Modeling prosodic features with probabilistic linear discriminant analysis for speaker verification

The use of continuous prosodic features is introduced into speaker verification. The whole prosodic contour is segmented over fixed-frame long with fixed-frame shift and the prosodic features are extracted using a basis consisting of Legendre polynomials. They are then modeled using the i-vector based approach followed by probabilistic linear diseriminant analysis (PLDA) to compensate for speaker and channel variability effects in the space of i-vectors. The experiments are carried out on the noisy conditions which are generated based on the extended condition 5 of the NIST 2010 Speaker Recognition Evaluation (SRE) dataset. The experimental results indicate that the prosodic features are noise-robust and the fusion of the prosodic features and the traditional Mel Frequency Cepstral Coefficients (MFCCs) can make significant performance improvement. Compared to the MFCCs system alone~ the fusion can provide up to 9% and 11% relative improvement respectively in equal error rate (EER) and minimum detection cost function (minDCF).