In this paper we propose the use of an iterative quadratic classifications algorithm as unsupervised training procedure in the frame-based approach to the recognition of non-stationary patterns , We present a comparative experimental analysis of the proposed algorithm and the well-known c-mean clustering algorithm. The efficiency of these two iterative clustering procedures is experimentally evaluated through their application in the robust recursive identification of the time-varying auto regressive (AR) model of speech signals. Means and standard deviations of the Mean Absolute Residual (MAR) criterion as well as trajectories of the estimated AR parameters show the superiority of the proposed procedure over the c-mean clustering algorithm.