Melodic Pattern Extraction in Large Collections of Music Recordings Using Time Series Mining Techniques

We demonstrate a data-driven unsupervised approach for the discovery of melodic patterns in large collections of Indian art music recordings. The approach first works on single recordings and subsequently searches in the entire music collection. Melodic similarity is based on dynamic time warping. The task being computationally intensive, lower bounding and early abandoning techniques are applied during distance computation. Our dataset comprises 365 hours of music, containing 1,764 audio recordings representing the melodic diversity of Carnatic music. A preliminary evaluation based on expert feedback on a subset of the music collection shows encouraging results. In particular, several musically interesting relationships are discovered, yielding further scope for establishing novel similarity measures based on melodic patterns.