Reducing Mismatch of WaveNet Vocoder for Variational Autoencoder based Voice Conversion

摘要

In this work, we proposed to utilize the selfreconstructionprocedure in the VAE-based VCframework to train or adapt the WaveNet vocoder foralleviating the performance degradation caused bythe mismatch between the training and conversionphases of the WaveNet vocoder. Evaluation resultsconfirm the effectiveness of the proposed method.