Reproduce “Variable Bitrate Residual Vector Quantization for Audio Coding” 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
Yunkee Chae, Woosung Choi, Yuhta Takida, Junghyun Koo, Yukara Ikemiya, Zhi Zhong, Kin Wai Cheuk, Marco A. Martínez-Ramírez and 3 more
VRVQ lets audio codecs vary codebook use across frames, using an importance map and binary masks to allocate quantization depth. A smooth surrogate gradient trains the controller with binary forward masks. On a DAC-based model, this surrogate improves rate-distortion results over the saturated-identity baseline; gains over DAC depend on metric, α, and bitrate, and do not establish lower encoded-file size.
Key insights
- Dynamically allocating codebooks using learned importance maps significantly improves the rate-distortion tradeoff compared to fixed-bitrate RVQ codecs, especially at bitrates above 3-4 kbps.
- The proposed smooth surrogate function for gradient estimation effectively addresses the non-differentiability of bit allocation, leading to more robust training and better performance than previous identity-based surrogates.
- Preprocessing steps like LUFS normalization can hinder the importance subnetwork's ability to learn effective content-based bit allocation, suggesting raw energy distribution is more informative.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.