not sure if someone also read the paper, in this paper they picks pseudo label via WER which means it already have the ground-truth, why not using those ground-truth as training goal but use pseudo label as goal?
wondering they must do the experiment but not reveal in the paper