Speech Processing Based on Large Heterogeneous Segments