This repository provides training and inference code for Descript Audio Codec, a high-fidelity general neural audio codec. It uses residual vector quantization with a GAN-based architecture to achieve state-of-the-art compression ratios. The model operates at 44.1 kHz and supports mono/stereo audio, and serves as a universal model working across speech, music, and environment sound domains. It is designed as a drop-in replacement for EnCodec in audio language modeling applications.