Exactly. Underfitting would be if the model doesn't pick up on the fact that outro silence is labeled differently from regular silence and transcribes them the same