汉语音节混淆网络的生成与重打分算法研究

Research on Algorithm of Mandarin Chinese Syllable Confusion Network Generation and Rescoring

针对目前混淆网络生成算法速度与精度不能兼顾的不足,提出一种新的汉语音节混淆网络生成的方法.本算法采用类似轴对齐算法,对音节网格每次提取一条局部路径与参考路径对齐,根据每次对齐路径与参考路径长度不同,采用不同的策略生成混淆网络,并在生成混淆网络之后对其应用一种新的解码框架进行重打分.实验表明,该算法生成的混淆网络精度较高,时间复杂度优于轴对齐算法,且重打分后的识别率有显著提高.

As the current generation algorithm of confusion network cannot take speed and accuracy into account at the same time,this article proposes a new method to generate syllable confusion network of Mandarin Chinese.Similar to Pivot Alignment Algorithm,this method choose a local path from syllable lattice every time,and then according to the different length of the path and reference path,adopt different strategy align this path to the reference path generating confusion network.After generate confusion network,a new decoding frame is used to rescore them.The experimental results show that our algorithm can obtain better precision and complexity than Pivot Alignment Algorithm.Meantime,syllable accurate is improved obviously after rescoring.