An orthogonal GMM based speaker verification system

This paper describes a new speaker verification system based on orthogonal Gaussian mixture modeling (GMM) techniques combined with maximum a posteriori (MAP) adaptation. In most of the GMM based speaker verification systems, the variance of each component is constrained to be diagonal for its computational simplicity. However, this approximation inevitably introduces performance degradation. To reduce this deleterious effect, we combine the orthogonal transform and MAP adaptation to build our system. The transform matrix is shared according to which mixture of the Speaker Independent (SI) model the current mixture component is adapted from. This gives significant system performance improvement at the cost of introducing only a small additional computational load.