【作者】 徐明； 胡瑞敏； 黄云森；

【Author】 Xu Ming 1,2+, Hu Rui-min 2, Huang Yun-sen2 1(Computer Science Institute, Wuhan University,Wuhan 430000, China) 2(Information Center, Shenzhen University, Shenzhen 518060, China)

【机构】 武汉大学计算机学院； 深圳大学信息中心；

【摘要】 介绍了基于语音音素识别技术的语音评价实现方法,包括Tri-Phone建模、分词、Viterbi搜索、音素识别、路径回溯等过程;还介绍了利用语音能量、发音节奏和基频参数进行语音综合评价的做法;最后,介绍了系统的运行结果。更多还原

【Abstract】 In order to guide people to study spoken language,research on how to assess people’s pronunciation is of real meaning and practical value.Hence,we study the pronunciation assessment techniques mainly based on natural speech phoneme recognition,the idea is as follows:Firstly,the phonene unit sequence and their change points are recognized;Secondly,how phoneme sequence change and how long they last can be derived; Thirdly, the effect of speech strength and tone on pronunciation are added;Finally, the complete information of pronunciation assessment in continuous speech can be represented. In this paper we explored the pronunciation assessment techniques for continuous natural speech based on speech phoneme recognition. The method based on speech recognition consists of five major parts, namely, triphone modeling, search space organization, Viterbi searching, speech phoneme recognition, and path tracking back. Speech rhythm, energy and pitch parameters are also used to evaluate the similarity between the evaluative speech and the standard speech. We use HMM log-likelihood, rhythm, pitch, and magnitude as the features to generate a similarity score. At the end of this paper, outputs of such a system are illustrated.Our research and experiments indicate that it is effective to do pronunciation assessment based on speech phoneme recognition. For a concrete application of pronunciation assessment,we will prepare the accompanying vocabulary and organize a small or medium search space in order to make the system perform quickly and efficiently.更多还原