Hybrid GMM-UBM and SVM based systems for text-independent speaker verification in noisy environment

This paper proposes a hybrid Gaussian Mixture Model-Universal Background Model (GMM-UBM) based speaker verification system for mobile devices in variable noisy environments. We applied a Support Vector Machine (SVM) classier in two GMM-UBM based text-independent speaker verification systems to improve the performance. In the first system, both hypothesized speaker and UBM models are trained independently using Maximum Likelihood (ML) estimation (GMM-UBM system). In the second, the hypothesized speaker model is derived by the UBM using Maximum a posteriori (MAP) adaptation (ADAP-GMM-UBM system). A Voice Activity Detection (VAD) algorithm based on bispectrum analysis is also applied to improve the performance. The methods are tested using the MIT Mobile Device Speaker Verification Corpus varying the number of the distributions in the hypothesized speaker models. The use of a single speaker-independent SVM with linear kernel function achieved an improvement up to 23.89% in the performance of the best GMM-UBM based system.