metamorworks - stock.adobe.com
Meta's new translation model is not quite ready for the enterprise, but the open source community could help it get there.
Facebook's parent company on Aug. 22 introduced SeamlessM4T, a multimodal and multilingual AI translation model. SeamlessM4T supports speech recognition for about 100 languages, speech-to-text translation for about 100 input and output languages, text-to-text translation for nearly 100 languages, and text-to-speech translation supporting almost 100 input languages and 35 output languages, according to Meta.
Meta released the translation model publicly to researchers and developers along with the model's metadata, named SeamlessAlign.
The releases come a year after Meta released its text-to-text translation model, which supports 200 languages.
Not ready for the enterprise
Meta's move to make SeamlessM4T open source shows that this model is not quite enterprise-grade yet, according to Mark Beccue, an analyst at Futurum Group.
"This is a research project," he said. "Enterprises aren't going to be on top of this right away. It's too early for it to be ready to go right."
Part of the reason it's too early is that AI bias is typically heavily present in translation models, needing a lot of intervention for the models to be consistently accurate.
However, universal translation could unlock revenue opportunities for Meta and its competitors in this market, Beccue said.
The tech giants stand to gain from a universal translation model mainly when it applies to e-commerce, digital goods and advertising.
In all three sectors, removing a language barrier leads to better communication, which could spur more sales.
"If you have dependable automatic universal translation, then sellers and buyers match up regardless of the language preference," Beccue said.
Meta's translation model benefits
Whether SeamlessM4T is ready for enterprise use, Meta's language translation model is still impressive because it's multimodal and multilingual, said Andy Thurai, an analyst at Constellation Research.
Mark BeccueAnalyst, Futurum Group
"Most of the current foundational models are single modal and mostly deal either with text or audio or video and images but not more than one," Thurai said. "This is one of the few models that crosses boundaries."
Moreover, SeamlessM4T won't face the same problems that other generative AI models have faced regarding originality, copyright and authenticity.
"There is no content produced," Thurai said. "It is mostly translated from the original source, so the issue is minimized." While certified translators are not in trouble yet, they may need to find ways to provide extra value to their clients, he continued.
However, fully automatic translation is not here yet, Beccue said. "That doesn't mean you can't semi-automate it or get it pretty good where these things work pretty well."
That's what Meta is looking to do: trying to make its model enterprise grade. The social media and e-commerce giant is counting on the open source community to help it get there and eventually reduce language friction in e-commerce, leading to more business.
"They're kind of saying, 'We need help,'" Beccue said. '"We're going to get people working on this, and they'll find ways that they might want to use it, and it'll benefit us.'"
Esther Ajao is a news writer covering artificial intelligence software and systems.