基于VTLN和频谱倾斜补偿的特征提取

With the development of artificial intelligence, speech recognition has been paid more and more attention. For speech recognition, speech feature should have characteristics of intraclass convergence and interclass discretization, but due to the influence of physiological pronunciation differences, speech features in addition to containing semantic information but also contains personalized information and the personalized information fuzzy the boundary between the features and reduce the distinguishing ability of features. Based on the research of the pronunciation physiological differences this paper has proposed a feature extraction algorithm based on vocal tract length normalization and spectraltilt compensation which comprehensively solves the effect of pronunciation differences on the position and amplitude of formant, experimental results show the effectiveness of the algorithm.?????