few languages that dominate the Internet.”
For direct speech-to-speech, additional data is even more important, according to the Facebook AI Research team. Most speech machine translation systems use text as an intermediate step, i.e. speech in one language is first converted into text, then translated into text in the target language, and then finally input into the text-to-speech system. Because of this, speech-to-speech systems are highly dependent on text, which limits their effectiveness and makes them difficult to scale to spoken languages.
The real-time translator that Meta is working on will be able to translate even those languages that do not have standardized writing systems. This speech approach will also make translation faster and more efficient, as it does not require additional speech-to-text steps.
bbabo.Net