A Differentiable Neural Network Approach To Parameter Estimation Of Reverberation

Differentiable Digital Signal Processing is a library and set of machine learning tools that disentangle the loudness and pitch of an audio signal for timbre transfer or for applying digital audio effects. This paper presents a DDSP-based neural network that incorporates a feedback delay network plugin written in JUCE in an audio processing layer, with the purpose of tuning a large set of reverberator parameters to emulate the reverb of a target audio signal. We first describe the implementation of the proposed network, together with its multiscale loss. We then report two experiments that try to tune the reverberator plugin: a "dark" reverb where the filters are set to cut frequencies in the middle and high range, and a "brighter", more metallic sounding reverb with less damping. We conclude with the observations about advantages and shortcomings of the neural network.