Although the ML-trained systems with 60 and
70 states yielded slightly higher recognition accuracies than the ones
with 50 states, our preliminary experiments found that the results of
MMI training on the 50-state system were generally better than the
other two