Towards estimation of the number of concurrent F0s

Abstract

Two central problems common to many multiple F0 estimation algorithms will be presented: The estimation and handling of the noise level and the modeling and representation of overlapped partials. Contrary to the widely used assumption that the noise is white Gaussian, it is proposed to use an adaptive and frequency-dependent noise level to be able to improve the description of sinusoidal partials and the expected residual. By using the frequency-dependent noise level, an F0 estimation algorithm can be guided to extract sources having designated harmonic-to-noise ratios. One of the major problems in multiple F0 estimation is partial collision. The assumption that is usually made to estimate the expected amplitudes of overlapped partials is the additivity of magnitude spectra or the additivity of power spectra. A mathematical investigation into the properties of overlapped amplitudes allows us to present an improved probabilistic description of the expected amplitudes. The implications of these two refinements for the estimation of the number of sources that are present in a recording will be discussed, and preliminary experimental results will be presented.