Speaker adaptation using regression classes generated by phonetic decision tree-based successive state splitting.