广播新闻语音的关键词检测系统.pdfVIP

第28卷第l2期 通 信 学 报 、，0ll28 No．12 2007年l2月 ．1oU／TIal on Communications December 2007 广播新闻语音的关键词检测系统 张鹏远，邵健，赵庆卫，颜永红 (中国科学院 声学所中科信利实验室，北京 100080) 摘 要：提出了一种两阶段的关键词检测策略，关键词表可即时配置、即时使用。与以往不同的是，解码是在整 个声学空间上进行的，而不仅仅是关键词所构成的声学空间。首先，利用维特比解码得到最优候选的音素序列。 然后，通过关键词列表，可以产生候选的关键词。最后，采用了基于声学模型的置信度与基于网格的置信度相结 合的方法来降低虚警。实验证明，新的框架下关键词检测的性能得到了显著的提高。 关键词：关键词检测；音节识别器；置信度计算 中图分类号：TP391．42 文献标识码：B 文章编号：1000—436X(2007)12—013l一05 Keyword spotting system for broadcast news ZHANG Peng—yuan，SHAO Jian，ZHAO Qing—wei，YAN Yong—hong (ThirddT Speech Lab，Institute ofAcoustics，ChineseAcademy of Sciences，Beijing 100080，China) Abstract：A two—step keyword spotting strategy was presented．This strategy allowed to change keyword list conven- iently．Different from the previous systems，search space was generated based on all Chinese syllables，not specifically for keywords．Phoneme recognition was performed without any lexical constraints．With 1-best phoneme sequence and keyword list，which generated keyword hypotheses．At last，two confidence measures were introduced adopted in the sys- tem：one based on acoustic model and the other based on phoneme lattice．For a decoded speech frame aligned to an HMM state，the acoustic confidence was calculated．Th e l~tice confidence made use of phoneme lattices generated by a phoneme recognizer．Th ese two confidence measures were combined using a weighting factor to obtain a hybrid confi— dence as they had different dynarnic scales．Experiments show that the proposed algorithms significantly improve the system perform ance on the broadcast news task． Key words：keyword