Query by Humming: Comparing Voices to Voices

In this paper, we describe a query by singing/humming (QBSH) system which compares voices to voices. Instead of comparing with MIDI database, our system compares humming queries directly with the songs sung or hummed by other people In our system, both humming queries and humming songs in the library are represented as sequences of <relative pitch, relative interval>. We use UDS string to represent notes. An improved harmonic product spectral (HPS) method is used for tracking pitch. Notes are segmented with a simple pitch segmenter. A modified approximate string matching method is used to measure the similarities between queries and humming songs in the library. Experiments reveal that voice-to-voice is a simpler and more effective way for QBSH.