Bayesian constrained frequency warping HMMS for speaker normalisation

Abstract

This paper presents a Bayesian constrained frequency warping technique. The Bayesian approach provides for inclusion of the prior information of the frequency warping parameter and for adjusting the search range in order to obtain the best warping factor dependent on HMMs. We introduce novel frequency warping (FWP) HMMs which are different warped versions of HMMs. Instead of frequency warping of the input speech we warp the spectrum of the HMMs. This is equivalent to HMMs which have both time and frequency warping capabilities. Experimentally FWP HMMs outperform the conventional constrained frequency warping approach. Furthermore, the best warping factor is estimated in two stages, a coarse stage followed by a fine stage. This method efficiently gauges the optimal warping factor and normalises the FWP HMMs.