A flexible and modular crosslingual voice conversion system

A cross-lingual voice conversion system aims at modifying the timbral structure of recorded sentences from a source speaker, in order to obtain processed sentences which are perceived as the same sentences uttered by a target speaker. This work presents the cross-lingual voice conversion problem as a network of related sub-problems and discuss several techniques for solving each of these sub-problems, in the context of a modular implementation that facilitates comparisons between competing techniques. The implemented system aims at high-quality cross-lingual voice conversion in a text-independent setting, i.e. where the training sets of sentences recorded by source and target speakers are not the same. New strategies are introduced, such as artificial phonetic maps, N -likelihood clustering and normalized frequency warping, which are evaluated through numerical experiments.