Improved model parameter compensation methods for noise-robust speech recognition

We study model parameter compensation methods for noise-robust speech recognition based on continuous density (CDHMM). First, we propose a modified PMC method where the adjustment term in the model parameter adaptation is varied depending on the mixture components of the HMM to obtain more reliable modeling. A state-dependent association factor that controls the average parameter variability of Gaussian mixtures and the variability of the respective mixtures is used to find the final optimum model parameters. Second, we propose a re-estimation solution of the environmental variables with additive noise and spectral tilt based on the expectation-maximization (EM) algorithm in the cepstral domain. The approach is based on the vector Taylor series (VTS) approximation. In our experiments on a speaker independent isolated Korean word recognition, the modified PMC show a better performance than the Gales´(1992) PMC and the proposed VTS is superior to both of them