Nonlinear prediction is a natural way to increase the quality of speech coders. Several approaches have been recently proposed in this direction ([1,2,3,4] are some examples) and most of them use neural networks as predictors. Nevertheless, the computational cost due to the network training is very high, since it usally involves a gradient descent-based nonlinear optimization process. In this paper we propose some improvements of our previous work reported in [3], all of them aiming at reducing the computational cost. Our predictor can be used in CELP-type coders and provides a 0.6 dB increase of the SEGSNR with respect to conventional CELP coders.