Voice recognition devices, such as Alexa and Amazon Echo, are cool devices to own at home. However, the energy they consume isn't cool at all, especially to the pockets. MIT researchers understand this need and decided to find a solution. The result of that quest was a voice recognition chip that guarantees less energy consumption by these smart devices.

MIT researchers collaborated with Quanta Computer, a Taiwan-based electronic and computer manufacturer, and created a voice recognition chip that is 90 to 99 more energy efficient than what is available on the market now.

ESource tested three voice recognition devices - Amazon Echo, Amazon Echo Dot, and Google Home - and found out that even when these devices are in "Idle Mode," they can consume up to 3 watts and up to 5 watts if they are in "Listening Mode." On the other hand, the new chip designed by MIT researchers only requires 0.2 to 10 milliwatts of power.

A Technological Breakthrough

Although the existing speech recognition software might look small, it can still drain the battery of any device if it runs non-stop. In order to solve this energy problem, the researchers had to find a more efficient bandwidth management.

A chip's memory bandwidth is measured by how often data is sent to the different nodes communicating inside a neural network. These nodes have 'weights' associated with them based on how often data is transmitted through them. That means the more data that passes through, the heavier the node becomes.

The new MIT-designed voice recognition chip compresses these weights and decompresses them once they're brought on-chip.

Aside from compressing the weights, the chip also brings in a single neural network at a time, but data passes through it from "32 consecutive 10-second increments." These 32 passes produce 384 output values which are locally stored in the chip. Then, each of these values is paired with 11 more values that go to the next layer of the node. This process is repeated from layer to layer.

The new voice recognition chip is not only energy efficient but will also miniaturize speech recognition devices more.