SNAC

A design insight rather than raw scale distinguishes SNAC from earlier neural audio codecs: its quantizers run at different temporal resolutions, so coarse tokens covering broad spans of time are sampled far less often than fine detail tokens. This multi-scale hierarchy compresses audio to very low bitrates while keeping token sequences short enough for language models to handle, which is why SNAC became the codec behind Orpheus TTS and several other recent speech LLMs. Three pretrained mono models ship on Hugging Face: a 19.8M parameter speech codec at 24 kHz and 0.98 kbps, and 54.5M parameter music and sound effects codecs at 32 kHz and 1.9 kbps and at 44 kHz and 2.6 kbps. Installation is pip install snac, and encoding or decoding is a few lines of PyTorch that run acceptably on CPU for the small speech model, with module definitions adapted from the Descript Audio Codec. Code and weights are MIT licensed with no usage restrictions. For anyone building an audio language model who needs compact discrete representations, it is one of the simplest codecs to adopt and among the most token-efficient available.