Google Announces On-Demand Speech Processing
We’ve often talked about the current limits of voice recognition (VR) software. With both Amazon Alexa and Google Assistant, we’ve reported VR error rates of around 10%. Several factors which influence this number, such as microphone quality, limited vocabulary, algorithm accuracy, VR training, etc.
Today Google announced an improvement in its VR – on-demand speech processing – that is expected to significantly reduce the processing time and bandwidth of required. Normally VR is done on servers through the cloud on behalf of connected voice devices.
What Google demonstrated today is the clever condensing of the VR software package’s size, its required footprint. Previously the VR kernel clocked in around 2Gigs, which is too large to easily host on current phones. The new condensed software is 1/25th that size, just 80 Megs.
The shrunk VR’s package fits much more comfortably on connected devices, eliminating the need to transmit voice recordings up for processing by distant servers. The software will be hosted locally on phones for immediate processing. Users can expect fewer demands on bandwidth, battery life and patience.
So long as Google respects the public’s expectations about timeout criteria, users won’t worry that their conversations are being continually analyzed and commoditized.
You can read the full article at Wired.
