# M3 Phoneme Relation Maps We show the attention map of a head with a fixed utterance on the left, and the PRM of that head over the entire speech corpus on the right. ## Layer 1 ![](https://i.imgur.com/0F9GHmY.jpg) ## Layer 2 ![](https://i.imgur.com/7WQp1Kw.jpg) ## Layer 3 ![](https://i.imgur.com/UTp9LEA.jpg) ## Layer 4 ![](https://i.imgur.com/9KUxkvp.jpg) ## Layer 5 ![](https://i.imgur.com/15bU3CW.jpg) ## Layer 6 ![](https://i.imgur.com/lGRcAOx.jpg) ## Layer 7 ![](https://i.imgur.com/qVqYny6.jpg) ## Layer 8 ![](https://i.imgur.com/ttlbVu9.jpg) ## Layer 9 ![](https://i.imgur.com/fKwwl8o.jpg) ## Layer 10 ![](https://i.imgur.com/cxhS2iV.jpg) ## Layer 11 ![](https://i.imgur.com/Cg3SPLo.jpg) ## Layer 12 ![](https://i.imgur.com/8wahlxt.jpg)