Fast Adaptation on Deepmixture Generative Network Based Acoustic Modeling

Abstract

Deep neural network (DNN) has achieved the state-of-the-art performance in automatic speech recognition (ASR). However, the meaning of parameters and neurons are hard to be interpreted in DNNs, which makes the regularizations and adaptation of DNNs difficult. In this work, we aim to do effective and efficient adaptation on a more interpretable model, deep mixture generative network (DMGN). Adapted means are first proposed to perform adaptation for DMGN. The speaker-dependent means are estimated in an unsupervised adaptation mode. Moreover, discriminative linear regression (DLR) is proposed to estimate more robust speaker-dependent means when lack of adaptation data. We evaluate our proposed methods on 50-hour subset of Switchboard. Experiments reveal that all proposed methods are better than speaker independent baseline, and a slight performance improvement is obtained compared with LHUC. In addition, we project the Gaussian mean of one senone and all inputs aligned to this senone to a 2 D graph. The illustration shows that after applying DLR, the mean is indeed transferred from an average point to the speaker specific center, which demonstrates the better explanation of DMGN again.