Classifiers Ensemble of HMM and d-Vectors in Biometric Speaker Verification

This paper presents a novel approach on text-dependent biometric speaker verification (SV) based on the ensemble of two feature extraction and classification processes using Hidden Markov Models in a Universal Background Model framework (HMM-UBM) and d-Vectors derived from a Deep Learning Network (DNN) structure. Once the individual SV systems are trained, a third classifier is trained/tuned over individual test scores in the same dataset using three different approaches for comparison purposes: Multilayer Perceptron (MLP), Support Vector Machine (SVM) with three different kernels, and a Fuzzy Inference System (FIS). Obtained results over a proprietary speech database in Spanish, indicate an improved performance, providing an Equal Error Rate (EER) within the range of 0.7%-2.54% when classifier ensembles are used, versus an EER of 3.6% and above obtained in average with individual classifiers. Results in detail corresponding to comparison of the several approaches used in this experimental work are further described.