Experimenting with a global decision tree for state clustering in automatic speech recognition systems