More polyglot than ever. Google shared on Tuesday, March 7, its first research elements for the construction of artificial intelligence, which would eventually be capable of translating its services into 1,000 languages. Currently, its online services like Google Translate are only available in 130 languages while in the world it lists more than 7000 languages.
This artificial intelligence project, presented for the first time last novemberhas been dubbed the “universal speech model” (for universal speech model, in English). Its particularity: being trained in both oral and written language.
Since then, the model has been trained in “12 million hours of speaking and 28 billion written sentences, covering more than 300 languages,” Google details in a progress report shared on its blog dedicated to its research projects. Artificial intelligence is especially used to create legends on YouTube, the group’s video platform.
hard to find data
The model can already perform automatic speech recognition in around 100 languages, whether they are widely spoken, such as English or Mandarin, or whether they are dialects. For some languages spoken by fewer than 20 million people, finding the data needed to train artificial intelligence is “very difficult,” Google says.
The American company is optimistic: “Our model learning process is effective in adapting to new languages and new data.”
Last November, Google had already said that the project would take several years. It will gradually be deployed in the different services of the technological giant, such as its digital keyboards or Google Translate.
Source: BFM TV
