For a long time, neural vocoders forced you to pick a side: autoregressive models like WaveNet gave you gorgeous audio but were painfully slow, while GAN-based waveform generators were fast but noticeably lower quality. HiFi-GAN closed that gap. It generates 22.05 kHz audio at 167.9x real-time on a single V100, hits mean opinion scores in the range of human recordings, and has become a default vocoder for a huge chunk of the modern TTS stack.

The review focuses on the design insight that made it work: speech is a sum of sinusoids with many different periods, so the discriminator needs to look at those periods explicitly. HiFi-GAN's multi-period discriminator (paired with a multi-scale discriminator) is what drives fidelity up while the generator stays lean. The walkthrough also covers how well it generalizes to unseen speakers for mel-spectrogram inversion, its end-to-end TTS behavior, and a small-footprint variant that runs 13.4x faster than real-time on CPU. If you're picking a vocoder for a TTS product in 2024 and want to understand why HiFi-GAN is still the sensible default, this vocoder retrospective is worth your attention.