Meta Unveils a New AI-Powered Hokkien Speech Translation System. As part of its Universal Speech Translator (UST) project, Meta stated that it had created the first AI-powered translation system for Hokkien, a language that is primarily spoken.

The innovation is especially significant because, up until now, machine translation technologies were only available for written languages. Until recently, the norm for training AI models was to feed them massive amounts of written material. However, researchers may now use various sources, including written and spoken data, to build a translation system for unwritten languages.

AI translation has concentrated chiefly on written languages until today. Yet almost half of the world's 7,000-plus surviving languages are primarily oral, lacking a standard or widely employed writing system. Consequently, it is hard to create machine translation tools using typical techniques that require massive amounts of text to train an AI model.

To solve this issue, Meta has developed the first AI-powered translation system for Hokkien, which is widely spoken throughout the Chinese diaspora. Due to the lack of a standard written form for Hokkien, Meta's technology allows Hokkien speakers to converse with English speakers.

The Universal Speech Translator (UST) project, which is part of Meta, is developing new AI techniques that will eventually enable real-time speech-to-speech translation across all existing languages, including those that are primarily spoken. The open-sourced AI translation system is a component of UST. The business holds that spoken communication can assist remove obstacles and unite individuals no matter where they are, including in the metaverse.

To construct the new system, Meta's AI researchers had to overcome numerous complicated obstacles conventional machine translation systems posed, such as model design, data collection, and evaluation. The blog entry reads, "We have a lot of work to do to expand UST to additional languages. However, the capacity to communicate fluently in any language has been a long-sought goal, and we're happy to be one step closer to realising it. Furthermore, we are open-sourcing our Hokkien translation models, assessment datasets, and research papers so that others can repeat and improve our work.

In addition, we can apply the concepts to numerous other written and unwritten languages. Meta also provides SpeechMatrix, a massive corpus of speech-to-speech translations extracted using the LASER technique for data mining. Then, researchers will be able to design their speech-to-speech translation (S2ST) systems based on the work of Meta.