WavTokenizer

Compressing a second of 24 kHz audio into as few as 40 discrete tokens, WavTokenizer pushes neural codec efficiency to roughly 0.48 kbps, or 0.9 kbps at 75 tokens per second, using a single quantizer where EnCodec and DAC stack several. Fewer, richer tokens matter because downstream audio language models spend context length on every token, and the ICLR 2025 paper argues this extreme compression still preserves enough semantic information for modeling while reporting strong reconstruction quality on UTMOS. Released checkpoints span small speech models trained on LibriTTS, a medium model trained on 10,000 hours, and large variants trained on 80,000 hours covering speech, music, and general audio, all downloadable from Hugging Face with encoder, decoder, and training scripts in the repository. Setup is a conda environment with pip-installed requirements, and the small models run without data center hardware. Everything is MIT licensed, permitting commercial use. Researchers building speech LLMs use it as a reference single-codebook codec and as a baseline for token-efficient audio representation work.