用于语音识别置信度的发音特征各维度分析和子集优化  

Analysis and subset selection of articulatory features for speech recognition confidence measures

作　　者：孙艳庆[1] 张晴晴[1] 周瑜[1] 赵庆卫[1] 颜永红[1] 

机构地区：[1]中国科学院声学研究所中科信利语音实验室,北京100190

出　　处：《声学学报》2011年第3期339-348,共10页Acta Acustica

基　　金：国家科技支撑计划(2008BAI50B03);国家自然科学基金(10925419;90920302;10874203;60875014)资助项目

摘　　要：提出了基于发音特征单个维度的置信度算法,并基于此对发音特征的各个维度展开分析。分析不仅验证了融合的必要性,同时也展示了发音特征各维度之间以及和隐马尔可夫模型之间的大量冗余。为了去除冗余,提出了用子集选择的方法进行优化。对比所有都用的情况,基于发音特征紧凑子集的语音识别置信度估计,在等错率上取得了12.7%的相对下降。把经过优化后的基于发音特征的语音识别置信度估计和基于隐马尔可夫模型的语音识别置信度进行融合,在保持集内识别率不损失的前提下,显著提高了语法外输入测试的拒识性能:在相同参数下,在开发集和测试集上分别取得了34%和35.3%的显著改善。Different articulatory properties are analyzed in terms of confidence measures using a separate AF-based confidence calculation method.The analysis not only verifies the necessity of assembly,but also demonstrates a great deal of redundancies between the articulatory properties and HMM.In order to reduce the redundancy,a subset selection method is proposed.Experiments are designed to verify the above assumptions.Compared with all used together,the confidence measures based on the compact subset of articulatory features get a relative decrease of 12.7%for EER.The optimized AF-based confidence is finally combined with the HMM-based confidence,and increases rejection rate for the out of vocabulary tests with no accuracy loss of the in vocabulary tests,and the relative improvement is 34%on the development sets and 35.3%on the testing sets with the same parameters.

关 键 词：置信度估计 语音识别 特征 发音 维度 优化 子集 隐马尔可夫模型 

分 类 号：TN912.34[电子电信—通信与信息系统]