Melody Extraction from Polyphonic Audio Signal Mirex2010

This paper considers the proposed algorithm submitted to the Music Information Retrieval Evaluation eXchange (MIREX) 2010 “Audio Melody Extraction” task. The proposed melody extraction algorithm can be divided into three steps: (1) a spectral analysis using a variable length window, (2) a pitch candidate estimation, and (3) a pitch sequence identification. In the first step, the short-time Fourier transform (STFT) with variable length window is performed to be robust against dynamic variation of melody line. In the second step, melody pitch candidates of each frame are obtained from weights of a harmonic structure in the spectrum. Furthermore, a melody pitch range estimation is also considered to reduce false-positive and computation. In the third step, a single pitch sequence (melody line) is selected based on the general properties of melody line.