Meta Is Training Its AI on the Bible and Other Religious Texts

The mother or father firm of Fb and Instagram says it has developed an AI-powered text-to-speech expertise that may determine over 4,000 languages. The purpose, Meta says, is to protect the world’s languages, and the tech large is utilizing the Bible and different spiritual texts to do it.

“Gathering audio knowledge for hundreds of languages was our first problem as a result of the biggest current speech datasets cowl 100 languages at most,” Meta stated in a submit asserting the venture. “To beat this, we turned to spiritual texts, such because the Bible, which have been translated in many various languages and whose translations have been broadly studied for text-based language translation analysis.”

In an accompanying analysis paper by the Meta AI core crew, the corporate says it obtained its knowledge from the Bible, together with authentic textual content and audio recordings from FaithComesByHearing.com, GoTo.Bible, and Bible.com.

The venture contains recordings of Bible tales, evangelistic messages, scripture readings, and songs in additional than 6,255 languages and dialects. Whereas most recordings function had been typically by male readers, the Meta says its fashions work equally properly for feminine voices.

A dataset of readings of the New Testomony, Meta stated, supplied greater than 1,100 languages that supplied 32 hours of knowledge per language on common.

Based on Broward School’s Lingua Language Middle, there are over 7,100 dwelling languages worldwide.

“Our consultations with Christian ethicists concluded that the majority Christians wouldn’t regard the New Testomony, and translations thereof, as too sacred for use in machine studying,” the Meta AI crew stated, including that the identical shouldn’t be true for all spiritual texts.

“There’s additionally the chance of non secular coaching knowledge biasing the fashions with respect to a specific world view,” Meta AI stated. “Nonetheless, our evaluation of the language generated by our fashions means that the language produced by the ensuing speech recognition fashions exhibit solely little bias in comparison with baseline fashions skilled on different domains.”

After its metaverse ambitions fizzled earlier this 12 months, Meta seems to have shifted its focus to synthetic intelligence, together with constructing an AI device to determine and separate objects in photos and an AI-powered device to assist manufacturers goal customers on its Fb and Instagram platforms.

Whereas the expertise continues to be in its early phases, Meta says it’s open-sourcing its knowledge and code in order that others can construct on, develop, and enhance the platform.

“Lots of the world’s languages are in peril of disappearing, and the constraints of present speech recognition and era expertise will solely speed up this pattern,” Meta stated. “We wish to make it simpler for folks to entry data and use units of their most popular language, and at present we’re asserting a collection of synthetic intelligence fashions that might assist them do exactly that.”