This paper describes some optimizations to our speech recognition system, which is based on semi-continuous Hidden Markov Models (SCHMM) of subword units. The optimizations pertain to codebook generation, Linear Discriminant Analysis (LDA), initialized training, and definition of subword units. The recognition rate of the continuous version of the system increased from 79% to 95% combining all of the optimization steps.
Keywords: SCHMM, LDA, continuous speech recognition, codebook generation, sub-word units