Statistical grammars can help us even more. In our example up to this point, we used a simple grammar, a pronoun, verb and noun. And that placed a strong limit of were we started and ended in our Viterbi trellis. >> But in real life, language is not so well segmented. Instead, we can record large amounts of language and affirm in the fraction of times need follows I, versus want following I, or want following we. >> We can then use these probabilities to help bias our recognition based on the expected distribution of the co-occurrence of these signs. In practice, using a statistical grammar divides the error rate by another factor of 4. We started with a fundamental error rate of E. When we used context training, we divide in half. And now with statistical grammars, we’re dividing it by a factor of 4 again. So, basically, we end up with our error rate divided by 8. This trick is one of the secret weapons in my research. I look for problems that might have a language-like structure and then apply these techniques to get my error rate down to something reasonable.