Analyzing the influence of pitch quantization and note segmentation on singing voice alignment in the context of audio-based Query-by-Humming

Abstract

Query-by-Humming (QBH) systems base their operation on aligning the melody sung/hummed by a user with a set of candidate melodies retrieved from polyphonic songs. While MIDI-based QBH builds on the premise of existing annotated transcriptions for any candidate song, audiobased research makes use of melody estimation algorithms for the songs. In both cases, a melody abstraction process is required for solving issues commonly found in queries such as key transpositions or tempo deviations. Full automatic music processes are commonly used for this, but due to the reported limitations in state-of-the-art methods for real-world queries, other possibilities should be considered. In this work we explore three different melody representations, ranging from a general time-series one to more musical abstractions, which avoid full automatic transcription, in the context of an audio-based QBH system. Results show that this abstraction process plays a key role in the overall accuracy of the system, obtaining the best scores when temporal segmentation is dynamically performed in terms of pitch change events in the melodic contour.