Perhaps part of the problem is that these files tended to be distributed in a way that led to fragmentation: someone might add lyrics for their own local music

Perhaps part of the problem is that these files tended to be distributed in a way that led to fragmentation: someone might add lyrics for their own local music files, but they might not share them back, or they might add them in a way that is unusual or unexpected to other users. Consensus becomes difficult to achieve. Strangely enough, source control foundries seem to have solved a similar problem, to a large extent. They bring together a large number of very different (authorship, content, language, ...) projects within one communal space, and while they provide plenty of agency for project owners to adjust details as they wish, they also allow for forking/editing and tend to lead towards an "eventual consistency". Perhaps a more-open, foundry-like model could work to gather consensus for lyrical annotation (potentially including natural language translations) of music.