摘要： 语速和插入删除错误是导致自动电话转接系统发生错误的重要原因。该文给出一种基于子词似然比(LLR)和子词驻留特征融合的语音确认方法减少上述错误。提出基于最小分类错误准则方法求取子词特征融合参数。电话转接系统实验结果表明，采用子词驻留特征能有效提高语音确认效果，与LLR方法相比，名称关键词的等错误率下降3.35%，数字串关键词的等错误率下降4.05%。

关键词: 语音确认, 似然比, 子词驻留概率, 最小分类错误

Abstract: Speech rate and insert/delete error are the main sources of recognition error occurrence. A speech verification method based on subword LikeLihood Ratio(LLR) and subword duration features fusion is presented to decrease above errors. Minimum Classification Error(MCE) criterion based estimation algorithm is proposed to estimate subword feature fusional parameters. Compared with the LLR method, experiment on telephone switch system shows that the subword duration feature can explicitly enhance the system performance with the relative Equal Error Rate(EER) reduction by 3.35% in name keyword verification and 4.05% in digital string keyword verification.□

Key words: speech verification, LikeLihood Ratio(LLR), subword duration probability, Minimum Classification Error(MCE)

中图分类号: 

• TP391.42