In this paper, we describe an experimental telephone based system that recognizes speaker-independent isolated words. The recognition method is based on discrete HMMs. We apply the following new techniques to the conventional discrete HMM method; interpolation of observation probabilities using Fuzzy Vector Quantization, multiple model construction, model training using expanded speech end-points, and state duration control using Gaussian windows. Experiments are carried out on Japanese digits spoken by 269 speakers (238 for training, 31 for evaluation). An improvement of about 4.5% in recognition accuracy is obtained with the new techniques.