Universal Adversarial Perturbations for Speech Recognition Systems