Rapid grow of the digital music content and service providers worldwide everyday increases the importance of music genre classification. Most genre classification still relies heavily on human effort. Signal processing combined with machine learning methods aims to solve this problem autonomously for decades. In this work, we introduce novel high-level features derived from song structures and examine their performance through both CNN and a Voting Classifier. Results show that these features alone increases the classification accuracy significantly compared to random prediction and has potential of use in combination with other various features.