In this paper we describe the procedure for establishing a robust set of phoneme prototypes suitable for Speaker Adaptable Isolated Word Speech Recognition (IWSR) systems. The procedure is based on a K-Means cluster analysis of a large multispeaker phoneme data base. The prototypes were tested on a large vocabulary IWSR system for Greek developed in the framework of the ESPRIT 2104 project, POLYGLOT-I [1], and the results obtained showed a significant improvement in the accuracy of the system.