We investigate a robust parametric approach to the estimation of time-varying AR parameters in speech. The time-varying parameters are assumed to be linear combinations of a set of basis time functions so that the model can be specified by constant parameters. The parameters of the linear combination of functions are obtained by the M-estimator defined by the minimization of an nonquadratic function of normalized residuals. We have demonstrated by computer simulation on synthetic data that the proposed method is less sensitive to variations of the pitch frequency.