Generating a complete song from lyrics and a style description is a demanding task: it requires long-range musical planning without sacrificing the fine detail of vocals and instruments. A new preprint on arXiv tackles this problem by exploring how to balance structural coherence with high-fidelity audio output.
The abstract notes that low-rate discrete semantic tokens make the planning problem tractable, but quantization limits the quality of the generated audio. The title of the paper points to a solution: joint discrete and continuous modeling, which likely combines the strengths of both representations to overcome the quantization bottleneck.
As the source is a single arXiv abstract, the details of the method are not fully available, but the stated direction is clear. The work addresses a core trade-off in song generation between long-range structure and fine-grained acoustic fidelity.