A Sound Decision

Speech recognition has emerged as one of the most natural and efficient ways to interact with computing devices in recent years. Not that this is a new idea, of course — we have been dreaming of talking to our computers since long before it was popularized in science fiction like the Star Trek television series. But the technology required to enable general speech recognition has only become practical recently.

It may not — and may never — completely take over as the primary way that we interact with our computers, but it has proven to be very valuable in certain use cases. Mobile applications, in particular, benefit from speech recognition. Being able to issue a voice command while on the go, or when occupied with other tasks like driving, can boost efficiency and even enhance safety.

But whenever one discusses mobile applications, issues related to energy consumption must also be considered. This is most definitely an important factor in speech recognition, as this technology is typically powered by artificial intelligence algorithms that can be very power-hungry. This is especially true since many applications require always-on listening, which can quickly drain batteries.

The cost and maintenance associated with continually recharging and replacing batteries grows quickly, rendering speech recognition impractical for many mobile use cases. Furthermore, the environmental impact of all of the associated e-waste is a serious concern. In the future, these concerns may be minimized, thanks to the work of a team of researchers at ETH Zurich in Switzerland. They have approached the problem from a completely different angle and developed a sensor that can recognize speech without an onboard source of power. Rather, the sensor is powered by the very sound waves that it is designed to detect.

The team’s device consists of a two-dimensional metamaterial arranged in a 7 x 7 lattice of cells. The material itself is not particularly important, but rather the structure of it is. These cells, made of silicone, are constructed with very specific geometries, and are connected to one another via small connecting bars that act like springs. This arrangement causes the entire structure to vibrate when it is struck by sound waves. That vibration can be converted to electricity, which in turn can power electronic components.

This is not just an energy harvester, however. The vibrations only occur under very specific conditions, which allow the structure of the device itself to perform speech recognition. A sensor can be tuned to vibrate when it encounters a sound wave produced by an individual saying the number “four,” for example, but not any other number. Once the metamaterial recognizes that keyword, it will also use the vibrations to generate enough power to report that fact to an external device.

As you might expect, finely-tuning these materials to respond to specific utterances is not easy. Accordingly, the researchers leveraged computer modelling and special algorithms to aid them in generating the microstructure designs.

Present prototypes built by the team are only capable of distinguishing a single word, but they believe that a new version should be able to recognize at least a dozen words. They are also working to shrink down the sensor, which is currently about the size of a palm.

With these enhancements, the researchers anticipate their sensors being deployed for many applications, including those in monitoring infrastructure and medical implants. Aside from not requiring an onboard source of energy, they are also environmentally-friendly, which is highly desirable when discussing large-scale deployments of sensor networks.