C4DM / Machine Listening Seminar : Will Wilkinson

QMUL School of Electronic Engineering and Computer Science

Centre for Digital Music Seminar Series

Lecture by William J Wilkinson, QMUL

Date/time: 11am, Tuesday, 11th December 2018

Location: QMUL Graduate Centre, First Floor, room GC114 Campus map: https://www.qmul.ac.uk/media/qmul/docs/about/Mile-End_map-May2018.pdf

Open to academics, students, alumni, public; all welcome. Admission is FREE, no pre-booking required.

Title: Probabilistic models for audio signals - an intro via time-frequency analysis

Abstract: For all their proclaimed benefits, probabilistic models are yet to take off in the signal processing community. But we are inching ever closer to the point at which specifying a probability distribution over an audio signal will be practical and useful. In signal processing we are interested in modelling functions of time, and Gaussian processes offer a principled way to specify distributions over functions whilst also encoding prior knowledge. In this seminar I’ll demonstrate how Gaussian processes provide a means through which to interpret existing models for time-frequency analysis, and to construct new models. The probabilistic framework allows us to adapt our time-frequency representation to a signal, quantify uncertainty about predictions, and generate new audio signals from the learnt distribution that maintain the correct amplitude and phase relationship.

Author info: Will Wilkinson is a PhD student in the Machine Listening Lab at the Centre for Digital Music, QMUL. He is interested in probabilistic models for natural sounds. See Will's research page for more about his work.