hifi-gan

Why we included this project

If you're building a text-to-speech system, the final step is usually converting a mel-spectrogram into actual audio. HiFi-GAN does that with a GAN that sounds natural and runs faster than older autoregressive vocoders. The repo ships pretrained models, so you can try it right away, and the small-footprint version is quick enough on CPU for many use cases. It's a solid open-source option to test before settling on a commercial vocoder.