An adaptive, pitch-synchronous analysis method is proposed for the simultaneous estimation of vocal tract and voice source parameters from speech waveforms. A time varying autoregressive model with exogenous input (ARX) is chosen for vocal tract modeling because of the capability of such a model for characterising both the formants and antiformants of the vocal tract. The Liljencrants-Fant model for the voice source is integrated into an iterative adaptive estimation procedure. Furthermore, an adaptive inverse filtering technique is put forward to obtain high accuracy estimation of the glottal source waveform, which is necessary for the intended application of the method to pathological voice analysis. The technique is evaluated and compared with a number of other approaches using synthetic speech containing additive noise at the source. The results illustrate the superior performance of the new method.