No. Because there would have been indtances in the data where silence was labelled correctly. But the model couldnt handle the null case, so it over fit on the outros. But generally it fit on the random error in the label of the null feature. Which is what overfitting is