语音识别中一种新的特征参数选择方法

Feature selection method based on orthogonal experiments for speech recognition

应用正交实验设计方案对语音识别中特征参数的选择进行了计划、分析与实验设计,目的是在大量的特征参数中选择出具有互补作用的特征参数。该方法有4个特点:1)实验方案的构造方法简单,而且得到的实验方案具有各特征参数搭配均衡的特点;2)实验结果分析方法计算简便,计算结果的物理含意明显;3)只需很少的比较实验就可以找到识别性能较好的特征参数组合;4)已有的实验结果对后续实验方案的设计有很好的指导作用。实验结果表明正交实验设计用于特征参数选择是有效的。

A new feature selection method was developed for speech recognition using orthogonal experiment design. The purpose was to identify which features cooperate. This method has 4 advantages: 1) The experiment scheme is simple and satisfies the orthogonality condition. 2) The analysis of the experiment results is convenient and the physical meaning of the analysis result is obvious. 3) This method can find feature combinations with better recognition performance using fewer experiments. 4) The experimental results provide useful information for additional experiments. The experimental results show that orthogonal experiment design facilitates feature selection.