IterVocoder: Fast High-Fidelity Speech Synthesis via GAN-Guided Iterative Refinement

Recent progress in neural vocoders has demonstrated impressive advances in natural speech synthesis. Among them, denoising diffusion probabilistic models (DDPMs) and generative adversarial networks (GANs) stand out due to their ability to produce high-fidelity audio. However, DDPMs typically require a large number of iterative steps, and GANs often suffer from training instability. To reconcile these limitations, we propose IterVocoder, a novel non-autoregressive neural vocoder that unifies fixed-point iteration and adversarial learning. By applying a deep denoising network iteratively and enforcing consistency through adversarial objectives at each refinement stage, IterVocoder achieves high-quality waveform synthesis in just a few iterations. Experimental results show that IterVocoder can synthesize speech with perceptual quality on par with human speech while being over 200× faster than autoregressive models. This makes IterVocoder a practical solution for real-time neural vocoding applications.