GPGAN-VC: Enhancing Voice Conversion using Gradient Penalty

This research introduces Gradient Penalty Generative Adversarial Network-Voice Conversion (GPGAN-VC), an upgraded version of the StarGAN-VC framework-a leading nonparallel VC method. Unlike traditional VC methods, which require parallel data, non-parallel multi-domain approaches, such as StarGAN-VC, have gained popularity due to their ability to function without this limitation. The main innovation in GPGAN-VC is including contrastive learning techniques in the GPGAN-VC Discriminator (D). Instead of using the typical structure, our approach introduces a gradient penalty mechanism. This modification not only increases the consistency of the training process but also effectively prevents over-fitting, resulting in a more robust and reliable VC model. We evaluated our generated sample using both subjective and objective metrics. Subjectively, our sample’s Mean Opinion Score (MOS) was 3.71 on a 5-point scale, achieving a higher performance than existing VC architectures. Objectively, we assessed the genreted samples with Perceptual Evaluation of Speech Quality (PESQ), Mel Cepstral Distortion (MCD), Modulation Spectra Distance (MSD), Signal-to-Noise Ratio (SNR),Fréchet Audio Distance (FAD), and Root Mean Square Error (RMSE) metrics, which scores 1.13, 5.70, 0.08, -0.84, 1.78 and 0.04 respectively, which demonstrating that our model outperforms best-in-class non-parallel VC models.