Google set up an autonomous speech recognition system, which is faster and more accurate than comparable systems that are connected to the Internet. While many counterparts and remain on paper, the new system has already been launched and tested on the Nexus 5.
Currently, Google App has a very limited capacity offline. More complex commands to be sent to the server, and where to undergo treatment. This leads to errors because of unreliable networks.
An alternative is to “build a speech recognition system which operates locally on the mobile device. However, such a system could be inaccurate, and can consume a significant amount of memory and other resources.
Through various methods of machine learning, the Google created the system volume of 20.3 MB, which is 7 times faster than systems that are connected to the Internet, and the error rate was only 13.5%. It has been implemented and tested Nexus 5 with quad-core processor at a frequency of 2.26 GHz and 2 GB of RAM.
To achieve this size and saving of resources, the system uses a single model for dictation and voice commands. Additional compression techniques have been used. The system was trained by exposure to 3 million anonymous voice samples (approximately 2000 hours). Each speech sample as a distorted version 20 created by extracting the noise from a YouTube video.
The document notes that such a system is not limited to smartphones and wearable devices can be used. Let’s hope that these improvements will appear on smartphones in the near future.