Google is improving voice dictation on Pixel phones in a major way.
The company today revealed an all-neural, on-device Gboard speech recognizer that promises to speed up your voice dictation. The new recognizer uses RNN transducer (RNN-T) technology that's small enough to reside on a phone, meaning there's no network latency since you're not using a network for your transcription. And since it lives on your phone, this improved speech recognizer works offline, too.
Google also notes that its improved Gboard voice dictation works at the character level, which means that as you speak, it outputs words character by character as if someone was typing out the words as you spoke in real time.
In a post on its AI blog, Google goes into much greater detail about how it developed this RNN-T technology and how it managed to get the final speech recognizer down from 450MB to just 80MB in size to help it fit on smartphones.
This new all-neural, on-device Gboard speech recognizer is now rolling out to all Pixel phones in American English. Google says that given the trends in the industry like the convergence of specialized hardware and algorithmic improvements, it hopes to expand to more languages soon.