Modelpublic

Respair/dune_codec

sourceHugging Facecc-by-4.0updated 6d agoView on Hugging Face

1likes71downloads

Model Card

Details

Dune (doo-neh دونه as I like to call it, or just the english dune) is a fast and compact 12.5hz speech tokenizer trained on tens of thousands of hours of multilingual data. <br>

the encoder is based on nvidia's nano codec architecture that compresses your audio to 22khz FSQ tokens; the decoder, using a different design then reconstructs your input to high quality 44.1khz.

Batched Extraction / Inference

fill in the path to your data in dune_extraction.py, then run it. <br><br>

bash

~$ python dune_extraction.py

Important note

this is strictly a speech tokenizer, trained only on human speech; it won't do well with music.