Multipitch estimation based on correlation between a harmonic structure model and an observed spectrum for mixed audio signals

Abstract

In the research of sound source separation and resynthesis, multipitch estimation is an important topic. We propose a multipitch estimation technique for monaural audio signals. It introduces a method of calculating multiple pitches within a frame for a pitched sound signal. First, a harmonic structure is modeled for a pitch M (Hz), using the cases of odd integer multiples of M, even integer multiples of M, and odd integer multiples of M/2. Then correlations are derived between each of the harmonic structure models and an observed spectrum. Harmonic structure models are expressed as the summation of Gaussian functions in the frequency domain, where each harmonic has a peak at its center frequency. An evaluation function is dynamically defined to test how well arbitrary values of M fit the observed spectrum. The evaluation function is created by searching for weighted coefficients for the correlation values. Next, a number of maximal peaks equal to the number of sound sources are selected from the spectrum and defined as the estimation pitches. This method was applied to a mixed piano signal, and a mixed audio signal that includes flute, clarinet, and cello. The results are good except at octave unisons.