SiFi-GAN: Combining Source-Filter Modeling and Upsampling-Based High-Fidelity Neural Vocoder for Fast and Pitch-Controllable Speech Synthesis

Achieving speech synthesis that is simultaneously high in fidelity, fast in generation, and flexible in control remains a core challenge in neural vocoder research. Our previous work, unified Source-Filter GAN (uSFGAN), demonstrated that incorporating source-filter modeling into Generative Adversarial Networks (GAN)-based vocoders can significantly improve controllability over fundamental frequency (F0). However, its architecture operates on high-temporal-resolution features, resulting in substantial computational inefficiency compared to efficient upsampling-based models such as HiFi-GAN. To overcome these limitations, we propose Source-Filter HiFi-GAN (SiFi-GAN), which combines the efficiency of HiFi-GAN with the F0 controllability introduced by source-filter modeling. SiFi-GAN adopts a hierarchical architecture in which the filter-network is conditioned on source excitation representations generated by a separate source-network, emulating the pseudo-cascade structure of human speech production. Experimental results show that SiFi-GAN outperforms both HiFi-GAN and uSFGAN in singing voice quality, while also achieving faster synthesis speed. Consequently, SiFi-GAN is more suitable than uSFGAN for integration into real-world applications and end-to-end speech synthesis systems.