[R] Nix-TTS 🐤: An incredibly lightweight text-to-speech via non end-to-end distillation

Hi, Reddit!

Excited to share with you guys, Nix-TTS 🐤! Our latest research in lightweight neural Text-to-Speech.

We've seen how synthetic voices generated by recent neural TTS are getting more and more natural, but most of the time the models suffer from slow CPU inference and are not end-to-end, which requires an additional vocoder model.

**Nix-TTS 🐤** is an incredibly lightweight end-to-end TTS model achieved by **applying non end-to-end knowledge distillation** to a powerful yet large-sized generative TTS teacher model. …

distillation machinelearning speech text text-to-speech