Convolutive Sparse Non-negative Matrix Factorization for windy speech

This paper presents a method for suppressing wind noise in a single channel recording of speech outdoors. This proposed method is based on the convolutive extension of Sparse Non-negative Matrix Factorization, in which a convolutive model is used to attenuate the acoustic effect of wind noise based on wind noise codebook estimated from a recording of pure wind noises. The extended method is convolutive in time domain, thus the potential interframe information of the speech can be exploited to get a more effective result. In the paper, we first exploit the characteristics of wind noise, then focus on the extension method of SNMF and demonstrate its effectiveness on wind noise reduction.