Pixel users get offline voice recognition with Gboard
Google has crested something of a summit in voice recognition processing on board smartphones. Until now, if you have been using Google’s Gboard on a Pixel device, real-time voice processing for spoken to written text translation required off-board processing with either a Wi-Fi or cellular data connection for processing. Google has announced that since it first introduced neural learning to process a human voice, it has continued to develop the software architecture to the point where it can work entirely offline using the neural processing capabilities of its Pixel devices to work.
Its new recurrent neural networks transducer (RNN-T) model has been made compact enough that it can reside completely on a phone removing network latency and patchy connections as a impediments to using the technology anytime. It works in real-time outputting word letter-by-letter, just as though someone was typing out what you say as you say it. Google has been able to achieve this by trimming down the data processing requirements from 2 GB, which required a wireless connection for off-boar processing, to just 80 MB so that it can be handled on board.
Although initially on available for Gboard users with Pixel devices and using American English only, Google hopes that the research techniques that it has developed will be adopted for use in more languages and more devices and for use in other contexts as well.