Meta releases AI model for translating speech between dozens of languages

Facebook mum or dad firm Meta Platforms on Tuesday launched an AI mannequin able to translating and transcribing speech in dozens of languages, a possible building-block for instruments enabling real-time communication throughout language divides.

The firm mentioned in a weblog publish that its SeamlessM4T mannequin may help translations between textual content and speech in almost 100 languages, in addition to full speech-to-speech translation for 35 languages, combining know-how that was beforehand accessible solely in separate fashions.

CEO Mark Zuckerberg has mentioned he envisions such instruments facilitating interactions between customers from across the globe within the metaverse, the set of interconnected digital worlds on which he’s betting the corporate’s future.

Meta is making the mannequin accessible to the general public for non-commercial use, the weblog publish mentioned.

The world’s largest social media firm has launched a flurry of largely free AI fashions this yr, together with a big language mannequin referred to as Llama that poses a severe problem to proprietary fashions offered by Microsoft-backed OpenAI and Alphabet’s Google.

Zuckerberg says an open AI ecosystem works to Meta’s benefit, as the corporate has extra to realize by successfully crowd-sourcing the creation of consumer-facing instruments for its social platforms than by charging for entry to the fashions.

Nonetheless, Meta faces related authorized questions as the remainder of the trade across the coaching information ingested to create its fashions. In July, comic Sarah Silverman and two different authors filed copyright infringement lawsuits towards each Meta and OpenAI, accusing the businesses of utilizing their books as coaching information with out permission.

For the SeamlessM4T mannequin, Meta researchers mentioned in a analysis paper that they gathered audio coaching information from 4 million hours of “raw audio originating from a publicly available repository of crawled web data,” with out specifying which repository.

A Meta spokesperson didn’t reply to questions on the provenance of the audio information.

Text information got here from datasets created final yr that pulled content material from Wikipedia and related web sites, the analysis paper mentioned.

Stay on high of know-how and startup news that issues. Subscribe to our each day e-newsletter for the most recent and must-read tech news, delivered straight to your inbox.