We present an improved algorithm for a robust speech analysis-synthesis method based on an auto-regressive with exogenous input (ARX) speech production model proposed previously. The speech analysis-synthesis method is capable of making an automatic estimation of vocal tract (formant) and voice source parameters from a speech utterance, generating accurate formant values even for very high-pitched voices. The improved algorithm presented in this paper incorporates aperiodic components included in the voice source signal, taking the dynamic nature of the speech production process into account. Perceptual experiments show that implementation of the aperiodic components in the analysis-synthesis is very effective in improving the perceived quality of synthetic speech, particularly for soft voices, typical of female voice quality.