What happened
The plan was a second generation of Dark Choir singers with sharper identities than the register-displacement cast. Eight models were trained. In a blind listening test they collapsed toward the same voice with effects on it rather than eight distinct people.
Why
The training data was speech, not singing. For a singing project that is the wrong corpus, and the mistake was not flagged before eight models were trained on it. A second cause was in the identity method itself: shifting and texturing a reference voice with signal processing homogenises identities instead of separating them.
What was decided
- The eight models are kept as evidence and are not part of the cast.
- The first-generation undead-register choir stays valid.
- The next attempt trains on the singing corpus only and moves identity into the generative model rather than into post-processing.
- One singer gets validated by ear before any scaling to eight.
Why publish a failure
Because the process is the product here. The cast that eventually works will be the one that passed a test the previous one failed, and that only means something if the failure is on the record.