Voice Conversion Using Non-parallel Corpora Based on Enhanced Variation Auto-encoder

This paper proposed a novel enhanced variational auto-encoder(EVAE) for voice conversion using non-parallel corpora. Firstly, the source speech was encoded into a speech code with Gaussian distribution through the encoder , then the decoder reconstructed the speech code to the specified target speech. Finally, the generated target speech was optimized through the EVAE. The EVAE was one input corresponding to one output and this made the algorithm of this paper had better denoising ability. In addition, this article also introduced a cyclic training method to improve the target orientation of the converted speech. The experimental results showed that compared with the basic variational auto-encode voice conversion system without an enhanced network, the enhanced conversion system was about 10.3% lower in the objective evaluation of spectral distortions in inter-gender voice conversion. Improvements in the similarity and clearness of subjective evaluation standards had also been achieved. The result shows that the novel algorithm proposed in this paper can make the converted speech have a good target orientation. At the same time, the voice quality is better.