Development of NU non-parallel Voice Conversion System 2018

摘要

This paper introduces the NU non-parallel voice conversion (VC) system developed at Nagoya University for SPOKE task of Voice Conversion Challenge 2018 (VCC2018). The goal of the SPOKE task is to develop VC systems without the requirement of parallel training data. In the previous version of the system, we developed a deep neural network (DNN) based cascade VC system to convert the source voice into the target voice via the TTS voice as the reference. However, the two stages conversion caused performance degradation. As a result, in this paper, we propose a compensation AutoEncoder technique to compensate the mismatch between the output of first stage conversion and the input of the second stage conversion. In addition, we also investigate the use of deep mixture density network (DMDN) to avoid the DNN-based limitations of the lack of ability to predict variance and the unimodal nature. The objective evaluation results shows the effectiveness of DMDN and the potential improvement of compensation AutoEncoder.