Abstract
This thesis proposes an improved framework for improving both the efficiency and the effectiveness of a query by singing/humming (QBSH) system. The proposed framework is based on three methods. Method 1 uses golden section search to reduce the computation time in traditional linear scaling (LS) algorithm. Method 2 assigns different weights for rests (in both database songs and in queries) so that these rests now have less effect on computing the weighted distance. Method 3 utilizes a sorted error vector to ignore the LS distances that are overly large and only considers the rest of the LS distances in the computation. This reduces the effect of pitch deviation in a short time span, probably due to the singer being out of tune or errors in pitch track-ing. The proposed framework improves the baseline system in both the computation time reduction (via scheme 1) and recognition accuracy (via schemes 2 and 3) of LS-based QBSH. Our experiment shows an error reduction rate of 21.4% in accuracy and 49.3% decrease in computation time on the MIR-QBSH dataset.