In this paper, we propose a wind noise reduction method based on various types of non-negative matrix factorization (NMF) approaches. Since wind noise has highly non- stationary spectral characteristics that are difficult to remove using stochastic oriented methods, it is more effective to use template matching-based methods.

We first investigate whether the quality of enhanced output varies depending on the ratio of the size of speech models to that of noise models. We especially show that the optimal ratio is related to the signal-to-noise ratio (SNR) of the input signal. Based on the results of analysis, we propose an efficient algorithm for adaptively changing the size of speech and noise NMF models in each analysis frame. Since the proposed algorithm takes into account the trade- off relationship between speech distortion and noise reduction, its output quality becomes very natural. The experimental results also confirm the superiority of the proposed algorithm to conventional template matching based algorithms.