Towards a Hybrid Deep-Learning Method for Music Classification and Similarity Measurement

Abstract

Large repository of music that can be accessed or downloaded over the Internet, provides a new way of trading or sharing. However, the technologies for features based Music Information Retrieval (MIR), which is a multidisciplinary field of research, are not well established. Existing MIR techniques and products suffer from either limited capabilities or poor performance. In this paper, we proposed a data model that describes the music information using both Music Definition Language (MDL) and Music Manipulation Language (MML), and supports extensible hybrid methods for music classification and similarity measurement. With proposed musical data model, we further developed a hybrid method that combines both contour and rhythm features, and employed an Artificial Neural Network (Unsupervised Kohonen Self-Organized Map) based classification mechanism that maps variations of music pieces to their corresponding originals using a new vector/matrix format defined as MDL. The proposed hybrid method based on a deep-learning mechanism and a new similarity measurement method has been introduced to fulfil analysis on the music classification and their similarity scores. The test results demonstrate that an accuracy of around 70% in the experiments has been achieved.