Discriminative speaker adaptation in Persian continuous speech recognition systems

Abstract

In this paper, the use of discriminative criteria such as minimum phone error (MPE) and maximum mutual information (MMI) is investigated for discriminative training HMM models for Persian speech recognition system. Discriminative training criteria have been successfully used to train acoustic models, so these criteria are expected to improve the estimation of linear transforms for speaker adaptation. MPE criterion is used to estimate the discriminative linear transforms (DLTs) for mean transforms. Experiments on Farsdat corpus show considerable improvements of discriminative training against ML trained models and MPE training outperforms MMI training on test data. Furthermore, MPE-based DLT reduces the word error rate in comparison to MLLR adaptation.