Wind Noise Reduction with a Diffusion-based Stochastic Regeneration Model

In this paper we present a method for single-channel wind noise reductionusing our previously proposed diffusion-based stochastic regeneration modelcombining predictive and generative modelling. We introduce a non-additivespeech in noise model to account for the non-linear deformation of the membranecaused by the wind flow and possible clipping. We show that our stochasticregeneration model outperforms other neural-network-based wind noise reductionmethods as well as purely predictive and generative models, on a dataset usingsimulated and real-recorded wind noise. We further show that the proposedmethod generalizes well by testing on an unseen dataset with real-recorded windnoise. Audio samples, data generation scripts and code for the proposed methodscan be found online (https://uhh.de/inf-sp-storm-wind).