A Revisit to Feature Handling for High-quality Voice Conversion Based on Gaussian Mixture Model
Hitoshi Suda, Gaku Kotani, Shinnosuke Takamichi, Daisuke Saito · 2018
This paper discusses influences of handling acoustic features on the quality of generated sounds in voice conversion (VC) systems based on Gaussian mixture models (GMMs). In the context of improving the quality of VC, mapping models, which are used to convert acoustic features, have been widely discussed. Nevertheless, the components other than the mapping models have rarely been studied. The experimental results show that the quality of VC depends on not only the models but also the methods of analysis and synthesis of utterances. This paper also investigates filtering methods for synthesis. In order to avoid buzzy sounds generated from vocoders, differential-spectrum compensation is applied as an alternative method of synthesizing waveforms. Although mel log spectral approximation (MLSA) filtering is traditionally used for differential-spectrum compensation, the experimental results indicate the approximation in MLSA filtering degrades the quality of the synthesized speech. In order to avoid this approximation, this paper introduces an alternative filtering method, which is named SP-WORLD, inspired by the WORLD vocoder framework. The subjective experiments demonstrate that SP-WORLD is comparable to MLSA filtering, and outperforms it in some cases.