Khala: A Two-Stage Framework for High-Fidelity Music Generation Using Acoustic Token Language Models

A research paper on arXiv (2605.01790) introduces Khala, a novel framework for music generation that models both structure and fidelity within a single deep acoustic-token hierarchy. The system uses a 64-layer residual vector quantization (RVQ) acoustic representation and a two-stage coarse-to-fine generation process. A backbone model generates coarse acoustic tokens for the full track, followed by a super-resolution model that refines finer tokens layer by layer in parallel over time, resulting in a fixed 62-step inference process. This approach aims to improve lyric alignment and fine-detail reconstruction, challenging the common design pattern of separate representation spaces for structure and fidelity.