Deep Neural Network-Based Music Information Retrieval

Music classification is essentially a pattern recognition problem, including two aspects:feature extraction and classification. Audio data has high redundancy, high-dimensional features, and must be extracted through the feature extraction method in order to effectively reduce redundancy and dimensions. Feature extraction is performed by analyzing the audio signal to obtain a time-varying audio signals characteristic parameters. Different feature extraction methods can affect the results of the follow-up of music classification Different feature extraction methods to extract the characteristic parameters directly affect the results of the follow-up of music classification. So feature extraction is a key step to music classification task.Deep learning, as a new feature extraction technique, made a series of success in speech signal processing field. Based on excellent achievements of deep learning in speech signal processing area, this paper combine the theory of deep learning and music classification, try to find more suitable acoustic music features for music information retrieval.In this paper, we gave an introduction to the concepts and methods of music information retrieval and deep learning. Then studied draw on how to use deep learning for music information retrieval problem. We proposed a novel algorithm which uses deep belief network for music emotion classification and added convolution to the algorithm. We compared the features extracted through proposed approach to the MFCC feature, and got a better result.