The purpose of this paper is tuning-free SMAPLR approach. The one of the important issues of SMAPLR approach is deciding occupancy threshold for the tree structure according to the number of adaptation data and control parameter for the hierarchical prior setting for appropriately incorporating prior information. For that purpose, we employ variational Bayesian linear regression (VBLR) approach. VBLR uses variational lower bound as an objective function. Using variational lower bound, model structure and contribution of the prior information can be decided without deciding experimentally. In this paper, we employ the VBLR adaptation in normalized feature-space. We first perform the feature space SMAPLR (fSMAPLR) to normalize the feature space. Then, VBLR is performed in previously normalized feature space. Experiments on large vocabulary continuous speech recognition using the Corpus of Spontaneous Japanese (CSJ) corpus confirm the effectiveness of the proposed method compared with other conventional adaptation methods.

Index Terms: speaker adaptation, SMAPLR, VBLR, normalized feature space