Exemplar-based voice conversion using non-negative spectrogram deconvolution

In the traditional voice conversion, converted speech is gener-ated using statistical parametric models (for example Gaussian mix-ture model) whose parameters are estimated from parallel training utterances. A well-known problem of the statistical parametric meth-ods is that statistical average in parameter estimation results in the over-smoothing of the speech parameter trajectories, and thus leads to low conversion quality. Inspired by recent success of so-called exemplar-based methods in robust speech recognition, we propose a voice conversion system based on non-negative spectrogram de-convolution with similar ideas. Exemplars, which are able to cap-ture temporal context, are employed to generate converted speech spectrogram convolutely. The exemplar-based approach is seen as a data-driven, non-parametric approach as an alternative to the tradi-tional parametric approaches to voice conversion. Experiments on VOICES database indicate that the proposed method outperforms the conventional joint density Gaussian mixture model by a wide margin in terms of both objective and subjective evaluations. Index Terms — Voice conversion, exemplar, non-negative ma-trix factorization, non-negative matrix deconvolution, temporal in-formation 1