MUDslop

Fight, loot, gamble, play chess, get into relationships, sing, and verb your way to the deepest floor of the dungeon.

Played entirely with voice control (instead of typing), the piece uses Nvidia's Parakeet speech-to-text running locally — a pre-trained model, already fixed in its biases and failures, with no cloud connection and no further learning.

A trivial action like 'move north' becomes impossible when your voices compete with friends & passerby, or you're forced to clearly enunciate and rid yourself of your accent to battle with the parser disrupting the illocutionary act. Language models tend towards scale & generality, but MUDs are finite & rule-bound, stretching this dungeon as a space of confinement for you and of language.

Voice in western games is often barking militaristic commands — but instead of dehumanizing in-game characters, your failure makes you vulnerable as the models mishear you, the world kind of resisting the supposed agency it has handed to you. (Etymology-wise: the word dungeon originally referred to the most secure place of a castle, and is now a place of underground & torture.)

All in all, something about the evil trap of language models and the endless friction of wanting to be understood.