LeVo

LeVo is a lyrics-to-song model that treats songwriting as two jobs at once. Its language model, LeLM, predicts two kinds of audio tokens in parallel: mixed tokens that capture vocals and accompaniment together for tight vocal-instrument harmony, and dual-track tokens that encode the vocal line and the backing track separately for cleaner, higher-quality synthesis. Two decoder-only transformers handle the two token streams, with a modular training strategy that keeps them from interfering with each other.

What sets LeVo apart is its post-training: a multi-preference alignment step built on direct preference optimization, trained on semi-automatically constructed preference data to improve musicality and instruction following. The paper reports that it beats existing open-source song models on both objective and subjective measures and holds its own against industry systems. The project demo page even lines its outputs up against Suno v4.5, Mureka, YuE, DiffRhythm and ACE-Step for direct comparison. One caution: the official code repo ships with a non-standard custom license, so commercial use needs a careful read first.

• Paper (arXiv 2506.07520): reports outperforming open-source lyrics-to-song methods on objective and subjective metrics, competitive with industry systems
• Demo page A/B comparisons vs Suno v4.5, Mureka O1, YuE, DiffRhythm, ACE-Step
• All headline claims author-reported; no independent verification