文章基本信息

标题：Heuristic Improvements of the HMM Use in Isolated Word Speech Recognition
本地全文：下载
作者：D. Dimov ; I. Azmanov
期刊名称：Cybernetics and Information Technologies
印刷版ISSN：1311-9702
电子版ISSN：1314-4081
出版年度：2007
卷号：7
期号：3
出版社：Bulgarian Academy of Science
摘要：The paper presents the authors’ experience with HMMs (Hidden Markov Models) used for isolated word speech recognition in Bulgarian. Two methods provoked by experiments are discussed, namely: (i) the precise quantization of Gaussian probability density function (G-pdf) modeling the HMM states’ output, and (ii) a method for averaging a set of HMMs trained for different versions of a given word. A universal threshold is evaluated for switching between the “large” and “slim” models of G-pdf defined here. Experimental results of the threshold usage for averaging of HMMs are also reported.rwicz
关键词：Speech recognition; HMM; Gaussian pdf; speech cepstrum.