OPTIMIZING VARIANCE FLOORING IN HMM-BASED SPEAKER VERIFICATION
Håkan Melin · 1998
Adaptive variance flooring has previously been suggested as a method to help preventing over-fitting of variance parameters during training of HMMs. With this method, the average variance over some calibration data set is multiplied by a constant variance flooring factor to produce the variance floor. In this paper, experiments with adaptive variance flooring is presented, in which we try to optimize the value of the flooring factor, and we look at the database dependence of that optimal value. A small modification to the original method is also suggested, in which the variance floor is made word-dependent. 1. INTRODUCTION For convenient use of an automatic speaker verification (ASV) system, the system should require only little enrollment data with maintained good recognition accuracy. One problem with using small training data is the risk of over-training, that is, parameters of the speaker model are over-fitted to the particular training data. For (Gaussian) HMM-based speaker mode...