[ad_1]
For several versions, Gboard for Android is working on a "faster voice input" that works offline. Today, Google formalizes this feature on Pixel phones and details the "end-to-end, fully neural, integrated voice recognition system" that it has created.
Google notes that a "revolution" in speech recognition began in 2012 with significant improvements in accuracy through deep learning. The main goal of various architectures was to reduce the time needed to transcribe a user's speech, or latency. Google notes how "an automated assistant feels much more useful when he responds quickly to requests."
Google's latest development is an end-to-end, fully neural integrated Gboard voice recognition system when users press the microphone icon in the upper right-hand corner. It runs completely offline and is only 85 MB in size, compared to previous models which were 2 GB and 450 MB later.
This means there is no more latency or wasted time on the network – the new discovery system is always available even when you're offline. The model works at the character level. So when you speak, it displays the words character by character, as if someone were typing what you say in real time and exactly as you wanted with a keyboard dictation system. .
<img title = "Google rolls out Gboard speech transcript in real time, on device, for Pixel phones" class = "aligncenter wp-image-276389 size-large" src = "https://9to5google.com/wp- content / uploads / sites / 4/2019/03 / Pixel-Gboard-new-speech-transcription.gif? w = 1000 "alt =" Pixel Gboard New Speech Transcript” width=”1000″ height=”563″/>
In addition to offline access, this new system generates one character at a time, instead of one word at a time.
The RNN-T recognition module generates characters one by one, as you speak, with white space where appropriate. To do this, it uses a feedback loop that returns the symbols predicted by the model to predict the following symbols, as described in the figure below.
Today's advancement is due to the merger of several components of the voice recognition system. A single neural network "directly maps an input audio waveform to an output phrase."
The new device voice input system is currently being deployed on Pixel, Pixel 2 and Pixel 3 phones in US English. To turn it on, go to Gboard Settings> Voice Recording> Faster Voice Recording. Google expects this to translate into more languages and later on in other cases of use.
Check out 9to5Google on YouTube for more information:
[ad_2]
Source link