This is an official implementation of the LyriCAR:A Difficulty-Aware Curriculum Reinforcement Learning Framework For Controllable Lyric Translation.

dataset model paper

The dataset is derived from DALI, a large collection of synchronized audio, lyrics, and vocal notes. From 6,984 English songs after filtering, we extracted lyrics and constructed 9,600 paragraphs per stage.

Table: Difficulty distribution and dataset size of each stage.

LyriCAR achieves state-of-the-art performance across all automatic metrics (BLEU, COMET) on the EN–CH lyric translation task, outperforming all baselines—including the strong Qwen3-8B model. It also obtains the highest scores on our multi-dimensional reward evaluation, indicating that the model has truly internalized musical-linguistic alignment patterns rather than relying on shallow correlations. Notably, LyriCAR-SD further reduces training steps by 34% compared to LyriCAR-SS while delivering superior translation quality.

Table: Comparisons with previous SOTA, where LyriCAR-F means the full-data version, LyriCAR-SS and LyriCAR-SD mean the staged static and staged dynamic versions respectively.

💡 Reward trajectories of LyriCAR-SS, LyriCAR-F and LyriCAR-SD.

💡Reward variance(LyriCAR-SD)