Fast model alignment for structured statistical approach of non-parallel corpora voice conversion

This study proposes a fast model matching algorithm of structured approach in Non-parallel corpora voice conversion. Most of conventional non-parallel corpus-based voice conversion method requires joint training which is computationally intensive and extremely inconvenient in system expansion. Existing structured approach of Non-parallel corpora voice conversion without joint training suffers from the imprecision in model alignment because of the simplified model matching algorithm, so we proposed a fast matching algorithm between statistical acoustic models of source-target speaker in structured approach of Non-parallel corpora voice conversion in this paper. In the proposed method, a Structured Gaussian mixture model (SGMM) is used to describe distribution of Linear Predication Cepstrum Coefficients (LPCC) and distribution structure of voices, then the structured distributions of source and target speaker are matched through Hill Climbing algorithm so that the conversion function is derived.