Our goal is to push the general performance of music genre recognition forward and introduce a new method for pre-processing which allows for faster experimentation and model tuning in the future. We experimented with two different musical representations: mel-spectrograms and manually extracted features.