I agree that muting or muffling allows the presumably original recording (of "Laurel") to come through better. I was hearing a nasally high-pitched "yanny" with a very faint low frequency monotone "laurel" until I heard it on a smaller speaker from across the room, when it was suddenly a normal speaking voice saying "Laurel".

Surprised there wasn't more discussion here about it that I could find. Seems like this was possibly found in a study of adversarial examples for audio codecs/speech processing systems?