• Most phone-based systems have 3 state HMM models regardless of the task

• Syllable models are typically chosen to be proportional to average duration for that task

• The initial model choice is empirical

• May have impact on the ability of the trained model to fit the training data