The application of Missing Data Techniques (MDT) to increase the noise robustness of HMM/GMM-based large\r\nvocabulary speech recognizers is hampered by a large computational burden. The likelihood evaluations imply\r\nsolving many constrained least squares (CLSQ) optimization problems. As an alternative, researchers have proposed\r\nfrontend MDT or have made oversimplifying independence assumptions for the backend acoustic model. In this\r\narticle, we propose a fast Multi-Candidate (MC) approach that solves the per-Gaussian CLSQ problems\r\napproximately by selecting the best from a small set of candidate solutions, which are generated as the MDT\r\nsolutions on a reduced set of cluster Gaussians. Experiments show that the MC MDT runs equally fast as the\r\nuncompensated recognizer while achieving the accuracy of the full backend optimization approach. The\r\nexperiments also show that exploiting the more accurate acoustic model of the backend does pay off in terms of\r\naccuracy when compared to frontend MDT.
Loading....