Voice Conversion of Tibetan Amdo Dialect Based on Cycle-Consistent Adversarial Network

This paper proposes a Voice Conversion (VC) method of Tibetan Amdo Dialect based on Cycle-Consistent Adversarial Network (CycleGAN). Non-parallel VC performance based on CycleGAN is almost similar to parallel VC. The model used in this paper is a improved version of CycleGAN, which has been improved in terms of network structure and adversarial losses. In terms of network structures, the model uses a 2D Convolutional Neural Network (CNN) as a generator and uses PathGAN instead of FullGAN as an discriminator. In order to reduce the negative impact of excessive smoothing, a new discriminator is introduced to improve the adversarial losses. The improved VC model used in this paper shows a good effect than the Traditional CycleGAN model. Objective evaluation shows that VC of Tibetan Amdo Dialect based on CycleGAN is close to the spectrogram of target speakers. Subjective assessments mainly include naturalness and similarity. In terms of naturalness, the MOS score of the converted voice quality reaches 3.33. In terms of similarities, the converted voice is closer to the voice of the target speaker.