Extraction of spectral peak parameters using a short-time Fourier transform modeling and no sidelobe windows

A new method which improves the estimation of frequency, amplitude and phase of the partials of a sound is presented. It allows the reduction of the analysis-window size from four periods to two periods. It therefore gives better accuracy in parameter determination, and has proved to remain efficient at low signal-to-noise ratios. The basic idea consists of using a parametric modeling of the short-time Fourier transform. The method alternately estimates the complex amplitudes and the frequencies starting from the result of the classical analysis method. It uses the least-square procedure and a first-order limited expansion of the model around previous estimations. This method leads us to design new windows which do not have any sidelobes in order to help the convergence. Finally an analysis algorithm which has been built according to the observed behavior of the method for various kinds of sound is presented.