Facebook-owned Meta Platforms on Tuesday unveiled an artificial intelligence model capable of translating and transcribing conversations in dozens of languages, setting a potential building block for tools that enable real-time communication such as cross-language family barriers.
The company said in a blog post that its model (Seamless MT4) can support text-to-speech and vice versa in over 100 languages in addition to full speech-to-speech translation in more than 35 languages. Available on select models only. Meta’s CEO, Mark Zuckerberg, said he envisions these tools facilitating interactions between users from around the world in the Metaverse, a group of connected virtual worlds.
The post said it would make the meta template publicly available for 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 at Microsoft-owned OpenAI and Alphabet-owned Google. Zuckerberg says the open AI ecosystem is in Meta’s interest because the company makes more money by building consumer-to-consumer tools on its social platforms than by charging to view and use models.
But Meta, like the rest of the industry, faces legal questions about the training data used to feed its models.
For the unconstrained MT4 model, the META researchers, in a paper, collected four million hours of audio training data “from raw audio obtained from a publicly available Internet data repository,” without identifying any commodities.
A Meta spokesperson did not respond to questions about the source of the audio data.
The research paper said the textual data came from databases established last year that used content from Wikipedia and related sites. (Al Arabiya Net)
“Award-winning beer geek. Extreme coffeeaholic. Introvert. Avid travel specialist. Hipster-friendly communicator.”
