An iterative optimization procedure to derive reliable sub-word templates from training speech material is described. At each step sub-word occurrences are located and new templates computed in such a way that the overall distance of speech material from the language model approaches an absolute minimum. Different options are discussed and experimental results presented.