Among thousands of spoken languages are used today, few ofthem are studied by the speech recognition community [3]. Oneof the major hurdles of automatic speech recognition (ASR) system deployment in new languages is that ASR system relies on a large amount of training data for acoustic modeling. Usually, to build a reasonableacoustic model for a largevocabulary continuous speech recognition (LVCSR) system, tens to hundreds of hours of trainingdata are required, which makes a full-fledged acoustic modelingprocess impractical especially for under-resourced languages.This motivates us to investigate methods to automatically transfer well-trained acoustic models to under-resourced languages.