一种有效的汉语连呼数字识别算法

模板匹配技术用于连呼数字识别,其存在的主要问题是巨大的计算量。本文在分析了汉语连呼数字发音参数特征的基础上,提出了一种集预分割和非预分割方法为一体的连呼数字识别算法,该算法的运算量与多级匹配方法,相比有了大幅度下降,本文同时还探讨了语音参数考模板的建立方法对识别性能的影响及通过加入基音信息进一步提高识别性能的途径。

AN EFFECTIVE ALGORITHM FOR CHINESE CONTINUOUS DIGIT RECOGNITION

Enormous computation is one of impartant problems of template matching used in connected word recognition. In this paper, we present a new hybrid-technique with incorporating aspects from both segmenting and nonsegmenting methods for Chinese digit string recognition based on the parameter analysis of digit speech. Compared with two-level or multi-leyel DP-mstching approaches, the computational complexity is largely reduced by dividing the reference templates into two groups. The effects of three kinds of creating reference patterns are disscussed. Final, pitch discriminative features are first applied to "post-correction" in order to improve recognition accuracy.