Music SimMetric |  |
A System for Quantifying Music Similarity through Digital Signal Processing
Digital music formats are fast becoming the pervasive mode of music consumption. Technologies like peer-to-peer networking and perceptual audio encoding have enabled even casual music enthuisiasts to amass digital music collections of thousands, if not tens-of-thousands of song titles. Advances in digital rights management (DRM) techniques have allowed online music stores to offer customers millions of song titles. On all levels, the amount digital music content available for consumption has grown to unmanageable proportions.This trend has prompted recent research in the area of content-based music information retrieval (MIR). This diverse body of research encompasses problems like automatic genre classification (Tzanetakis), automatic song summarization (Foote), and music similarity quantification (Aucouturier, Logan). This work describes Music SimMetric a system for deriving music similarity metrics from a set of music files using digital signal processing techniques. The system employs three distinct dimensions of similarity: timbral similarity, rhythmic similarity, and structural similarity, to place individual songs in a “music similarity space.”
SimMetric will be tested on a set of popular music files obtained from the iTunes online music store. The resulting similarity scores will be compared
to music similarity data provided by the Music Genome Project.
- Timbre Model The Timbre Model uses spectral analysis and statistical modeling to provide a picture of the timbres present in a music file. Specialized statistical methods are used to calculate a timbral similarity distance between music files. This modeling approach is based on the work of B. Logan, J. Aucouturier, and E. Pampalk.
- Rhythm Model The Rhythm Model uses a high-resolution self-similarity analysis to characterize the rhythmic patterns of a music file and represent them with a vector. Standard cosine distance measures are used to calculate a rhythmic similarity distance between music files. This modeling approach is based on the work of J. Foote.
- Structure Model The Structure Model uses a low-resolution self-similarity analysis correlated with novelty kernel to identify changes in song structure. This modeling approach is also based on the work of J. Foote.