Domain adaptation of lattice-free MMI based TDNN models for speech recognition

Abstract

The recent proposed time-delay deep neural network (TDNN) acoustic models trained with lattice-free maximum mutual information (LF-MMI) criterion have been shown to give significant performance improvements over other deep neural network (DNN) models in variety speech recognition tasks. Meanwhile, the Kullback---Leibler divergence (KLD) regularization has been validated as an effective adaptation method for DNN acoustic models. However, to our best knowledge, no work has been reported on investigating whether the KLD-based method is also effective for LF-MMI based TDNN models, especially for the domain adaptation. In this study, we generalized the KLD regularized model adaptation to train domain-specific TDNN acoustic models. A few distinct and important observations have been obtained. Experiments were performed on the Cantonese accent, in-car and far-field noise Mandarin speech recognition tasks. Results demonstrated that the proposed domain adapted models can achieve around relative 7---29% word error rate reduction on these tasks, even when the adaptation utterances are only around 1 K.