Kagome Kagome was sung at a Tokyo expo in 2011 by a silicone mouth with 9 motors and no face.
The room went quiet, because the sound was wrong in a way nobody could name, like a song being learned from inside a throat that had never owned one.
A microphone sat bolted to the right side of the frame, and everyone assumed it was there to take orders.
It was pointed at the mouth.
The machine was listening to itself.
Professor Hideyuki Sawada had given it nothing no recordings, no phoneme library, not one line of text, just an air pump for lungs, 8 vocal cords cut from silicone as soft as human mucous membrane, a tongue, and a resonance tube 180 mm long with a nasal cavity of 60 cm³ that opens for "m" and "n".
So it babbled.
200 random sounds fed into a 25 × 25 map of 625 nodes, 6 people in the room saying yes or no, and that was the entire training set smaller than what a single Instagram ad gets tested on.
It learned 5 Japanese vowels the way an 8-month-old does, make a noise, hear the noise, move the tongue 1 mm.
The clip went around as horror, one headline asked if the robot apocalypse had started, and 15 years later it still gets reposted every few months with the same caption about nightmares.
All of them missed who it was built for.
6 deaf teenagers came to that lab having never heard their own voices, only felt them in the bone, and they watched the silicone tongue take the shape of a vowel, copied it with their own, and 5 of 6 started speaking clearer.
Every AI voice on your phone was trained on millions of hours of humans talking, and not one of them has lungs.
This mouth has never read a word.
It breathes, listens, and tries again.