Combination of improved Katz and mutual information for speech recognition based on Lattice

In the language model based on Chinese, only the number of occurring count of n-gram can not account for the reliability. Some n-grams have strong meaning in Chinese although their occurring count is low in training data. Combined the number of count with mutual information, whether the syllables in n-gram are highly associated can be better described. Further, in most smoothing approaches, the discount idea is widely adopted, that is for less reliable n-grams, the probabilities calculated by training data may be over-estimated, so they should be discounted. In this paper, not only the over-estimated n-grams are discounted, but also the most reliable n-grams which may be under-estimated during training procedure are enhanced. For all these modification is based on Katz which is the most commonly used method in speech recognition, the proposed approach is named as improved Katz. From the experiment results, it can be drawn that the performance of improved Katz is better that only Katz.