Acknowledgements

Alignment

The default aligner is the multi-task model by Jiawen Huang, Emmanouil Benetos and Sebastian Ewert, released under the MIT licence at github.com/jhuang448/LyricsAlignment-MTL.

Huang, J., Benetos, E. and Ewert, S. (2022). Improving Lyrics Alignment Through Joint Pitch Detection. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2022), pp. 451-455.

A second aligner, which also checks every track for an untranscribed intro, is the phoneme-level model by Kilian Schulze-Forster and colleagues, released under the MIT licence at github.com/schufo/lyrics-aligner.

Schulze-Forster, K., Doire, C. S. J., Richard, G. and Badeau, R. (2021). Phoneme Level Lyrics Alignment and Text-Informed Singing Voice Separation. IEEE/ACM Transactions on Audio, Speech and Language Processing, 29, pp. 2382-2395.

Vocal separation

The vocal is isolated from the mix with Hybrid Transformer Demucs by Meta AI Research, released under the MIT licence at github.com/facebookresearch/demucs.

Rouard, S., Massa, F. and Defossez, A. (2023). Hybrid Transformers for Music Source Separation. ICASSP 2023.

Everything else

PyTorch, librosa, NumPy, SciPy, NLTK, the CMU Pronouncing Dictionary, FFmpeg, Next.js, React, Supabase and the Unbounded, Hanken Grotesk and DM Mono typefaces, each under its own permissive or open licence. The fonts are served from our own servers.

The three models were trained by their authors on academic datasets. LyriSync does not hold, and has never held, any of that data. How we handle your own audio, and for how long, is in the Privacy Policy. The words you paste are used only to time your track.