Non-parallel Voice Conversion with Generative Attentional Networks

Non-parallel voice conversion is a challenging task because the converted speech contents are not included in the target speaker dataset. Given the good performance of generator adversarial network in non-parallel voice conversion, CycleGAN-VC is used as the baseline system here. This work proposes the CycleGAN -CAM architecture with additional class activation map (CAM) to convert spectra. The attention is achieved by giving more weight to features with greater differences between speakers so that it can focus on the regions that can distinguish the source and the target features. In the loss function calculation, we incorporated the CAM loss function so that the architecture can adjust the weights automatically. The experimental results show that we have achieved better results than the baseline system in Mel-cepstrum distortion (MCD) and Mean Opinion Score (MOS).