Iterative Deepening for Melody Alignment and Retrieval

For melodic theme retrieval there is a fundamental tradeoff between retrieval performance and retrieval speed. Melodic representations of large dimension yield the best retrieval performance, but at high computational cost, and vice versa. In the present work we explore the use of iterative deepening to achieve robust retrieval performance, but without the accompanying computational burden. In particular, we propose the use of a smooth pitch contour that facilitates query and target representations of variable length. We implement an iterative query-by-humming system that yields a dramatic increase in speed, without degrading performance compared to contemporary retrieval systems. Furthermore, we expand the conventional iterative framework to retain the alignment paths found in each iteration. These alignment paths are used to adapt the alignment window of subsequent iterations, further expediting retrieval without degrading performance.