Efficient melodic query based audio search for Hindustani vocal compositions

Time-series pattern matching methods that incorporate time warping have recently been used with varying degrees of success on tasks of search and discovery of melodic phrases from audio for Indian classical vocal music. While these methods perform effectively due to the minimal assumptions they place on the nature of the sampled pitch temporal trajectories, their practical applicability to retrieval tasks on real-world databases is seriously limited by their prohibitively large computational complexity. While dimensionality reduction of the time-series to discrete symbol strings is a standard approach that can exploit computational gains from the data compression as well as the availability of efficient string matching algorithms, the compressed representation of the pitch time series itself is not well understood given the pervasiveness of pitch inflections in the melodic shape of the raga phrases. We propose methods that are informed by domain knowledge to design the representation and to optimize parameter settings for the subsequent string matching algorithm. The methods are evaluated in the context of an audio query based search for Hindustani vocal compositions in audio recordings via the mukhda (refrain of the song). We present results that demonstrate performance close to that achieved by time-series matching but at orders of magnitude reduction in complexity