Time.news – It is the first step towards a speech translation system of an unwritten language powered by Artificial Intelligence and it was done with Hokkien, one of the official languages of Taiwan, spoken by Chinese immigrants but lacking a standard written form. The system was presented today by Meta.
“This technology – says Palo Alto – allows those who speak Hokkien to have conversations with people who speak English. To develop this new speech-only translation system, Meta’s AI researchers had to overcome many challenges posed by traditional machine translation systemssuch as data collection, model design and evaluation ”.
The goal of the Meta researchers is to make “the Hokkien translation system part of a universal voice translator”: an open-source model in which “the code and data used to train the AI community” will allow “other researchers to carry out this work”.
In its current stage, according to the researchers, this approach “allows a person who speaks Hokkien to speak to an English-speaking person. Although the model is still in development and we can only translate an entire sentence at a time, it is a step towards a future where simultaneous translation between languages will be possible”.
… Even in the metaverse
Meta is committed to the development of an Artificial Intelligence system capable of providing real-time speech to speech translation in all languages: even those mainly spoken. The project is called Universal Speech Translator and was presented in February by Meta.
Meta
“Until now – explained Mark Zuckerberg’s company – the translations carried out by Artificial Intelligence have mainly focused on written languages. Yet more than 40% of the more than 7,000 languages currently in use are mainly oral and do not have a standard written form or a widely adopted writing system ”. Among the objectives of the initiative “to help remove barriers and bring people closer wherever they are, even in the Metaverse”.
Data collection
“Gathering enough data was a major obstacle that faced us when creating the translation system for Hokkien”, a language with little resources, ie there is “no range of data available to train the system”.
To overcome this obstacle, the researchers explain that “they have exploited Mandarin as an intermediate language: we have translated a vocal content into English (or hokkien) into a text in Mandarin and then we translated it into Hokkien (or English) and added to the data for model training. This method has greatly improved the performance of the model by exploiting data from similar languages with more resources available ”.
Data mining
The techniques developed for Hokkien “can be extended to many other languages, both written and oral. For this purpose, we are about to release the oral matrix, a vast corpus of speech-to-speech translations developed with Meta’s innovative data mining technique, called LASER, which will allow researchers to create your own speech-to-speech translation systems and rely on our work to progress ”.