Melody-Lyrics Matching with Contrastive Alignment Loss

📔 ArXiv    📔 HAL    💻 Code    💾 Dataset

This is a companion page for the following paper:

C. Wang, M. Olvera, and G. Richard, “Melody-Lyrics Matching with Contrastive Alignment Loss”, IEEE/ACM Transactions on Audio, Speech and Language Processing (TASLP), vol 34, pp 3560-3571, 2026.

Listenable synthesized singing renditions of the example lyrics in Fig. 1:

Rank 1
Rank 2
Rank 3
Rank 4

We provide more examples to supplement the example (Fig. 8) in the paper. Each example includes (1) Playable MIDI notes; (2) Reference (Ref) and top-1 non-perturbed lyrics retrieved by our proposed MLM-CAL (MC), the length-informed (LI) method, and random baseline (Ra) all aligned with the melody in terms of words; (3) the same set of lyrics, but aligned with the melody at the syllable level.

Ref
MC
LI
Ra
Ref
MC
LI
Ra
Ref
MC
LI
Ra
Ref
MC
LI
Ra
Ref
MC
LI
Ra
Ref
MC
LI
Ra
Ref
MC
LI
Ra
Ref
MC
LI
Ra
Ref
MC
LI
Ra

Citation

If you use our work in your research, please cite our paper:

@article{wang2025melody,
  title={Melody-Lyrics Matching with Contrastive Alignment Loss},
  author={Wang, Changhong and Olvera, Michel and Richard, Ga{\"e}l},
  journal={IEEE Transactions on Audio, Speech and Language Processing},
  volume={34},
  pages={3560-3571},
  year={2026},
}