Acknowledgements

The models

The assistant hears and speaks through Kyutai’s open models: their speech recogniser, their text-to-speech model and the Mimi codec. The Rust ports, the kernels, the voice and the rendering head are this project’s work on top of theirs. The language model is Qwen, run through Apple’s MLX.

The voices

The assistant’s voice was inverted from public domain and CC0 recordings and blended so that no one person’s voice survives in it. The people who made those recordings gave their voices to the commons, not to a synthetic voice on a website, so they are not named here and their recordings are not on this site. It could not have been made without them, and I am grateful.

The method

The voice was found by textual inversion (Gal et al., 2022), applied to a voice tensor instead of an image concept. The mouth on the front page reads formants back into a tongue position after Stevens and House (1955), with vowel targets from Peterson and Barney (1952). Pitch tracks are pYIN, from librosa; word times are Whisper’s.

The help

I used Claude, Anthropic’s model, as a tool throughout the research and the engineering: to search the literature, to write and check code, to run experiments overnight, and to edit the essays I had drafted. The measurements, the listening and the decisions are mine, and so is responsibility for anything wrong here.

The design

This site would not look the way it does without Kat, the Poet Engineer. Her experimental interface design on X is where I found the idea that the figures for a project still being worked out should be instruments rather than illustrations. The plates on these pages are built in that spirit.

The type

Handjet by Rosetta for the titles, Sligoil by Ariel Martín Pérez for the body, and Martian Mono by Evil Martians for the readouts. All three are under the SIL Open Font License and served from here.

The figures

Every figure is drawn from a measurement: the session traces, the training logs, the scored samples, and the one utterance the second essay follows. The measurements and the tools that draw them will be published with the source.