Facebook owner Meta’s AI speech translation tool for Hokkien language draws interest in China
- The Facebook owner says it has developed an artificial intelligence system for translation between spoken English and Hokkien
- Hokkien is a primarily oral language widely spoken in parts of mainland China, Taiwan and Chinese diaspora communities across Southeast Asia
Existing AI-powered speech translation systems mostly focus on written languages and rely on transcriptions to train their AI models. Hokkien, however, is a primarily oral language that lacks a standard written form, according to Meta.
Another challenge is that there are far fewer Hokkien speakers than, say, English or Spanish speakers, which means there is insufficient training data available.
Meta’s speech-to-speech translation system overcame the problem by first translating English to Mandarin text, and then converting it into Hokkien, and vice versa, before adding that into training data.
When performing real-time translation, the system converts Hokkien into units of sounds and generates waveforms from them. The waveforms are then converted into Mandarin text, and eventually translated into sound units of English.
The system, which Meta said is the first of its kind, stoked curiosity among netizens in mainland China, where a video showing Meta CEO Mark Zuckerberg and AI researcher Chen Peng-Jen using the tool to talk to each other in English and Hokkien, respectively, was shared by several local media outlets on Chinese social media platforms.
“Impressive. WeChat doesn’t even support voice input in Hokkien,” wrote one user of the Tencent Holdings super app who is shown as being based in Fujian province.
Some mainland internet users also encouraged Meta to take on the bigger challenge of developing a translation system for the local vernacular of the Wenzhou city in eastern Zhejiang province, which is famously difficult for even native Chinese speakers to grasp.
Other online commenters, however, see Meta’s new translation system as a potential threat to China amid intensifying tensions with the US, fearing that the technology could become a useful tool for the US government and military.
While Meta has made its Hokkien translation system open source in the hope that others would use it for more languages, the AI model is “still a work in progress” and can translate only one full sentence at a time, the company wrote.