摘要
本征音话者识别方法能够在一定程度上补偿因文本无关造成的语音类失配,但它并没有涉及另一个重要的失配因素——信道失配.本文提出了一种在本征音方法基础上补偿信道失配的方案.首先用本征音方法进行语音类失配补偿,然后采用WCCN(类内方差规整)进行信道失配补偿,从而得到经过语音类失配补偿和信道失配补偿的话者因子并将其作为话者模型,最后采用余弦评分方法进行性能评测.实验表明,本文方法在等误识率和最小检测代价函数上具有较好表现,同时本文方法对话者建模所需要空间较小.
The eigenvoice-based speaker verification method can compensate for voice mismatch in text- independent speaker verification applications, but it does not compensate channel mismatch, which also exerts a negative impact on the verification. Therefore channel mismatch compensation based on eigenvoice method was proposed. First, eigenvoice was adopted to compensate voice mismatch, then WCCN was applied to compensate channel mismatch. After these compensations, the speaker factor was computed and acted as speaker model. Based on the speaker factor model, Cos-score calculation was conveniently used to test verification operation. The experiment results show better performance, with an improvement by 22.85% at EER and 31.22% at MinDCF, while compared with GMM-UBM-SVM, an improvement was achieved by 9. 14% at MinDCF. Meanwhile, the new method needs less storage space, which benefits practical applications.
