Fuzzy ASC-GAN: A Fuzzy C-means Audio Similarity CycleGAN on Articulation Disorder Voice Conversion
Yiwei Huang, Sheng‐Wen Chang, Chiun-Li Chin, Guang–Tao Lin, Pei‐Hsin Chang, Wan-Xuan Lin, Jun-Ren Chen, Zhi-Huan Zheng, Geng-Kai Wong · 2023
This paper introduces an innovative approach, Fuzzy Audio Similarity CycleGAN (Fuzzy ASC-GAN), turned at addressing articulation disorders and enhancing dysarthric voice conversion (DVC). By integrating Fuzzy C-means clustering and CycleGAN-VC2, the proposed method focuses on refining both training and transformation processes. Notably, Fuzzy ASC-GAN incorporates a unique strategy that leverages the audio similarity of timbres to adjust the training data of CycleGAN-VC2, ensuring accurate and effective voice conversion. Through repeated and rigorous parameter adjustments and comprehensive performance comparisons, Fuzzy ASC-GAN emerges as the standout performer, achieving an average accuracy of 93.35% across diverse S2T models. The method's distinct advantage lies in its ability to tailor voice conversion about specific characteristics of individual with articulation disorders. This highlights the potential of Fuzzy ASC-GAN to significantly enhance communication for individuals with dysarthric voices, marking a significant advancement in the realm of voice conversion techniques.