Evaluation of Chord and Chroma Features and Dynamic Time Warping Scores on Cover Song Identification Task
Abstrakt
Cover song identification has been a popular task within music information retrieval in the 20th century. The task is to identify a different version or performance of a previously recorded song.
Unlike audio search for an exact matching song, this task has not yet been popularized among users, due to an ambiguous definition of a cover song and the complexity of the problem. With a great variety of methods proposed on the benchmarking challenges, it is increasingly difficult to compare advantages and disadvantages of the features and algorithms.
We provide a comparison of three levels of feature extraction (chroma features, chroma vector distances, chord distances) and show how each level affects the results. We further distinguish five scores for dynamic time warping method, to find the best performance in conjunction with the features.
Results were evaluated on covers80 and SecondHandSongs datasets and compared to the state-of-the-art.