Melody Extraction from Polyphonic Audio Signal Mirex2009

This paper describes the proposed algorithm submitted to the MIREX 2009 “Audio Melody Extraction” task. The algorithm addresses the task of extracting the predominant melody pitch from a polyphonic audio signal. The algorithm extracts the melody pitch in three steps. In the first step, transient analysis is performed on the polyphonic audio signal to determine the analysis frame length, and then a fixed number of pitch candidates are obtained by ranking the weights of the harmonic structure of the windowed signal. In the second step, a single dominant pitch sequence (melody line) is selected from the many possible pitch sequences based on the following properties of melody line: (1) while the nominal dynamic range of a singing vibrato is ± 60∼200 cents, it is only± 20∼30 cents for instruments; (2) melody transitions are typically limited to one octave; (3) a rest during singing is often longer than 50ms. In the third step, a smoothing process is performed to refine the estimated pitch sequence.