Today during I/O 2019, Google unveiled its next-generation Assistant that’s been, essentially, rebuilt from scratch. Somehow, someway, Google’s engineers were able to smush the amount of required data to process requests to just 0.5GB. That’s down from 100GB. Now, all of the required information to process requests you make can live on your phone.
Because everything’s hardware-based and so efficient, with the next-gen Assistant, Google says it’ll process requests up to 10 times faster than it can now. No network connection is required and saying “Hey Google” before every request is no longer needed. It’s a really amazing achievement that should improve the overall experiece of using the Assistant on your phone.
To power the Google Assistant, we rely on the full computing power of our data centers to support speech transcription and language understanding models. We challenged ourselves to re-invent these models, making them light enough to run on a phone.
Today, we’ve reached a new milestone. Building upon advancements in recurrent neural networks, we developed completely new speech recognition and language understanding models, bringing 100GB of models in the cloud down to less than half a gigabyte. With these new models, the AI that powers the Assistant can now run locally on your phone. This breakthrough enabled us to create a next generation Assistant that processes speech on-device at nearly zero latency, with transcription that happens in real-time, even when you have no network connection.
Google says the next-gen Assistant will debut first on the Pixel 4 this fall.