Spherical Lattice Vector Quantization in Neural Audio Coding

In neural audio coding, latent space quantization is often trained together with the rest of the model. In this work, we investigate the use of algebraic vector quantization (VQ) in a shape-gain approach, to modify the residual vector quantization (RVQ) method in the DAC neural audio codec. Results on speech signals show that the overall performance of DAC with the proposed modifications is nearly equivalent to the original DAC model, while offering significant advantages: virtually no codebook storage, fast nearest neighbor search and indexing, codebook training limited to one gain optimization per quantization stage.