Using the state-of-the-art BERT technology, or Bidirectional Encoder Representations from Transformers – a pre-training model for natural language processing that was open-sourced in 2018 – Google is looking to significantly improve the voice recognition and continued conversation technology in Nest speakers and smart displays. To spare you from all of the geeky details, which you can read on the AI Google Blog, the gist is that Assistant is about to get a lot less contextually ignorant. These updates are now available for alarms and timers on Google smart speakers in English in the U.S. and expanding to phones and smart displays soon.
Teaching Assistant unique names has always been a challenge. Over the next few days, you can teach it to enunciate and recognize the names of your contacts and how you pronounce them. Assistant will remember this without keeping a voice recording and be able to intelligently utilize it the next time you ask it to call or text someone. This will first be available in English and then will expand to other languages shortly after.
The most significant improvements to Assistant using the BERT technology come in the form of its ability to recognize context. You can see all of this in the video below, and it’s pretty darn cool. Currently, setting two separate timers for different things means that Google won’t know which one to cancel should you ask it to. With new context awareness, it will be able to recognize timers and alarms based on their list order (i.e. first or second timer), or their names (i.e. “boiled potatoes timer”, “homework timer”, etc.)
Lastly, BERT is being applied to Assistant to better help it understand how to respond to follow-up questions you ask, meaning that back-and-forth conversations will feel more natural. This means that you’ll spend less time re-stating the subject or topic of your conversation and more time speaking fluidly as you would with another human.
This is huge, but it feels as though it should have been able to do this all along, right? These improvements mean that Assistant can respond nearly 100% accurately to these types of tasks now. In the near future, Google will be bringing these improvements to other use cases as well, and I can’t express how happy this makes me. All too often I’m frustrated by Assistant’s lack of common sense, so this should prove to be the one thing that transforms that.