Active Learning for Speech Recognition: the Power of Gradients