Meta Platforms Releases Model AI For Translation And Speech Transcription In Many Languages

Meta Platforms, released an AI model that can translate and write speeches in dozens of languages. (photo: twitter @meta)

JAKARTA - Meta Platforms, Facebook's parent company, has released an AI model capable of translating and writing speeches in dozens of languages. This AI model can form the basis for the development of communication tools in real-time across language differences.

In a blog post, the company announced a SeamlessM4T model that could support translations between text and speech in nearly 100 languages, as well as full speech-to-pidato translations for 35 languages, combining previously available technologies in separate models.

Meta CEO Mark Zuckerberg has stated that he envisions tools like this will facilitate interactions between users from around the world in the metaverse, a network of interconnected cyberspaces where he is betting on the company's future. Meta makes this model available for public use without commercial purposes.

The world's largest social media company has released a number of AI models, mostly for free, this year, including a large language model called Llama which is a serious challenge for exclusive models sold by Alphabet's Microsoft-backed OpenAI and Alphabet's Google.

Zuckerberg said that the open AI ecosystem provides benefits for Meta, as the company has more advantages by gathering contributions in the creation of consumer tools for its social platforms than charging access fees to those models.

However, Meta faces legal questions similar to the entire industry regarding the training data used to create its models.

In July, comedian Sarah Silverman and two other writers filed a copyright infringement lawsuit against Meta and OpenAI, on charges that the companies used their books as unlicensed training data.

For the SeamlessM4T model, Meta researchers collected audio training data from 4 million hours of "treal audio coming from a publicly accessible web data repository," without specifying which repositorys were used.

The text data comes from a dataset created last year that took content from Wikipedia and related websites, as written in their research papers.