Voice Conversion from Tibetan Amdo Dialect to Tibetan U-tsang Dialect Based on StarGAN-VC2

This paper proposes a conversion method from Tibetan Amdo dialect to Tibetan U-tsang dialect based on StarGAN-VC2. Recently, StarGAN-VC has gained increasing public attention due to use of a single build network. However, this method exits a difference between the converted result and the real speech. In order to alleviate this problem, the model proposes in this paper has been improved in the training part and the network architecture respectively. The training part proposes a source and target conditional adversarial loss, which allows all source domain data to be converted into target domain data. In the aspect of network architecture, a modulation-based dynamic approach is proposed, which can transform acoustic feature modulation in a specific way. Subjective and objective evaluations indicate that StarGAN-VC2 is superior to StarGAN-VC in naturalness and similarity.