Fusion of SNR-Dependent PLDA Models for Noise Robust Speaker Verification Results Methods Hard-Decision SNR-Dependent PLDA Xiaomin PANG and Man-Wai MAK Dept. of Electronic and Information Engineering, The Hong Kong Polytechnic University Motivation of Methods SNR Distribution in NIST 2012 SRE Introduction Motivation In practical speaker verification, additive and convolutive noise cause mismatches between training and recognition conditions, degrading the performance. Methods A fusion system that combines a multi-condition PLDA model and a mixture of SNR-dependent PLDA models is proposed to make the verification system noise robust. Key Findings Results on NIST 2012 SRE show that (1) the SNR- dependent PLDA models can reduce EER, (2) the fusion system is more robust than the conventional i-vector/PLDA systems under noisy conditions, and (3) the SNR-dependent PLDA models are insensitive to Z-norm parameters. Decision Weights These histograms suggests that the test utterances exhibits a wide range of SNR. Soft-Decision SNR-Dependent PLDA System 1: Fusion of SNR-independent and hard-decision SNR-dependent PLDA System 2: Fusion of SNR-independent and soft-decision SNR-dependent PLDA System 3: Fusion of SNR-independent, hard- and soft- decision SNR-dependent