Abstract
A query-by-singing/humming (QBSH) System is a song searching system that enables the user to find a song by singing or humming the melody. The sung or hummed melody is compared to the songs in the database to find the closest match. Two common comparing methods are dynamic time warping (DTW) and linear scaling (LS). The former method has the advantage of higher recognition accuracy, while the latter is more efficient in terms of computation time. The goal of this research is to improve current LS-based QBSH system so that the comparing time is reduced while not sacrificing too much recognition accuracy. In this thesis, the following approaches are taken to speed up the system: avoiding comparing the repeated patterns in a song, reducing the sample rate, comparing using the scaling ratios around 1.0 first, and comparing the songs with higher hit rate first. The experimental results show that the former two methods (avoiding comparing the repeated patterns in a song, and reducing the sample rate) are able to reduce the comparing time effectively, while the other two methods cannot speed up the system evidently in some conditions.