Speaker Recognition Using Temporal Decomposition of LSF for Mobile Environment

摘要

A novel approach to speaker recognition in mobile or IP network environment is described. In this approach, we use decoded line spectral frequency (LSF) parameters directly from compressed speech packets instead of using parameters from decompression and analysis procedure. Furthermore, we reduce the number of LSF series based on a restricted temporal decomposition method. Consequently, proposed approach gets more than three times faster than a traditional speaker recognition approach without losing any accuracy according to our experiments.