Gradient-based musical feature extraction based on scale-invariant feature transform

We investigate a novel gradient-based musical feature extracted using a scale-invariant feature transform. This feature enables dynamic information in music data to be effectively captured time-independently and frequency-independently. It will be useful for various music applications such as genre classification, music mood classification, and cover song identification. In this paper, we evaluate the performance of our feature in genre classification experiments using the data set for the ISMIR2004 contest. The performance of a support-vector-machine-based method using our feature was competitive with the contest even though we used only one fifth of the data. Moreover, the experimental results confirm that our feature is relatively robust to pitch shifts and temporal changes.