During atrial fibrillation (AF), time-frequency analysis of atrial signal has been applied to describe fibrillatory frequency trends. Recently, temporal changes in spectral shape have been investigated using the spectral profile technique. This profile is computed recursively by fitting each short-time log-spectrum to a spectral template, using amplitude scaling and frequency shifting. The purpose of the present study was to develop a Gaussian mixture model of the spectral profile in order to characterize the shape of AF waveforms. A novel index is introduced, the so-called harmonic index (HI), which reflects properties of the fundamental frequency peak and related harmonics peaks as estimated from the model. The index was tested on recordings from 9 patients with persistent AF, obtained before and after exercise testing. The HI succeeded in monitoring the response to exercise, i.e. change in the spectral profile to a less harmonic pattern, which is consistent with a reduction in AF organization (HI: 0.61+/-0.11 vs. 0.50+/-0.19, rest vs. exercise; p<0.05).