Singing Melody Extraction from Pop Songs Using a Novel Feature and Viterbi Search

Extracting singing melody from audio signals is one essential component for musical information retrieval. One can identify melody of different types of music easily even without professional training, but it is difficult for computer to achieve this goal. In this paper, we present a new feature which provides not only a convenient way to observe music but also a tool to extract singing melody. Also, we propose pitch likelihood, note and energy transition probabilities to perform Viterbi search. The experimental results show that our method can extract singing melody accurately.