Speaker normalization using dynamic frequency warping

In an effort to reduce the degradation in a gender-independence isolated word recognition performance caused by variation character among different speaker, a dynamic frequency warping approach to speaker normalization is investigated. There are a lot of discrepancy in frequency domain which caused by vocal tract length difference among different speakers. Dynamic frequency warping (DFW) is an exact analog of dynamic time warping (DTW) which is used to reduce the discrepancy frequency scale of speech and normalize the frequency accurately. In this paper, the DFW method is to be introduced to normalize the frequency scale of speech and then applied it to a gender-independence isolated word recognition system. The results of experiments show a large improvement in average word error rate.