Discriminative training of HMM using MASPER procedure

The main focus of the article is on incorporating discriminative training into MASPER multilingual training procedure by some necessary modifications. Next the performance of discriminative rules like Maximal Mutual Information (MMI) and Minimal Phone Error (MPE), application of I smoothing technique, setting up convergence parameter, benefits of discriminative training for different hidden Markov models (HMM), etc. are tested and evaluated. Moreover an overview of discriminative training strategies and their relations to the classical Maximum Likelihood (ML) estimation is given. All experiments have been accomplished on Slovak part of MobilDat training database that contains wide range of noises and specific GSM distortions. Achieved results show that discriminative training if properly adjusted can improve performance over ML training on average by 5% depending on the model complexity, training strategies and deployment scenarios. Finally, MPE when properly set may outperform MMI, however it is prone to higher sensitivity to the set parameters, used models and application domain.