Abstract In traditional Katz approach, the discount coefficients may be greater than 1, or can not be calculated in some serious conditions. To avoid the above problem, the smoothing idea in log domain of Simple GoodTuring combined with backoff model is proposed in the modified Katz approach. The proposed approach is further applied in speech recognition system based on lattice. The analysis of the effects on the structure and performance of lattice with different language model is given. Experiments show that compared with traditional Katz approach, the modified Katz approach can enhance the system performance. The best recognition rate can achieve 60.90% for the corpus from interview program.
2011, Vol. 24 
