Towards End-to-End Speech Recognition with Deep Convolutional Neural Networks