A Bayesian Network Approach for Combining Pitch and Reliable Spectral Envelope Features for Robust Speaker Verification

Abstract

In this paper, we provide a new approach in the design of robust speaker verification in noisy environments using some principles based on the missing data theory and Bayesian networks. This approach integrates high-level information concerning the reliability of pitch and spectral envelope features in missing feature compensation process in order to increase the performance of Gaussian mixture models (GMM) of speakers. In this paper, a Bayesian network approach for modeling statistical dependencies between reliable prosodic and spectral envelope features is presented. Within this approach, conditional statistical distributions (represented by GMMs) of the features are simultaneously exploited for increasing the recognition score, particularly in very noisy conditions. Masked by noise data can be discarded and the Bayesian network can be used to infer the likelihood values and compute the recognition scores. The system is tested on a challenging text-independent telephone-quality speaker verification task.