Music information retrieval using Query-by-humming based on the dynamic time warping

Abstract

Query-by-humming (QBH) is a content-based music information retrieval system that use humming as its query. Naturally, human cannot hum or sing perfectly. They may sing in different pitch, key, or tempo thus can affect the retrieval performance of the system. It becomes a challenge to build an effective QBH system despite all of the query imperfection. Many approaches have been proposed to develop QBH systems, such as contour approach, hidden Markov model, and dynamic time warping. One of a simple yet promising method for building a QBH system is the Dynamic Time Warping. In this paper, we describe our works on building a QBH system using the Dynamic Time Warping to search for Indonesian songs stored in the MIDI format. To evaluate the system, we use accuracy rate for humming transcription evaluation and mean reciprocal rank for retrieval evaluation. Most of retrieval results are in the first rank thus shown that the Dynamic Time Warping is effective to be applied on this task.