I'll be giving a talk at DataBeersBCN next Tuesday. The title is "Quick-and-dirty facts about deep learning", and in it I'll expose a series of facts and myths about these models.

Raga is the melodic framework of Indian art music. It is a core concept used in composition, performance, organization, and pedagogy. Automatic rāga recognition is thus a fundamental information retrieval task in Indian art music. In this paper, we propose the time-delayed melody surface (TDMS), a novel feature based on delay coordinates that captures the melodic outline of a raga. A TDMS describes both the tonal and the temporal characteristics of a melody, using only an estimation of the predominant pitch. Considering a simple k-nearest neighbor classifier, TDMSs outperform the state-of-the-art for raga recognition by a large margin. We obtain 98% accuracy on a Hindustani music dataset of 300 recordings and 30 ragas, and 87% accuracy on a Carnatic music dataset of 480 recordings and 40 ragas. TDMSs are simple to implement, fast to compute, and have a musically meaningful interpretation. Since the concepts and formulation behind the TDMS are generic and widely applicable, we envision its usage in other music traditions beyond Indian art music.

S. Gulati, J. Serrà, K.K. Ganguli, S. Senturk, & X. Serra. Time-delayed melody surfaces for raga recognition. Proc. of the Int. Soc. for Music Information Retrieval Conf. (ISMIR), pp. 751-757. Aug 2016. [MTG] [ISMIR]