A novel approach to classification of peaks of audio spectra is presented. In extending previous work on detecting transient spectral peaks we here investigate into the classification of sinusoidal and noise peaks. The classification is based on descriptors derived from properties related to time-frequency distributions: mean time, duration, instantaneous frequency and normalized bandwidth. In contrast to existing methods, the descriptors are designed to properly deal with non-stationary sinusoids, which considerably increases the range of applications. The experimental investigation shows superior classification results compared to the standard correlation-based approach.