Ensemble knowledge distillation of self-supervised speech models