DTV-Based Melody Cutting for DTW-Based Melody Search and Indexing in QbH Systems

Melody analysis is an important processing step in several areas of Music Information Retrieval (MIR). Processing the pitch values extracted from raw input audio signal may be computationally complex as it requires substantial effort to reduce the uncertainty resulting i.a. from tempo variability and transpositions. A typical example is the melody matching problem in Query-by-Humming (QbH) systems, where Dynamic Time Warping (DTW) and note-based approaches are typically applied. In this work we present a new, simple and efficient method of investigating the melody content which may be used for approximate, preliminary matching of melodies irrespective of their tempo and length. The proposed solution is based on Discrete Total Variation (DTV) of the melody pitch vector, which may be computed in linear time. We demonstrate its practical application for finding the appropriate melody cutting points in the R∗-tree-based DTW indexing framework. The experimental validation is based on a dataset of 4431 queries and over 4000 template melodies, constructed specially for testing Query-byHumming systems.