1,000 Languages

Generated by author using Midjourney

Last year, Google announced the 1,000 language recognition / translation initiative. More on it is here:

Universal Speech Model (USM): State-of-the-art speech AI for 100+ languages

Currently, they’ve managed to release support for 100 languages for speech transcription. That’s quite a fete.

Google has access to a significant amount of content for voice through YouTube videos. It’s been using this source to mine training sets for all sorts of models, including speech recognition.

Now the question is what will come when 1,000 languages are transcribed? Will we have a universal translator like Star Trek? Will we have the equivalent of a deep fake to dub our voice with a synthetic version of our own in another language?

We are getting pretty close to these technologies being possible. However, the most exciting outcome might be customized training to learn new languages based on our current language ability and large language models.