We present a feature space transformation method for personalization. This method includes a generalization of i-vector based clustering that allows parameter tying of sub-loading matrices. This method trains i-vector parameters from utterances of a device, uncovering a low dimension space for clustering variability within a device. We show through empirical results impacts of parameters of the generalized i-vector method. We conducted recognition experiments on an internal large vocabulary voice search system for gaming. The method achieved significant reductions of word error rates by 28%, compared to a per utterance adaptation system.

Index Terms: speech recognition, personalization, adaptation, i-vector