Formant model estimation and transformation for voice morphing

In this paper we consider the estimation and mapping of timevarying formant model parameters and orders for voice transformation. The model order is the number of perceptually significant formant trajectories estimated from an analysis of the poles of “over-modelled’’ linear prediction models of the source and target speech. A 2-D HMM with NF left–to-right states across frequency and M states across time is used to classify formant observations into NF sequential formant clusters. A formant-based non-uniform frequency warping method is proposed for voice transformation. In this method speech spectrum is divided into NF+1 formant bands. A transformation is estimated for each formant band of a phoneme model. Multi-mixture Gaussians are used to model the distribution of parameters in each formant band. The voice mapping yields perceptually high quality results.