Reproduce “HiFi-Codec: Group-residual Vector quantization for High Fidelity Audio Codec” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dongchao Yang, Songxiang Liu, Rongjie Huang, Jinchuan Tian, Chao Weng, Yuexian Zou
HiFi-Codec, a neural audio codec, introduces Group-Residual Vector Quantization (GRVQ) to achieve high reconstruction fidelity. It demonstrates superior audio quality using fewer discrete codebooks than state-of-the-art codecs like Encodec, while also releasing an open-source training toolkit.
Key insights
- Splitting latent features into independent groups for RVQ (GRVQ) allows initial quantizers to capture more distinct information in parallel, reducing the total number of codebooks required.
- Concentrating more information in early quantization layers enables comparable or superior fidelity with fewer overall residual quantization steps, benefiting downstream generative models.
- Open-sourcing training code and pre-trained models for state-of-the-art codecs (AcademiCodec) democratizes access and accelerates research in neural audio compression.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.