The voice dictation module of the Android Gboard keyboard now works without latency and without the need for a data connection. To achieve this, Google has managed to implement a very compact neural network system on mobiles.
Google has just updated its virtual keyboard for Android mobile and including its voice dictation function. The big news is that rather than exploiting its remote servers for voice recognition and transcription, the Internet giant implements this ability directly in the phone. No need to have a network connection to perform the operation. The Gboard now integrates a neural network system capable of performing the operation.
The result of the demonstration is bluffing since there is more latency. The words are displayed character by character at the same time as the dictation. For the moment, this option is available only for the English-American keyboard and only on Pixel smartphones. Originally, to be able to analyze and transcribe voice dictation, the system occupied more than 2 GB of capacity. Google has managed to reduce this neural network to 450 MB, then compact it so that it does not exceed 80 MB!
Google, new apostle of data privacy
Casually, this example marks a new philosophy of Google for Android and mobiles. Rather than exploit remote servers, the arrival of embedded speed processors in smartphones, even positioned at the entry level, allows now to deal with operations using artificial intelligence directly on the mobile.
For Google, embedding the AI on the smartphone solves the problem of confidentiality of the data, since they will not transmit on its servers. This is why the firm multiplies this kind of practice on its various tools. This is already the case with the automatic responses implemented in the Gmail application. These answers are generated directly by the algorithms on the smartphone and not by the servers.