On the Evaluation of Rhythmic and Melodic Descriptors for Music Similarity

In exploratory studies of large music collections where often no ground truth is available, it is essential to evaluate the suitability of the underlying methods prior to drawing any conclusions. In this study we focus on the evaluation of audio features that can be used for rhythmic and melodic content description and similarity estimation. We select a set of state-of-the-art rhythmic and melodic descriptors and assess their invariance with respect to transformations of timbre, recording quality, tempo and pitch. We create a dataset of synthesised audio and investigate which features are invariant to the aforementioned transformations and whether invariance is affected by characteristics of the music style and the monophonic or polyphonic character of the audio recording. From the descriptors tested, the scale transform performed best for rhythm classification and retrieval and pitch bihistogram performed best for melody. The proposed evaluation strategy can inform decisions in the feature design process leading to significant improvement in the reliability of the features.