LoSATok: Low-Dimensional Audio Tokenizer for Cross-Domain Understanding and Generation

Researchers propose LoSATok, a low-dimensional audio tokenizer designed to unify audio understanding and generation across domains. Traditional unified tokenizers encode both semantic and acoustic details in high-dimensional continuous latents, increasing the modeling burden for Diffusion Transformers (DiTs). LoSATok introduces a Semantic Bottleneck that compresses 1280-dimensional semantic encoder features into 128 dimensions, regularized by a time-relation loss for temporal feature consistency. A dual-level semantic supervision method leverages both high- and low-dimensional semantic signals, enabling joint capture of semantics and acoustic details within a compact latent space. The approach is motivated by the observation that high-dimensional semantic features are compressible. The work is published on arXiv under ID 2605.27840.