Music Emotion Classification Using Deep Belief Networks

• Main Authors: CHI HSU, 許霽
• Other Authors: Jyh-Jong Tsay
• Format: Others
• Language: en_US
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/18411289638893719664

碩士 === 國立中正大學 === 資訊工程研究所 === 102 === Deep learning which is capable of learning features from raw data is one of the most important methods in recent machine learning research with remarkable successes in many practical object recognition tasks, such as face recognition, traffic sign recognition, or even Dogs vs. Cats image recognition. It have been used by many winners in recent Kaggle Data Science competitions. In this thesis, we propose to study the effectives of features learned by deep learning for music data. We studied two unsupervised feature learning systems, Restricted Boltzmann Machines (RBMs) and Deep Belief Networks (DBNs), and carried extensive experiments for two important tasks in music information processing, music genre classification and emotion recognition. The results show that unsupervised learned features outperform traditional features such as Mel-frequency cepstral coefficients (MFCCs) which has been widely used to represent audio data.