The target of this project is to filter out wind noise from speech segments, giving great importance to the fact that speech won't be distorted in the process. For this purpose, we reviewed existing solutions and focused on filtering by matrix factorization. We studied this method and characterized the parameters that lead to the best results of the algorithm. We then chose aspects we would like to improve, such as using the signal's time dependence and low frequencies reconstruction. For the time dependence we generalize the existing algorithm according to an existing article, while for the low frequencies reconstruction we developed a new method that includes learning the characteristics of the speaker. These solutions have resulted in improved signal-to-noise ratios and better conservation of speech characteristics as required. In addition, we offered a more appropriate real-time processing and explored other ways to improve the algorithm, such as using wavelets transform.