No, I don’t think it applies here. The semantics and speech patterns were generated by a human, Whisper just transcribed them.

There is some risk that Whisper transcribed inaccurately, but that’s less model collapse and more “the dataset is bad”.