Revisiting the Entropy Semiring for Neural Speech Recognition