WaveNeXt 2 unifies GAN and diffusion vocoders, trains in 32 hours

WaveNeXt 2, presented at ICASSP 2026 by researchers from NICT and JAIST, is a unified neural vocoder framework built on ConvNeXt blocks. It tackles a key limitation of prior work: most vocoders are locked into either GAN or diffusion paradigms. State-of-the-art models like Vocos and the original WaveNeXt only operate as GANs and struggle with multi-speaker scenarios. Diffusion vocoders, while faster to train, suffer from slow CPU inference. WaveNeXt 2 introduces residual denoising and sub-modeling, where multiple sub-models sequentially refine the waveform, enabling a single architecture to work seamlessly with both GAN and diffusion training.

Experimental results on multi-speaker datasets confirm the performance gains. The GAN variant (GAN-WaveNeXt 2) is substantially faster than HiFi-GAN and WaveFit. The diffusion variant (Diff-WaveNeXt 2) achieves inference speed superior to FastDiff with just 4 sampling steps, while maintaining competitive synthesis quality. Crucially, Diff-WaveNeXt 2 trains in only 32 hours, making it highly practical for teams with limited compute. This efficiency, combined with its dual-mode flexibility, positions WaveNeXt 2 as a strong contender for real-time speech synthesis in resource-constrained environments.