Extracting singing melody in music with accompaniment based on harmonic peak and subharmonic summation

In content-based music applications, music melody is one of the most basic features. In this paper, we proposed a melody extraction method for music with accompaniment based on spectral peak peaking and subharmonic summation. We conducted the performance evaluation of the proposed method against the Cepstrum method, the harmonic overtone detection (HOD) based method and the YIN algorithm using the MIREX data set. Our evaluation results showed that the proposed method is accurate, fast and interference-resistent. The overall pitch accuracy of the proposed method is as high as 92.3% for clean music, and 51.6% for music with accompaniment of equal strength. We also implemented this method in a prototype online Karaoke system and demonstrated its real-time performance in a practical system.