Segment-based Melody Matching for Query By Singing/Humming

Segment-based melody matching approaches are proposed to solve the problems of pitch instability, tempo inconsistency, and puff noise of singing/humming in Query By Singing/Humming system. Instead of comparing frame by frame from the beginning and trying different scaling ratio, our segment-based approaches reduce the computation complexity by jumping to the possible segment boundary and using the segment ratio as scaling ratio. Our methods predict segment boundary candidates in query singing/humming and songs in database by detecting the abrupt change of pitch curve, match segment pair, scale each segment linearly and calculate the distance. The top 10 recognition rate of our method can reach 70.6%. Our method reduce computational load and performs better in situations of puff noise and inconsistent tempo than popular global Linear Scaling method.