Conventional linear prediction of speech suffers from oversimplification in both vocal tract modeling and voice source modeling. We have presented a six-parameter model for the glottal source and a method for simultaneous estimation of voice source and vocal tract parameters based on an ARMA model combined with the glottal model (henceforth the GARMA method ). This paper describes further effort to refine our previous proposal. Taking into account the non-Gaussian nature of the innovation for voiced speech sounds, the robust signal processing techniques are introduced as an effective means of separating the voice source from the vocal tract transfer function. By minimizing the sum of appropriately weighted errors rather than the sum of squared errors, the proposed method is more robust against glottal excitation as well as random noise interferences. Therefore higher synthesized speech quality can be obtained by the robust GARMA method.