Content-Based Features in the Composer Identication Problem CS 229 Final Project
Sean Meador · 2008
Classification of digital music (also called Music Information Retrieval, or MIR), is a long-standing problem in machine learning, one with many potential real world applications. For example, providing accurate recommendations is extremely important to online music websites (Pandora, iTunes Genius, etc.) as this is the primary means by which users discover new music. Many classification schemes have therefore been proposed, with some attempting to group music according to characteristics like mood and genre, and others attempting to identify the particular artist or composer who created the work. Due to the complex nature of audio waveforms, however, nearly all music classification algorithms begin with a feature extraction step. Raw digital audio is essentially unusable for direct training on a learning algorithm; therefore, it is first necessary to process the data and distill key identifying features from the audio clips. Until recently, researchers were getting very good results using only relatively low-level signal features of the audio [5]. However, after the discovery of a design flaw now known as the “album effect”, the performance of such classifiers has dropped dramatically [1]. In this paper, we focus on classical composer identification, and propose a content-based feature set which addresses the limitations of current classifiers caused by the album effect.