3D-printed ‘cocktail party listener’ picks out voices in a crowd
While humans are uniquely able to selectively listen to individual voices in a group of multiple speakers, accomplishing the same feat with computers has been an ongoing challenge, with only partial success from the use of artificial intelligence or multiple microphones. However the new device – and associated algorithm – developed by the Duke researchers is claimed to be 96.7% accurate when run with three overlapping sound sources.
The device itself is a 3D-printed acoustic metamaterial disc, about the size and shape of a pizza. It comprises 36 honeycombed "pie slices," or wedges, with openings around the edges that channel sounds to a microphone located in the center of the device.
Each of the channels – or ‘waveguides’ – in the device is designed to acoustically filter the sound that passes through it in a certain way. Once the resulting combined sound is picked up by the center microphone and sent to a computer, an algorithm is used to then separate and identify – or reconstruct – the multiple individual sounds that originated from different sources and directions.
The researchers are currently exploring ways to shrink the device by a factor of ten or more to make it practical for more applications. The approach is seen as having potential uses in speech recognition and hearing aids, as well as other acoustic imaging and sensing applications. For more, see the paper published in the Proceedings of the National Academy of Sciences: Single-sensor multispeaker listening with acoustic metamaterials.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
