Exploring GMM-derived Features for Unsupervised Adaptation of Deep Neural Network Acoustic Models

International audienceIn this paper we investigate GMM-derived features recentlyintroduced for adaptation of context-dependent deep neural networkHMM (CD-DNN-HMM) acoustic models. We present an initial attemptof improving the previously proposed adaptation algorithm by applyinglattice scores and by using condence measures in the traditional max-imum a posteriori adaptation (MAP) adaptation algorithm. ModiedMAP adaptation is performed for the auxiliary GMM model used in aspeaker adaptation procedure for a DNN. In addition we introduce twoapproaches - data augmentation and data selection, for improving theregularization in MAP adaptation for DNN. Experimental results on theWall Street Journal (WSJ0) corpus show that the proposed adaptationtechnique can provide, on average, up to 9:9% relative word error rate(WER) reduction under an unsupervised adaptation setup, compared tospeaker independent DNN-HMM systems built on conventional features