Speaker normalization with the band-pass transform

This paper presents a new spectral transformation for speaker normalization called the "band-pass transform" (BPT). From the framework of the bilinear transformation (BT), the BPT is a new frequency warping resulting from a mapping of a prototype band-pass (BP) filter into a general BP filter. The BPT offers two degrees of freedom enabling complex warpings of the frequency axis. Estimation of the BPT parameters is performed using the Nelder-Mead algorithm. Our experimental results include a study of the BPT performance. BPT performs better than other VTLN methods and offers a gain of 1.13% absolute on Hub-5 English Eva101 set.