Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory announced last week the development of a new, low-power microchip designed specifically for processing continuous automatic speech recognition in automobiles, robots, Internet of Things (IoT) and wearable devices. The key feature of these new chips is their ultra-low power consumption. They use only 0.2 to 0.10 milliwatts of power for speech recognition tasks compared to the 1 watt of power currently necessary to process the same data in devices such as cellular phones, resulting in a power savings of up to 95 percent.
The MIT research team was able to achieve significant efficiencies in microchip power consumption in a few ways. One method was to include 3 separate “voice activity monitoring” circuits, each with differing degrees of complexity and assigned to different tasks. One of the more simple circuits monitors the ambient noise in the area to determine if it is actual speech. If it detects speech, a larger and more complex circuit engages for further processing. In addition, the MIT researchers focused on compressing the data that passes through the circuits for more efficient processing, as well as storing and processing speech recognition data locally on-chip. This minimizes the amount of data that has to be processed and retrieved from the cloud, which requires much more power.
This innovation in low-power microchip technology has huge implications for many IoT and wearable devices that are required to harvest their energy from the environment, or have to endure a long duration with no access to direct power. Currently, even the most energy-efficient components in use would still drain a device’s batteries quickly if voice recognition software was running continuously without interruption. However, by deploying MIT’s new low-power circuit devices could now have continuous and uninterrupted automatic speech detection and processing taking place. Eventually, most IoT devices such as sensors will be deployed with long-lasting, pre-charged batteries with no intention of ever recharging them and no need for an external power source. These devices could have a service life of five to ten years, after which time they are simply replaced. In a situation such as this, MIT’s low-power microchip would be extremely valuable.
Making embedded speech recognition ubiquitous in a wide-range of electronic devices is a serious goal of this MIT project, so this evolution in microchip technology was necessary to help realize that goal in a timelier manner. This development also follows current trends in technology overall that include embedded voice recognition in many consumer products, as well as the miniaturization of everything from IoT sensors and controllers, to consumer wearables like watches, fitness trackers, and eyeglasses. With many of these devices having form factors too small to connect to a keyboard or external display, the next logical and natural way to interface with them will be through voice commands and speech recognition software.
To Learn More: