Towards Physically Interpretable Parametric Voice Conversion Functions

摘要

Typical voice conversion functions based on Gaussian mixture models are opaque in the sense that it is not straightforward to establish a link between the conversion parameters and their physical implications. Following the line of recent works, in this paper we study how physically meaningful constraints can be imposed to a system operating in the cepstral domain in order to get more informative conversion functions. The resulting method can be used to study the differences between source and target voices in terms of formant location in frequency, spectral tilt and amplitude in specific bands.