Abstract: A measure was studied to solve the problem of lacking corpus for a Chinese voice search system. First, semantic class induction was done from the existing corpus using a novel similarity measure which is based on cooccurrence probabilities. Clustering with the new similarity measure outperformed that with the widely used distance measure based on Kullback-Leibler divergence in precision, recall and F₁ evaluation. Then corpus was generated using induced semantic classes and structures. Finally, generated corpus were used to do language model adaptation and improve the result of character recognition from 85.2% to 91%. The experiment results show that the problem of lacking corpus for a new voice search system can be solved through semantic class induction, template generation then in-domain data generation.