Environmental warping for in-car speech recognition

Summary form only given. We present an environmental warping approach to reduce the mismatch between the acoustic conditions during training and recognition. The idea of this approach is to map the log mel-filter-bank (MFB) vector obtained from the speech in a test driving condition into the one in the target driving condition, in which the acoustical models are trained. The mapping function is obtained by training multilayer perceptron (MLP) based neural network. In our in-car isolated word recognition experiments under 12 real car environments, the proposed approach obtained an average relative word error rate (WER) reduction of 47.6% and 17.5%, compared to the original speech and conventional speech enhancement methods, respectively.