A Stable and Robust Emotional Voice Conversion Framework Using CycleGAN

Emotional Voice Conversion (EVC) aims to modify speech to convey various emotions while maintaining the speaker's identity and the content of their language. In this study, we propose a CycleGAN-based framework for EVC, specifically targeting the bidirectional transformation between neutral and sad emotional states. The approach integrates adversarial and cycle-consistency losses, enhanced by spectral normalization, to achieve stable and high-quality emotional conversion in the mel-spectrogram domain. The model is implemented using PyTorch and evaluated its performance on the CREMA-D and IEMOCAP datasets. The results demonstrate the effectiveness of our method, showing improved emotional expressiveness and perceptual quality measured by Mel-Cepstral Distortion (MCD) and cosine similarity. Compared to baseline models, our framework achieves superior performance, maintaining speaker identity while offering expressive and natural emotional transformations. These findings highlight the potential of our framework for applications in affective computing and human-computer interaction. Future work will focus on broadening the spectrum of emotions and enhance real-time capabilities.