论文部分内容阅读
随着现代社会信息的全球化,双语以及多语混合的语言现象日趋普遍,随之而产生的双语或多语语音识别也成为语音识别研究领域的热门课题。在双语混合语音识别中,主要面临的问题有两个:一是在保证双语识别率的前提下控制系统的复杂度;二是有效处理插入语中原用语引起的非母语口音现象。为了解决双语混合现象以及减少统计建模所需的数据量,通过音素混合聚类方法建立起一个统一的双语识别系统。在聚类算法中,提出了一种新型基于混淆矩阵的两遍音素聚类算法,并将该方法与传统的基于声学似然度准则的聚类方法进行比较;针对双语语音中非母语语音识别性能较低的问题,提出一种新型的双语模型修正算法用于提高非母语语音的识别性能。实验结果表明,通过上述方法建立起来的中英双语语音识别系统在有效控制模型规模的同时,实现了同时对两种语言的识别,且在单语言语音和混合语言语音上的识别性能也能得到有效保证。
With the globalization of information in modern society, bilingual and multilingual mixed language phenomenon is becoming more common. The consequent bilingual or multilingual speech recognition has also become a hot topic in the field of speech recognition. In bilingual mixed speech recognition, the main problems faced by the two: First, to ensure that the bilingual recognition rate under the premise of the complexity of the control system; the second is to effectively handle the original language of the insert language caused by non-native accent phenomenon. In order to solve the problem of bilingual mixing and reduce the amount of data required for statistical modeling, a unified bilingual recognition system is established by phoneme hybrid clustering. In the clustering algorithm, a new two-pass phoneme clustering algorithm based on confusion matrix is proposed, and the method is compared with the traditional clustering method based on acoustic likelihood criteria. For the non-native speech recognition in bilingual speech Low performance, a new bilingual model correction algorithm is proposed to improve the recognition performance of non-native speech. The experimental results show that the Chinese-English bilingual speech recognition system established by the above method can effectively control the size of the model and at the same time realize the recognition of the two languages at the same time, and the recognition performance in single-language speech and mixed-language speech can also be obtained Effective guarantee.