Qualcomm on-device speech recognition reaches 95 percent accuracy
Qualcomm’s on-device voice recognition system now recognizes words and phrases with 95 percent accuracy, a company representative told an audience at the Re-Work Deep Learning Summit in Boston, according to VentureBeat.
Aritifical intelligence researcher Chris Lott said that the system consists of both a convolutional neural network (CNN), and a recurrent neural network (RNN), the latter of which uses its memory to process inputs. The method contrasts with most current speech recognition technologies, in which only key phrases are recognized by the device, with the remainder recorded and transmitted to the cloud for processing. In addition to the privacy benefits of having speech processed on the device which records it, rather than in a company cloud, Lott said the system provides much faster feedback, and works offline.
“There’s a push to do the whole end-to-end system in some neural net fashion,” Lott said. “It’s something that’s going to make interacting with devices more natural.”
VentureBeat points out that Google’s offline speech recognition system revealed in 2016 was 7 times faster than its online system, and delivered 86.5 percent accuracy while running the 20.3-megabyte software on a smartphone.
Qualcomm chips are already found in many smartphones, and the company recently implemented the forthcoming Android P with its Snapdragon Mobile Platforms, including the Snapdragon 845, which features a secure processing unit (SPU) intended to secure biometric transactions.