My plan was to do some procgen on prompts and then potentially feed that into a language model (or a Madlibs-style template) and a sufficiently-tuned TTS system, but it all seemed like a lot of work for very little benefit.
Viewing post in Transformative Meditations jam comments
It can probably be done Madlibs-style if you have the music, voices and background noises playing separately. Randomize all 3 lists and play them on top of each other. Reorder one of the lists again whenever its last track is done playing. I think the problem with this setup is that whatever trance effect it may have will be weakened due to the random nature, and then weakened again when the person's brain realizes that the commands have started repeating.