Modified-prior plda based speaker recognition system

To reduce the negative impact on the performance of speaker recognition systems due to the duration mismatch between enrollment utterance and test utterance, a modified-prior PLDA method is proposed. The probability distribution function of i-vector was modified by incorporating the covariance matrix with duration of each utterance of each speaker during the PLDA training, which further improved the discriminant capability of speaker classification. To evaluate the robustness of the proposed modified-prior PLDA method, extensive experiments were performed on NIST SRE10 core-core task (female part) in duration mismatch conditions and NIST 2014 i-vector machine learning challenge. Experimental results demonstrated that the duration-based modified-prior PLDA method achieved better performance compared with the traditional PLDA.

Duke Scholars

Author Ming Li DKU Faculty