A vector-quantizer based method of speaker normalization

As an effort to reduce the performance decline of speaker independent speech recognizers due to inter-speaker variations of vocal tract length among population, a method of speaker normalization based on vector quantization is proposed. In this paper, presented is an iterative method of constructing the ´normalized´ codebook that can be used as a text independent warp factor estimator for LVCSR system. Given the normalized codebook, the warp factor is estimated by searching the best fitting warped version of feature vectors of a given utterance. Throughout the whole process of normalized codebook construction and warp factor estimation, neither acoustic, nor phonetic knowledge is made use of The effectiveness of the proposed method is investigated by performing recognition experiments. The results showed more than 4% improvements in word level accuracy.