Detecting Features for a Music Retrieval System

A music retrieval system inputs an audio file and outputs metadata, such as the music title and the singer's name. Many music retrieval systems recognize music with the original key and tempo. However, they are not suitable for music with different keys or tempos. This paper proposes a system to identify arranged songs by extracting and quantifying their beats, melodic line, and chords. The proposed method in this system compares differences in melodic lines and a vector dictionary of chords considering transpositions. We can ignore the tempo difference between the original and target sounds since we analyze melodic lines and chords in each beat. This method achieved an 89% recognition rate when we extracted chord progression but a 21% recognition rate when we extracted melody.