ChatGPT may be the hot topic at the moment, but Google has been working on developing AI models for years now. One of those projects is an AI language model capable of supporting over 1,000 different languages.
And, according to a new update, Google has taken a “critical first step” in building it. Right now the model supports over 100 different languages; just 10% of what Google intends for it to do.
Google announced its plans to create the language model, which it’s dubbed the “Universal Speech Model” (USM) back in November. Apparently, the USM is a “family of state of the art speech models “ with 2 billion parameters and is trained on 12 million hours of speed and 28 billion sentences of text from over 300 languages.
Functions will include automatic speech recognition, particularly for those languages that typically don’t get the same resources as widely spoken tongues like English or Mandarin.
It’s also already in use on YouTube to generate closed captions and subtitles for videos. Although only 73 of the USM's aforementioned current 100 languages are supported on the site at present.
Those of you interested in the extreme technical details of the USM’s training regime can check them out over on Google’s blog post.
Google’s not the only company in the AI translation business right now, but the prospect of applying machine learning to translation is an exciting one. Having just got back from MWC in Barcelona, I can tell you the Spanish I learned in high school is not up to scratch. Google Translate was helpful, but not particularly convenient when you’re in the middle of a conversation.
It’s not entirely clear what Google plans to do with the USM, but the possibilities are seemingly endless. If the system can receive and translate speech in real time, whether that comes out as text or artificial speech, it could prove invaluable. More so if it’s also able to translate text quickly and efficiently.
More so if we have augmented reality glasses that can recognize and translate for us without necessarily having to be prompted first. Just as long as the AI can also recognize what not to translate. The last thing we need is background conversations constantly being translated and pumped into our ears.
But there's still a long way to go before the USM hits Google’s lofty 1,000 language goal. A hundred languages is a great start, but there are many more than that in existence around the world. So we’re just going to have to wait and see what happens.
Google I/O 2023 isn’t that far away, and with a rumored 20 AI products in the pipeline we may well hear more about them and the USM during the keynote.