The problem with pretty much all voice assistant software and voice recognition applications is that they need a connection to a server in order to process your request. Sometimes we have to wonder if this is completely necessary, especially when there are certain requests that are small and simple enough where a connection isn’t really required.
Turns out Google has been exploring a similar idea as well, and the company has recently developed an offline voice recognition system (via 9to5Google) that is small and lightweight enough where it can be run on a 3-year old Nexus 5 handset, and weighs in about 20.3MB. To top things off, this system is said to be 7x faster than current systems and only has a 13.5% word error rate.
How did Google achieve this? By sampling 3 million anonymous voices taken from Google search. The voices were then further distorted into 20 different variations by extracting noise from YouTube, thus allowing the system to recognize different voices even better. Google also relied on a single model for the dictation and voice commands.
Now to be fair, Google Now does come with some offline commands albeit limited. With this system, Google will be able to help speed things up which we reckon would be good for markets with limited internet capabilities, or just in terms of overall efficiency. Google also notes in their paper that this system could be used beyond phones, and that it could find its way into wearables as well. Whether or not Google has plans to launch such a system remains to be seen, but we wouldn’t mind seeing it make its way onto Android.