LeVo: High-Quality Song Generation with Multi-Preference Alignment

Shun Lei, Yaoxun Xu, Zhiwei Lin, Huaicheng Zhang, Wei Tan, Hangting Chen, Jianwei Yu, Yixuan Zhang, Chenyu Yang, Haina Zhu, Shuai Wang, Zhiyong Wu, Dong Yu · 2025

Read LeVo: High-Quality Song Generation with Multi-Preference Alignment by Shun Lei, Yaoxun Xu, Zhiwei Lin, Huaicheng Zhang, Wei Tan, Hangting Chen, Jianwei Yu, Yixuan Zhang, Chenyu Yang, Haina Zhu, Shuai Wang, Zhiyong Wu, Dong Yu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-07520.txt.

Description

Shun Lei and co-authors present a song generation system in a 2025 paper. Turning lyrics into full songs is hard: voice and backing must blend, quality data are scarce, and taste is personal. Their framework pairs a language model for song structure with a codec that renders the audio.

The model writes two token streams at once, one mixed track for harmony and one split track that keeps voice and backing apart for clarity. Training runs in modular stages so the streams do not fight. A preference tuning step, fed by semi automatic ranking data, sharpens musicality and instruction following. Tests beat open rivals on measured and human scores and rival industry systems.

Catalog record

CREATORShun Lei, Yaoxun Xu, Zhiwei Lin, Huaicheng Zhang, Wei Tan, Hangting Chen, Jianwei Yu, Yixuan Zhang, Chenyu Yang, Haina Zhu, Shuai Wang, Zhiyong Wu, Dong Yu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-07520

Browse the full library · About the Tower of Babel