Exploiting sparsity in stranded hidden Markov models for automatic speech recognition

We have recently proposed the stranded HMM to achieve a more accurate representation of heterogeneous data. As opposed to the regular Gaussian mixture HMM, the stranded HMM explicitly models the relationships among the mixture components. The transitions among mixture components encode possible trajectories of acoustic features for speech units. Accurately representing the underlying transition structure is crucial for the stranded HMM to produce an optimal recognition performance. In this paper, we propose to learn the stranded HMM structure by imposing sparsity constraints. In particular, entropic priors are incorporated in the maximum a posteriori (MAP) estimation of the mixture transition matrices. The experimental results showed that a significant improvement in model sparsity can be obtained with a slight sacrifice of the recognition accuracy.