Open-source attribution

LrcSong runs on open-source AI models. Most of them carry licences that ask nothing of us beyond keeping the licence with the code. A few are published under Creative Commons Attribution 4.0, which asks for credit. This page is that credit.

It is not an inventory of every model we run — it names the works whose licences ask to be named, and it names them here rather than on each tool page because a single model can sit behind several tools.

We run these models on our own servers to produce your output. We do not redistribute the model files or the recordings they were trained on, and we use each one unmodified.

The Voice Changer, Voice Cloner, and AI Cover Studio convert a voice using FreeVC (freevc24) by Jing-Yi Li (OlaWod), distributed under the MIT licence.

FreeVC was trained on the CSTR VCTK Corpus — around 44 hours of speech from 110 speakers, recorded at the Centre for Speech Technology Research, University of Edinburgh, by Christophe Veaux, Junichi Yamagishi and Kirsten MacDonald. The corpus is licensed CC BY 4.0 and is available from Edinburgh DataShare.

Veaux, C., Yamagishi, J. and MacDonald, K. (2017). CSTR VCTK Corpus: English Multi-speaker Corpus for CSTR Voice Cloning Toolkit.

Worth knowing: FreeVC is zero-shot. The voice you hear back is shaped by the reference audio you supply — no VCTK speaker's voice is reproduced in your output.

When a job aligns lyrics to audio word by word, the aligner is chosen by the language we detect. For Turkish, that model is mpoyraz/wav2vec2-xls-r-300m-cv7-turkish by mpoyraz, published under CC BY 4.0. It is used unmodified, and only to place timestamps — it never changes a word of your lyrics.

WhisperX ships aligners for 23 languages. We run 16 of them. The seven we hold back are held back on licence grounds, listed below — not because they do not work. Every other language, and those seven, keep Whisper's own word timings, which are looser.

WhisperX picks each aligner from one of two built-in tables: five languages come from torchaudio's bundles, and the rest from a table of 34 models on Hugging Face. We checked the weights licence of every model in both, and they are not uniform:

• English uses WAV2VEC2_ASR_BASE_960H, which is MIT — the one torchaudio bundle we can use commercially.
• French, German, Spanish and Italian use torchaudio's VoxPopuli bundles, which are CC BY-NC 4.0 — non-commercial. We run a paid product, so we do not run them. These four languages keep Whisper's own word timings instead.
  We should say plainly that we ran these four for a long time before noticing. WhisperX checks the torchaudio table first, and every licence review we did looked only at the 34-model table — so four non-commercial aligners sat in the blind spot of a page whose whole purpose is to not have one. We stopped them on 17 July 2026.
• In the 34-model table, 27 are Apache-2.0, which asks only that the licence text travel with the model. A 28th, the Galician aligner, states Apache-2.0 in the text of its model card, though its machine-readable licence field is empty.
• The Vietnamese aligner, nguyenvulebinh/wav2vec2-base-vi, is CC BY-NC 4.0 — non-commercial. We run a paid product, so we do not run it. Vietnamese jobs keep Whisper's own word timings instead, which are looser.
• The Danish aligner, saattrupdan/wav2vec2-xls-r-300m-ftspeech, is trained on Danish Parliament broadcasts and carries the Parliament's own terms of use rather than a standard open licence. Those terms are in Danish and we have not read them, so we do not know whether they permit what we do. We do not run it until we know. Danish jobs keep Whisper's own word timings.
• Three — Hebrew, Hindi and Croatian — declare no licence at all. Their cards say nothing, so we say undeclared rather than guess. No licence is not a permissive licence; it means the author granted nothing to anyone, so we do not run them either. (Hindi and Croatian were never used here regardless — they fall outside the 23 languages above.)
• The last two are worth separating out, because they are not the same problem as the Vietnamese and VoxPopuli models. Those we know we may not use. Danish and Hebrew we simply cannot tell — and on a paid product, the honest default for "we cannot tell" is not to. If you are the author of either and can tell us the terms, we would like to turn them back on.

None of them ever touch your words. An aligner only decides when a word is sung, so a language that loses one loses timing precision — never a lyric.

If you built something we run and this page has your name wrong, your licence wrong, or your work missing, that is a bug and we want to fix it — email support@lrcsong.com. An attribution page is only worth the accuracy of its worst line.