The metric used to measure the goodness of the network is the so-called perplexity. The precise definition requires a lot of math, but it is roughly the average number of choices that the model may have after each word; if we consider spoken English, about one million words, currently the best network models have perplexity 247, a really big number of branching for each word.