Abstract
This thesis discusses the methodology of speeding up melody recognition. The general algorithm for melody recognition now used is “Dynamic Time Warping”, shorted as DTW. While the database increases, the time cost for DTW will largely increase as well. As a result, we propose “step by step” recognition flow for acceleration. In first step, we introduce the differences of semitones between test and referenced songs, and by setting the threshold, we can save certain number of songs which are more similar with test songs. Then, we reduce the pitch vector by the parameter, frame rate, and the time cost will be reduced in proportion of frame rate. Different number of songs in the database will be kept according to different survival rate after this stage. And so on, we can combine several methods for melody recognition and design the system to reduce the time cost and increase the recognition rate. There are several stages in this combined system, and for mathematics analyzing, it requires some data: the recognition rate of stage and single song comparison time cost correspond to survival rate of stage. After achieving this information, we could know the approximate parameters setting for stages with dynamic programming. This system can achieve the balance between speed and performance. With this design, the stages will guarantee the recognition rate and speed up.