Tuesday, the company, Meta Platforms, which owns the Facebook platform, launched an artificial intelligence model that has the ability to translate conversations and transcribe them in dozens of languages, laying a potential building block for tools that enable real-time communication, breaking barriers between languages.
The company said in a blog post that its model (Seamless MT4) can support translation from text to speech and vice versa in nearly 100 languages, in addition to full speech-to-speech translation in more than 35 languages, combining technologies that were previously available in separate models only.
Mark Zuckerberg, CEO of Meta, said he envisions these tools facilitating interaction between users from all over the world in Metaverse, a collection of connected virtual worlds on which he is betting the company’s future. The post stated that Meta will make the template publicly available for non-commercial purposes.
The social media giant has released a handful of free AI models this year, including a giant language model called Llama that poses a serious challenge to models sold and registered to Microsoft’s OpenAI and Alphabet’s Google.
An open AI ecosystem, Zuckerberg says, is in Meta’s interest, as the company will gain more by crowdsourcing the creation of consumer-to-consumer tools on its social platforms rather than charging for the right to view and use models.
But Meta faces legal questions, like the rest of the industry, about the training data used to feed its models. Comedian Sarah Silverman and two other authors in July sued Meta and OpenAI, accusing the companies of using their books as training data without obtaining consent.
For the Seamless MT4 model, META researchers said in a paper, they collected audio training data from four million hours of “raw audio sourced from a publicly available web data repository,” without identifying any repository.
A Meta spokesperson did not respond to questions about the source of the audio data. The research paper stated that the text data came from databases established last year that used content from (Wikipedia) and related sites.