Compasses: Expressive interpolations of spelling and sound

Compasses is a collection of minimalist neologistic poems produced with a machine learning model of spelling and phonetics, published recently in Andreas Bülhoff\u27s sync series. This talk situates the work in the context of my own practice as a computational poet, and in the broader context of avant-garde poetry as a whole. I describe the methodology and models used to produce the work, along with the open source framework (called Pincelate) I created to facilitate this work and other similar work. The model used to produce Compasses has two parts: a speller, which spells words based on how they sound, and a sounder-out, which sounds out words based on how they\u27re spelled. In the process of sounding out a word, the sounder-out produces a fixed-length numerical vector known as a hidden state : a condensed representation of a word\u27s phonetics. The speller can then use the phonetic information contained in this hidden state to produce a plausible spelling of the word. In Compasses, I use this model to generate new imaginary words that exist in the negative phonetic spaces between the phonetic hidden states corresponding to names of members of well-known quartets. The expressive potential of English spelling—and its tenuous but rich relationship with phonetics—lies at the intersection of the interests of minimalist poets, nonsense poets, and sound poets alike. Poems like Aram Saroyan\u27s lighght, Lewis Carroll\u27s Jabberwocky and Elsa Von Freytag-Loringhoven\u27s Eigasing Rin Jalamund (for example) operate only by virtue of the reader\u27s knowledge of the conventions of English spelling. Using Compasses as a model, I suggest that computational models of these conventions—along with tools to manipulate those models—open up important new avenues for creation in these and other poetic genres