Dur = model size proportional to duration

• Analysis indicates the existence of an optimal topology larger acoustic units like syllables

• Use information theoretic measures to learn the structure

• The optimal model size is highly task-dependent so a data-driven approach is needed