Odd Challenges of Using Deep Learning in Designing a Feedback Delay Network Reverb

Explore the unique challenges of applying deep learning techniques to design a Feedback Delay Network (FDN) reverb in this 44-minute conference talk from the Audio Developer Conference (ADC) 2023. Delve into the intersection of digital signal processing (DSP) and machine learning, focusing on optimizing audio effects and synthesizer parameters. Examine the limitations in areas such as parameter inference from audio input and Differentiable Digital Signal Processing (DDSP). Investigate the complexities surrounding IIR filters, including stability, interpretability, and differentiability issues. Follow the speakers' research journey in modeling room Impulse Responses using FDNs, covering approaches from naive to advanced. Gain insights into machine learning challenges specific to DSP applications, including approximating common transformations, improving computational efficiency, managing feedback system explosivity, and attempting to differentiate the undifferentiable.