Meta develops AI translation system for a primarily spoken language

The new AI translator can convert spoken languages like Hokkien into spoken English, helping break down language barriers

Meta's AI translation system can translate spoken languages like Hokkien

Meta has developed a new AI translator that can convert spoken languages like Hokkien into spoken English in an attempt to knock down language barriers. Hokkien, a dialect of southern Min Chinese, is a primarily spoken language and does not have a standard writing system, which makes building translation tools for it a huge challenge.

The open-source translation system is a part of Meta’s Universal Speech Translator (UST) project and has made a breakthrough in this challenge. The company, formerly known as Facebook, hopes that this along with the other AI methods under development will eventually allow real-time speech-to-speech translation across hundreds of languages – even the spoken kind.

Languages like Hokkien are difficult to translate because machine translation tools require vast amounts of written text to train on, and such languages do not have a widely used writing system. To ease this problem, Meta used Mandarin – another Chinese language but with an ample supply of readily available training data – as an intermediary between English and Hokkien.

Also Read | Inside Zuckerberg’s metaverse struggles

Researchers from the project also actively worked with native Hokkien speakers to verify the accuracy of the AI translation models. Of course, since a speech-to-text system was not available, Meta focused primarily on speech-to-speech translation to build the translation system.

While the model is still a work in progress, it does allow someone who speaks Hokkien to converse with someone who speaks English. The only catch is that it can translate only one full sentence at a time in its current state. Meta is encouraging other developers to jump on the bandwagon by releasing technologies like SpeechMatrix to assist them in creating their own speech-to-speech translation systems or build on its work. The company has also open-sourced its Hokkien translation models along with the research papers associated with them.