摘要
为了解决基于词语的维吾尔语语音识别系统集外词过多的问题,采用形态分析生成的语素或数据驱动切分生成的统计子词代替词语作为识别系统的词典单元。在此基础上,提出一种根据语素识别系统和统计子词识别系统在声学模型训练数据上的音素错误率差别选择词语最佳分解结果,从而构建语素-统计子词联合词典的方法。在维吾尔语电话交谈式语音识别任务上比较各个系统的性能。实验结果表明,语素或统计子词的运用能有效缓解词语系统集外词过多的问题。与词典大小为200K的词语系统相比,55K的语素-统计子词联合系统使测试集上的音素错误率从45.4%下降到43.8%。
To handle the high out-of-vocabulary （OOV） rate problem of the word-based Uyghur speech recognition system, morphemes, which are obtained from morphological parsing, or statistical sub-words, which are leaned through data-driven splitting, are selected as the lexicon units. Then, according to the phoneme error rates （PERs） difference on the acoustic training data, we build a hybrid vocabulary with morphemes and statistical sub-words by selecting the optimal splitting re- sult for each word. Performances of these systems are compared in the conversational telephone speech transcription task. Experiment results suggest the use of morphemes or statistical sub-words can alleviate the OOV problem effectively. Com- pared to a 200K word-based system, a 55K hybrid system reduces the PERs from 45.4% to 43.8% on the test set.
出处
《重庆邮电大学学报（自然科学版）》
CSCD
北大核心
2013年第3期391-396,共6页
Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)
基金
国家自然科学基金(10925419
90920302
61072124
11074275
11161140319
91120001
61271426)
中国科学院战略性先导科技专项(XDA06030100
XDA06030500)
国家863计划(2012AA012503)
中科院重点部署项目(KGZD-EW-103-2)~~
关键词
黏着语
语音识别
集外词
词语分解
分解方法联合
agglutinative language
speech recognition
out-of-vocabulary
word decomposition
hybrid decomposition
