http://www.digitaljournal.com/technology/google-building-7x-faster-speech-recognition-that-works-offline/article/460133

Google's new speech recognition works offline, powered by a phone

Posted Mar 14, 2016 by James Walker
Google is building a speech recognition system that can run on a smartphone even while it is offline. The company has substantially trimmed down its speech libraries to maintain accuracy while running on a phone's limited processor and memory.
Cortana on an Android phone
Cortana on an Android phone
Microsoft
Voice recognition systems such as Siri, Google Now and Cortana are becoming increasingly more mature, giving smartphone users a quick and easy way to interact with their device while on the go or in the car. They all require an Internet connection though, making them unusable without a reliable network and data plan.
The app installed on the phone is merely a portal to the speech services running on the servers of Google, Apple or Microsoft. The phone collects the microphone input and sends it off for processing and analysis. A few moments later, the server sends back the words that the user spoke, giving the phone what it needs to respond to the command.
The algorithms needed to work out what the user said are too complex to run on a phone's limited hardware. The process can be sped up by offloading the number crunching to Google's powerful servers, making the system faster and saving space on the phone. This leads to a smoother voice recognition process, except for when it doesn't work because there is no Internet connection available.
Google has developed a solution. In a recent research paper, it explained how it has been able to condense its voice recognition algorithms so that they fit onto a smartphone, work reliably and don’t consume huge amounts of power. The result is a version of Google Now's voice analysis system that is 10 times smaller than the one running on the company's servers today.
Installed on a Google Nexus 5 test device, the system ran seven times faster than the Internet-connected original. It was trained by being exposed to 3 million anonymous voice samples sourced from Google Search.
The researchers observed a word error rate of 13.5 percent, a 5 percent increase from the error rate of the servers. There is room for improvement but the results clearly validated the proof-of-concept design, showing smartphones could soon take responsibility for processing their own voice commands.
In the near future, both systems could be combined to obtain the benefits of each one. Adding the offline voice recognition algorithms to Android would give smartphone owners a way to keep using voice recognition without a connection, albeit with decreased accuracy.
At home, the traditional approach of connecting to Google's servers could be used instead. When available, the increased speed and reliability of the server-side algorithms would be favoured. The offline alternative could be used as required when no stable connection is available.
The offline version of Google Now described in the research paper is feature-complete. Aside from a slimmed-down dictionary and the loss of some accuracy, it is capable of supporting all the features of the online algorithms. It has voice customisation and error detection, all running on the limitations of the phone's processor. It may be a little way off yet but offline voice recognition looks likely to appear in Android at some point in the future, making Google Now a more versatile assistant that is helpful outdoors as well as in the office.