|
|
论文题目 |
中英双语混合语音识别的研究 |
论文题目(英文) |
Development of a Mandarin-English Bilingual Speech Recognition System |
作者 |
张晴晴 |
发表年度 |
2008 |
卷 |
20 |
期 |
4 |
页码 |
391-396 |
期刊名称 |
重庆邮电大学学报:自然科学版 |
摘要 |
介绍了针对歌曲检索中出现的中英混合现象所开发的中英双语识别系统。在双语混合语音识别中 ,主要面临的 2个问题: ① 在保证双语识别率的前提下控制系统的复杂度; ②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 ( TCM) ,并将该方法与基于声学似 然度准则的聚类方法进行了比较。实验结果表明:利用 TCM进行音素聚类的识别性能优于基于声学似然度音素聚类的性能 ,最终得到的中英双语识别系统在纯英文测试集上的短语错误率 ( PER)相对基线单英文识别系统下降 7 . 19%;在双语混合测试集上 PER相对基线混合模型下降 13 . 78%;同时在纯中文测试集上保持了基线单中文识 别系统的性能。 关键词:双语识别;聚类算法;自适应 |
摘要_英文 |
TheMandarin2 English bilingual s peech recogniti on system which has been devel oped for theMandarin English phenomenon in song retrieval is introduced . The main difficulties to handle the bilingual speech recogniti on for real world applicati on are focused on two as pects : the first is to balance the performance on inter and intra - sentential language s witching and to reduce the comp lexity of the bilingual speech recogniti on system; the second is to effectively dealwith the matrix language accents in embedded language . In order top rocess the intrasentential language s witching and reduce the amount of data required to robustly esti mate statisticalmodels, instead of using two separatemonolingualmodels for each language, acompact single set of bilingual acoustic model derived by phone setmerging and clustering is devel oped . Hence,a novel Two pass phone clusteringmethod based on Confusi onMatrix (TCM) is presented and compared with the log like lihood measure method . Experi ments testify that TCM can achieve better performance . The phrase err or rate (PER) of MESRS for English utteranceswas reduced by 7 . 19% relatively compared to the baseline monolingual English system while the PER onMandarin utteranceswas comparable to that of the baseline monolingualMandarin system. The perfor mance for bilingual utterances achieved 13 . 78% relative PER reducti on . Key words: bilingual speech recogniti on; clustering algorithm; adap tati on |
|
|
|