One-Speaker Detection – Limited Data: The WCL-1 System

Our text-independent Probabilistic Neural Network (PNN)-based speaker verification system, referred to as WCL-1, a participant in the One-Speaker Detection Task of the 2003 NIST Speaker Recognition Evaluation (SRE), is presented here. The WCL-1 system is built on modular structure with a distinct PNN for each enrolled user. A gender-dependent universal background model, representing the world of non-users, is employed to counterbalance the scores produced by the users’ models. The year’s 2003 version of the WCL-1 system is based on our year’s 2002 system, which is considered here as a baseline. In the year’s 2003 WCL-1 system, prosodic speech features were added to the short-term spectrum, represented by Mel-scale-like cepstral coefficients (MFCCs), to account for the speaking style of the users. The frame energy and fundamental frequency of the speech signal are appended to the MFCCs to form the speech feature vector for every voiced speech frame.