Abstract: To improve on accuracy of Uyghur speech recognition, based on analysis of Uyghur characteristics, the framework of Uyghur speech recognition based on subword is developed for the first time. Pronunciation model of Uyghur word is given. How to build subword pronouncing dictionary, subword language model and acoustic model is described. Experiments are completed on a speech corpus and an unsupervised Uyghur word segmentation method is utilized to produce subwords. Experimental results show that Uyghur speech recognition based on subword can gain better recognition results.

Key words: Uyghur, word segmentation, subword unit, Hidden Markov Model(HMM), continuous speech recognition

摘要： 为提高维吾尔语语音识别的识别率，在分析维吾尔语特点的基础上，设计一种基于子字单元的维吾尔语语音识别总体结构，指出维吾尔语单词的发音模型，给出构建子字发音字典的方法，及其以子字单元为基础构建语言模型与声学模型的方法。在一个语音库上进行实验，采用一种非监督的词切分方法对维吾尔语单词进行词切分，生成子字。实验结果表明，基于子字单元的维吾尔语语音识别可以获得更好的识别结果。

关键词: 维吾尔语, 词切分, 子字单元, 隐马尔科夫模型, 连续语音识别

CLC Number: 

• TN912.34