Wind noise reduction using empirical mode decomposition

Abstract

One common problem of recorded sound outdoors is interfusion of wind noise, which has highly non-stationary characteristics. Although there are a lot of noise reduction methods which produce good results for general kinds of noises, most methods perform worse for wind noise due to its non-stationary nature. Therefore, wind noise reduction need special technique to overcome this non-stationarity. Empirical mode decomposition (EMD) is a relatively new method to decompose a signal into several nonlinear and non-stationary bases which are modeled as amplitude and frequency modulated sinusoids that represent wind noise well. Thus, EMD has a possibility to reduce wind noise from recorded sounds in entirely different way from ordinary methods. In this paper, a preliminary discussion of applying EMD to wind noise reduction is presented. Since EMD decomposes a signal into monocomponent bases, it is easier to treat them as analytic signals via Hilbert transform. Our method utilize this characteristics in order to reduce wind noise. The experiment is performed on female voice superimposed with wind noise and shows it possibility and effectiveness.