DAC (Descript Audio Codec)

The Descript Audio Codec, usually shortened to DAC, is a neural audio compression model that squeezes 44.1 kHz audio into discrete tokens at 8 kbps, a compression ratio of roughly 90x, while preserving fidelity across speech, music, and environmental sound. Built on an improved RVQGAN architecture, it outperforms earlier codecs such as EnCodec and SoundStream at comparable bitrates, and pretrained models are provided for 16 kHz, 24 kHz, and 44.1 kHz sample rates. Because it produces compact discrete codes, DAC is widely used as the tokenizer stage in audio language models and generative audio systems, serving as a drop-in replacement for EnCodec. The project ships a command-line interface for encoding and decoding, a Python API for programmatic use, and Docker images with GPU support. Descript releases the code and model weights under the MIT license, and its typical audience is researchers and engineers working on speech synthesis, music generation, and audio compression.