rvqgan

Definition

A family of neural audio codecs combining residual vector quantisation with GAN-style adversarial reconstruction losses; the best-known instance is Descript's Improved RVQGAN, shipped as the Descript Audio Codec. It sits in the same lineage as SoundStream, differing mainly in quantiser utilisation and discriminator design, and aims at high-fidelity reconstruction of speech, music and general audio from one model.