Lyrics fidelity: the take that sings its lyrics
A music model does not always sing what it was given. It leaves a line out, repeats one, or sings other words - and the same lyrics, score and settings sing better in one take than in the next. In 37 sung YuE2 takes of October 2026 the median take sang 89 % of its lyrics' words as written; one in six sang less than 75 %, and one verse was sung with 8 of its 65 words in one take and 64 of 65 in the next.
Song Brief and Cover Brief have a field for it, lyrics fidelity (the last one):
| Choice | What happens | Time |
|---|---|---|
| standard | one take, as before | the render |
| high - best of 2 takes | a second take with the next seed when the first sang less than 96 % of the words; the take that sings more of the lyrics as written is kept | up to twice the render |
| highest - best of 3 takes | as high, and a third take when neither of the two sang 96 % of the words | up to three times |
Higher fidelity makes the generation take longer: every further take is a whole render, and every take is heard once more.
How it works
The node Best Take · lyrics fidelity sits in the render section of every template with lyrics (YuE2 Song, YuE2 Cover, MiniMax Song, YuE2 DAW), after the render block - in the cover template after Check Vocals. The render block holds two more takes - the take seed + 1 and + 2 - that are rendered only when Best Take asks for them: never with standard; with high and highest a further take only while the best take so far sang less than 96 % of its lyrics as written - a take that sings 96 % is close enough, and the run goes on with it at once.
Each take is heard by the lyrics ASR - the same hearing Export Release uses for the subtitles, kept in its cache - and scored like the release record's as sung: the share of the lyrics' words heard as written. The take with the highest share goes on to the Master block; at a tie the earlier one. Best Take's summary and the release record list every take's share and the one kept, for example take 2 of 2 kept - 94 % of the lyrics heard as written (take 1: 81 %, take 2: 94 %).
An instrumental, a song without sung lines, or a workflow without a brief always takes the first take. A workflow saved before the option has no value for it and keeps standard.
What it does not change
- The documents: the lyrics, the style and the score stay exactly as approved - the takes differ only by their seed.
- The hearing is a measurement: an unclear voice can score lower than it sings, and a word the ASR mishears counts as not sung. What the kept take sings is one click away in the Plenio Player (As sung).
- The take seed shown is the first take's; the kept take's seed is in the report (+1 or +2).
Related: Seeds, Tags, variables and lyrics subtitles, Plenio Player (the lyrics as written and as sung).