Q2D2: A New Quantization Method for Neural Audio Codecs

Researchers have introduced Two-Dimensional Quantization (Q2D2), a novel quantization scheme for neural audio codecs that projects feature pairs onto structured 2D grids such as hexagonal, rhombic, or rectangular tiling. Unlike conventional methods like Residual Vector Quantization (RVQ), Vector Quantization (VQ), and Finite Scalar Quantization (FSQ), Q2D2 preserves the geometric structure of the latent space and improves correlation capture between features. This leads to enhanced representation learning, better codebook utilization, and lower token rates. The method achieves codebook sizes comparable to traditional approaches while improving audio compression efficiency. The paper is available on arXiv under identifier 2512.01537.