Interesting, this reminds me a bit of TTS models that had poor quality training data inserting commercial jingles or background music into their output, although I would agree that in this case it sounds intentional.