Melody Extraction based on Harmonic Coded Structure

Abstract

This paper considers a melody extraction algorithm that estimates the melody in polyphonic audio using the harmonic coded structure (HCS) to model melody in the minimum mean-square-error (MMSE) sense. The HCS is harmonically modulated sinusoids with the amplitudes defined by a set of codewords. The considered algorithm performs melody extraction in two steps: i) pitch-candidate estimation and ii) pitch-sequence identification. In the estimation step, pitch candidates are estimated such that the HCS best represents the polyphonic audio in the MMSE sense. In the identification step, a melody line is selected from many possible pitch sequences based on the properties of melody line. Posterior to the melody line selection, a smoothing process is applied to refine spurious pitches and octave errors. The performance of the algorithm is evaluated and compared using the ADC04 and the MIREX05 dataset. The results show that the performance of the proposed algorithm is better than or comparable to other algorithms submitted to MIREX2009.