1 paper · 1 filter
Wuyang Liu, Yanzhen Ren
Most audio tagging models are trained with one-hot labels as supervised information. However, one-hot labels treat all sound events equally, ignoring the semantic hierarchy and pro…