Wind noise poses a severe problemfor speech signal recording and processing in outdoor environments. Due to its highly nonstationary nature, classical single-channel noise reduction approaches often fail at correctly estimating the noise power. With more microphones being available in modern devices, wind-noise reduction can be considered a multichannel speech enhancement problem that is solved in a minimum mean-square error sense by the multichannelWiener filter (MWF). In this contributionwe propose to approximate the MWF in two steps. First, we apply blind channel identification to estimate the acoustical transfer functions in order to implement the spatial processing of theMWF. Our main contribution is then the derivation of a maximum-likelihood optimal computation of the spectral postfilter based on a short-time statistical model of the microphone signals. The proposed postfilter is evaluated in terms of segmental SNR and PESQ improvements.
展开▼
