Audio coding, generative modeling, and perception

Abstract:
The goal of audio coding is a bit boring. The best you can hope for is to create a sound identical to the reference sound. Generative modeling, on the other hand, can create unlimited amounts of highly realistic audio signals. In its purest form this tool will give you a different sound every time you ask, so why is it so popular in recent audio coding research? To explain this, I’ll give a few examples of its usage from our recent work, but also discuss how machine learning can be used to progress our understanding of one of the most basic test questions related to human auditory perception: are these two sounds different? Having means to reliably predict the answer would be all but boring

Speaker bio:
Lars Villemoes has a docent degree in mathematics from KTH and is currently a Principal Member of Research Staff at Dolby Sweden AB. His main research interests are in the application of machine learning, time-frequency analysis, and auditory modeling to audio signal processing.

Local host: Anders Friberg