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.
Bibliographic reference. Scagliola, Carlo / Vicenzi, Cesare / Carossino, Angelo / Sciarra, Donatella (1989): "Iterative optimization of sub-word templates for speech recognition", In EUROSPEECH-1989, 1079-1082.