论文部分内容阅读
针对传统的Katz方法会出现折扣系数大于1或者无法计算的情况,将SimpleGood-Turing中对出现次数对数域的平滑思想用于Katz方法中,结合回退模型,提出一种改进的Katz算法.将该方法应用于基于Lattice的语音识别系统中,分析不同语言学模型对生成的Lattice结构的影响和基于该结构的识别性能的影响.实验表明,应用改进的Katz算法针对访谈节目的识别性能最高可以达到60.90%,优于传统Katz方法.
For the traditional Katz method, if the discount coefficient is greater than 1 or can not be calculated, the smoothing idea of the number of occurrences in SimpleGood-Turing is applied to the Katz method. Combined with the backoff model, an improved Katz algorithm is proposed. The method is applied to Lattice-based speech recognition system to analyze the influence of different linguistic models on the Lattice structure generated and the recognition performance based on the structure.Experiments show that the improved Katz algorithm has the highest recognition performance for the talk shows Up to 60.90%, outperforming the traditional Katz method.