How MFCC features and other music features change with musical key?

Abstract

Nowadays there are abundant media resources for us to explore and access on the Internet. The physical medium of media such as CDs, DVDs and tapes has nearly been replaced by online media services. As one of the most popular media types, music is constantly accessible in everyday life. Music is also an important platform which facilitates communication among different populations and nations. At the same time, a new problem arises when we want to retrieve different types of media resources. The information contained in music is complicated and hard to maintain. No matter from personal view or business view, music information retrieval(MIR) is a promising future for the industry field, same for the automatic music genre classification. It triggered my project focus on MIR using content-based method, specifically music genre classification. A genre classification system can be used to organize a music database, or to suggest similar songs for a playlist. The goal is to identify the genre from the contents of the song, i.e. the audio signal. The Mel Frequency Cepstral Coefficients (MFCCs), which are calculated from the short-time Fourier transform of the audio signal, is one popular feature set used to represent the audio signal of a song. As it is based on a frequency spectrum, the feature changes when the same song is played in a different key. In this project, the invariance of MFCCs to key changes is studied. In addition, this project will investigate a way to extract MFCC-like features that remain constant even if the key of a song changes. In other words, the feature would remain the same even if the key of the song were to change. This project shall explore a way to train a GMM model which is invariant to key transformations. The effectiveness of the new feature will be tested by being applied to music genre classification.