Zivanovic et al.
57distributions of the different peak classes overlap, and the optimal determination of the decision boundaries depends on the specifi c application.The peak classifi cation method proposed in Zivanovic, Röbel, and Rodet (2004) uses descriptors that were designed to adequately characterize nonstationary sinusoidal signals. These descriptors have proven to lead to better classifi cation performance than other approaches devoted to sinudoidal detection / estimation (Thompson 1982;Rodet 1997). It was also shown in Zivanovic, Röbel, and Rodet (2004) that the peak classes can be characterized by distributions in the descriptor domains, similar to probability density functions. Once the distributions have been generated, a simple decision tree can be derived that allows the classifi cation of spectral peaks into sinusoids, noise, and sidelobes.The peak classifi cation method has been used successfully in a number of applications. As examples we mention polyphonic fundamentalfrequency detection (Yeh, Röbel, and Rodet 2005), adaptive noise-fl oor determination (Yeh and Röbel 2006), and voiced / unvoiced frequency boundary determination. Another interesting application lies in the pre-selection of the sinusoidal peaks to reduce the number of candidate peaks considered for partial tracking in additive analysis. A reliable classifi cation of noise peaks could reduce the number of incorrect connections, and, for probabilistic approaches like that described by Depalle, Garcia, and Rodet (1993), it would considerably reduce the computational cost.The major problem with the classifi cation scheme of Zivanovic, Röbel, and Rodet (2004) is the control of the classifi cation boundaries (classifi cation thresholds) that generally need adaptation for the specifi c problem at hand. Another problem is that the descriptor boundaries of the different classes will depend on the analysis window that is used. UpThe decomposition of audio spectra into sinusoids, transients, and noise can serve as a useful tool for improving the results of parameter estimation or signal manipulation applications. As has been shown for the case of transient detection (Röbel 2003) and sinusoidal and noise discrimination (Zivanovic, Röbel, and Rodet 2004), the classifi cation of spectral peaks is a benefi cial step toward the identifi cation of these signal components. Such a classifi cation scheme that makes optimal use of the information provided by spectral peaks can then be used to achieve a robust segmentation into higher-level signal components, for example, partials or unvoiced regions. Unlike the perceptual audio segmentation (Painter and Spanias 2005), which attempts to maximize the matching between the auditory excitation pattern associated with the original signal and the corresponding auditory excitation pattern associated with the modeled signal, we base our classifi cation purely on signal characteristics.The basis for spectral peak classifi cation is an adequate choice of criteria that would best describe sinusoidal and noise spec...