Meta Kembangkan AI Translation For Unwritten Languages

JAKARTA - Meta is developing their Universal Speech Translator system, which in the future is expected to train artificial intelligence to translate hundreds of languages in real time.

The tech giant claims to have produced the first artificial intelligence to translate the Hokkien language, which is a language used throughout the Tawinian, spoken but not written.

"This translation system is the first milestone for the Universal Speech Translator Meta AI project, which focuses on developing AI systems that provide real-time speech-to-speech translations in all languages, even spoken languages," Meta said in a statement.

The purpose of the Meta artificial intelligence project is not only to make it easier for residents of the metaverse to interact easily, but also to translate more languages that do not have a formal writing system.

According to Mark Zuckerberg's company, there are almost half of the approximately 7,000 languages known in the world, four of the ten of which are without accompanying written components.

This unwritten language raises a unique problem for modern machine learning translation systems, as they usually need to convert verbal words into written words before translating into new languages and returning text to speech.

"(To get around this) we used a translation of the speech-to-unit (S2UT) to convert input speech to the order of the acoustic unit directly on the track previously spearheaded by Meta," Zuckerberg explained as quoted by Engadget, Thursday, October 20.

Then, we produce waveforms from the unit. In addition, the Unit is adopted for a two-time code-decomposition mechanism in which the first path decoder produces text in the associated language (Mandarin), and the second path decoder makes a unit,' he added.

Currently, the Hokkien language translator is still in process because artificial intelligence can only translate one sentence at a time, but it has been released as an open source so that other researchers can develop it.

In addition, the company also released SpeechMatrix which is a large collection of speech-to-speech translations developed through innovative Meta natural language processing devices.