Non-parallel hierarchical training for voice conversion

Many research topics in speech processing face the same difficult problem, how to create cheaply (or quickly) a parallel corpuswhich associates the acoustic realizations of two speakers having pronounced the same linguistic content. Among those topics are voice conversion techniques and some aspects of speech and speaker recognition. In the context of voice conversion, we propose a new methodology to map the source speaker vectors with those of a target speaker, without any parallel corpus nor using DTW (Dynamic Time Warping). The proposed approach is based on a hierarchical decomposition of the source and target acoustic spaces. At each level, source and target class centroids of a reduced subspace are paired. We propose an evaluation of our algorithm when applied to GMM-based voice conversion on the ARCTIC database.