hifi-gan

Definition

A GAN-based vocoder that turns mel spectrograms into raw waveforms. Its discriminators examine the signal at multiple periods and multiple scales, which forces the generator to get the periodic structure of voiced speech right — the part that naive models smear. Because generation is feed-forward rather than autoregressive, it is fast, and it became the default back end paired with acoustic models like FastSpeech in TTS pipelines.