Reduced semi-continuous models for large vocabulary continuous speech recognition in Dutch

Due to the decoupling between the set of Gaussians and other hidden Markov model (HMM) parameters, semi-continuous-density HMMs (SC-HMMs) have more possibilities than continuous-density HMMs (CD-HMMs) to match the number of parameters in the model to the available training data. The computational load of the SC-HMMs, however, is huge compared to the load of their continuous counterparts, because of the large mixture-weighting vector and because of the fact that, for each frame, all Gaussians have to be evaluated. This paper describes the different steps taken to reduce the computational load of the SC-HMMs, resulting in faster and better models.