Exploiting Deep Neural Networks for Detection-Based Speech Recognition