Have you been using Google Now to manage your Android-based smartphone? It’s like the iPhone’s Siri. You can give basic commands without having to type like “Play music” or “Open Gmail” or “Send email” and such. Now Google has created an off-line speech recognition system for Google Now that is much faster and more accurate compared to its online version. This is the PDF that explains the technology in detail. Here is a relatively older video that explains what Google Now achieves for you:
When you are giving commands in Google Now, your smartphone needs to be connected to the remote server where all the commands are processed and executed. This slows down the entire process. Besides, no Internet connection means no voice control. Google has now created a 20.3 MB system that can be downloaded on your smartphone. Once installed and in use, it makes your Google Now experience 7x times faster (without an Internet connection) and has only 13.5 word error rate.
It has been tested on a Nexus 5 with a quad core 2.26GHz processor and 2 GB RAM.
The system has gone through rigourous testing. In order to reduce the size of the program the system uses a single model for dictation and voice commands.
The system was exposed to 3 million anonymous voice samples from Google search. Every sample had 20 distorted versions created by extracting noise from YouTube videos. The PDF paper linked above says that the system can be implemented on smartphones as well as smart wearable devices like smart watches.