Blind Clustering of Music Recordings Based on Audio Fingerprinting

Although multiple music recordings may sound identical to a human listener, the underlying representations of sound may differ due to the variations in their audio encoding and/or transmission methods. In contrast to the existing audio-fingerprinting techniques, which establishes the fingerprint of each source music to identify unknown, ldquodistortedrdquo audio clips, this paper proposes an unsupervised clustering framework to identify unknown, ldquodistortedrdquo audio clips derived from the same source music. By grouping together audio data derived from the same music, the human effort required to label music data can be dramatically reduced. This work develops methods to measure the similarities between audio clips and use hierarchical agglomerative clustering to group together audio clips that are similar to one another. Also proposed is a method based on the Rand Index to determine the optimal number of clusters automatically in relation to the number of source music recordings.