Curriculum Learning based Probabilistic Linear Discriminant Analysis for Noise Robust Speaker Recognition

This study introduces a novel Curriculum Learning based Probabilistic Linear Discriminant Analysis (CL-PLDA) algorithm for improving speaker recognition in noisy conditions. CL-PLDA operates by initializing the training EM algorithm with cleaner data (easy examples), and successively adds noisier data (difficult examples) as the training progresses. This curriculum learning based approach guides the parameters of CL-PLDA to better local minima compared to regular PLDA. We test CL-PLDA on speaker verification task of the severely noisy and degraded DARPA RATS data, and show it to significantly outperform regular PLDA across test-sets of varying duration.