descript-audio-codec

Why we included this project

This neural codec compresses 44.1kHz audio into discrete tokens at an 8 kbps bitrate, roughly a 90x reduction, while keeping artifacts low enough for high-fidelity reconstruction. That makes it a practical building block for audio language models, text-to-speech, and music generation. The trained weights are universal, so one model covers everything from speech to environmental sound without per-domain retraining, and it drops in as a replacement for EnCodec-style tokenizers in AudioLM or MusicGen workflows. The repository includes encode and decode CLI tools plus a Python API, with pretrained weights at 16kHz, 24kHz, and 44.1kHz, so you can compress and reconstruct files or plug the tokenizer into your own model directly. Training scripts are included too, if you want to adapt the codec to a custom sampling rate or dataset.