This paper presents a new and improved method for decomposing the speech signal into a periodic and an aperiodic part. The method developed here use sinusoidal modeling techniques introduced by other authors, but new ideas are presented here that achieve improved output quality. The method involves: (1) Separation of speech into an approximate excitation and filter components using Linear Predictive (LP) analysis; (2) Reconstruction of the two excitation components of the residual using a sinusoidal modeling techniques. That is, the periodic part is approximated by a sum of sinusoids whose frequencies are harmonics of the fundamental frequency, whose magnitudes and phases are assumed to be linear and third order polynomial respectively with respect to time. The aperiodic part is obtained by subtracting the reconstructed periodic part signal from the residual signal; (3) Finally, the periodic and aperiodic components of the excitation are obtained by combining the reconstructed frames of data using an overlap-add procedure. Then, the individual parts of the original signal can be obtained by using the corresponding parts of the residual to pass through the derived all-pole filter. This new method is a powerful tool for analysis of relevant features of the source component of speech signal.