语音识别特征参数选择方法研究

Study of Feature Selection for Speech Recognition

摘要: 语音识别系统的性能与识别器所用的特征参数密切相关 ,用正交实验设计进行特征参数选择 ,提出一种系统性的实用的特征参数选择方法研究了特征参数排列顺序对识别系统最终识别率的影响 ,讨论了这种特征参数选择方法的可扩展性与目前常用识别器相比 ,新方法的词错误率下降了 37% 实验结果表明 ,正交实验设计用于特征参数选择是有效的 ,特征参数排列顺序对识别率的影响可以忽略

Abstract: As well known, the performance of the recognizer depends on the features of recognizer In this paper, a systematic and practical method of the features selection based on the orthogonal design is proposed The influence of the features’ order on the performance of the recognition system and the generalization of the features selection method are studied The word error rate of the method is reduced 37% when compared with that of the traditional recognizer The experiment results show that the application of orthogonal experiment design is effective in feature selection, and the influence of the features’ order on the performance of the recognition system can be ignored