Institute of Information Science, Academia Sinica

Digital filters are essential building blocks in signal processing and many audio applications, ranging from linear prediction voice synthesis, audio effects, to system identification. With the introduction of deep learning and end-to-end optimisation, there has been growing interest in optimising these interpretable structures using gradient-based methods so they can be combined with data-driven approaches, such as neural network training. In this talk, we will start with a brief introduction of feedforward and recurrent filters and their differentiable use cases. We will then look into problems that have arisen in the literature, such as slow inference in automatic differentiation frameworks and the challenge of modelling parameter-varying systems, and present some solutions. The talk will conclude with a forecast of possible steps we could take, the unsolved challenges in this field, and where the low-hanging fruit are.

Chin-Yun Yu is a fourth-year PhD student at the Centre for Digital Music, Queen Mary University of London, working on knowledge-driven voice synthesis. He received his B.S. degree in computer science from the National Yang Ming Chiao Tung University, Taiwan, in 2018, and later joined the Music and Culture Technology Lab as an RA, supervised by Li Su. He was an independent audio researcher before joining C4DM in 2022, and many of his implementations are open-sourced on GitHub.  His research interests include all aspects of signal processing, music information retrieval, and maybe some deep generative models and spatial audio. He's currently working on a PyTorch package, philtorch, for fast gradient computations across a wide range of filters and linear systems.