Speaker adaptation based on pre-clustering training speakers