Marcel Kockmann, Luciana Ferrer, Lukas Burget, Elizabeth Shriberg, and Jan Cernocky
We describe recent progress in the field of prosodic modeling for speaker verification. In a previous paper, we proposed a technique for modeling syllable-based prosodic features that uses a multinomial subspace model for feature extraction and within-class covariance normalization or linear discriminant analysis for session variability compensation. In this paper, we show that performance can be significantly improved with the use of probabilistic linear discriminant analysis (PLDA) for session variability compensation. This system does not require score normalization. We report an equal error rate below 7% on a NIST 2008 task. To our knowledge, this is the best reported result to date for a prosodic system for speaker recognition. Fusion of this system with a state-of-the-art acoustic baseline system yields 10% relative improvement in the new detection cost function (DCF) as defined by NIST.
|Publisher||IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP)|