This paper proposes DCKD, a privileged knowledge distillation framework for target sound extraction that regulates the amount and flow of target information via neural codec and disentangled representation learning.