ADC26: Wave Sciences - GLIMPSE

We spoke to Keith from Wave Sciences about GLIMPSE a physics-based AI system that uses computational acoustics to separate sound sources in noisy and reverberant spaces.

The system treats sound fields in a way that is compared to optical focus. It focuses on a selected source and reduces foreground or background material, addressing the "cocktail party" problem. It is software-based and works with any microphone setup that has at least two mics. More microphones, and more scattered microphone positions, improve performance.

The demo uses a smartphone with three microphones recording a speaker about five metres away, with a music loudspeaker much closer to the phone. In the workflow described, you click on a word or source in the original recording and tell the system to focus there.

This version works in post and runs faster than real time. Licensing starts now as an API, with a C engine compiled for target processors and a machine-learning API. A streaming version is planned for next year.
More Info: https://wavesciencescorp.com/news