The author used htdemucs to separate the vocals from the mix, because using the full mix directly would cause the accompaniment to overwhelm the vocal conditioning signal and render lip sync ineffective.