Clustering Music Recordings by Their Keys

Music key, a high level feature of musical audio, is an effective tool for structural analysis of musical works. This paper presents a novel unsupervised approach for clustering music recordings by their keys. Based on chroma-based features extracted from acoustic signals, an inter-recording distance metric which characterizes diversity of pitch distribution together with harmonic center of music pieces, is introduced to measure dissimilarities among musical features. Then, recordings are divided into categories via unsupervised clustering, where the best number of clusters can be determined automatically by minimizing estimated Rand Index. Any existing technique for key detection can then be employed to identify key assignment for each cluster. Empirical evaluation on a dataset of 91 pop songs illustrates an average cluster purity of 57.3% and a Rand Index of close to 50%, thus highlighting the possibility of integration with existing key identification techniques to improve accuracy, based on strong cross-correlation data available from this framework for input dataset.