基于上下文相关置信度打分的语音确认方法

口语对话系统中,集外词的存在会引起很多识别错误,为了有效地发现并拒绝集外词,提高系统性能,研究利用置信度打分进行语音确认的方法,发现并拒绝识别错误。提出上下文相关的置信度特征,充分考虑当前待确认词与其前序词和后序词之间的相关性。实验结果表明:上下文相关的置信度特征能够很好地提高拒识性能,对符合识别文法的句子,错误拒绝率为2.5%或5%时,对比没有使用上下文相关的置信度特征时,错误接受率分别下降了29%和36%;基于置信度打分的语音确认策略在拒识性能上优于系统已有的在线垃圾模型。

Confidence scoring using context dependent features for word verification

In a spoken dialogue system,out-of-vocabulary words will lead to recognition errors.A confidence scoring strategy for word verification was developed to detect and reject such recognition errors.Context-dependent confidence features were used based on the information in the previous and following words.The results show that the context dependent confidence features greatly improve the word verification performance.For in-grammar sentences,the false acceptance rate is reduced by 29% at a false rejection rate of(2.5%) and by 36% at a rejection rate of 5%,compared with a system not using the context-dependent confidence features.The verification strategy based on confidence scoring also has better performance than the verification method using the online filler model.