How it is made

Every voice here is synthesised. That is said at the door, not in a footnote.

The declaration

The scripts for this show are written with machine assistance. The conversations are performed by an open speech model — Microsoft's VibeVoice, running locally — and no voice in the show is a clone of a real person's recording. Nothing is sent to a service; the whole thing is made on one computer, offline, including the transcription that checks it.

The method this house uses, and the reasoning behind declaring it, is published and citable: doi.org/10.5281/zenodo.21914929.

A listener who finds this out afterwards is entitled to feel misled. A listener who knows it going in can just decide whether the argument is any good.

Why a conversation, and not a reading

Nearly every open speech model synthesises one line at a time. Line by line, a four-person conversation becomes four monologues cut together: the prosody resets at every boundary, nobody's pitch answers anybody else's, and the room has no air in it. VibeVoice generates the whole conversation in one pass — up to ninety minutes, four speakers, turn by turn — so the third speaker's line lands because of how the second one finished theirs. That is the entire reason the show is built on it, rather than on the model with the prettier single voice.

The writing comes first

Every script for every episode in a season is written before a single one is recorded, so a later episode can quote an earlier one word for word. Recording is the cheap part; the show is won or lost on the page.

Two tiers, one published

A fast model exists only so a whole episode can be heard back while the writing can still change. Nothing it renders is ever published. Everything a listener hears comes off the large model at full precision — a setting arrived at by measurement, not by preference.

III

Every take is listened to

A take can be the right length, at the right level, and say the wrong words. So every take is transcribed back and compared to the line it was given; anything that misreads more than a quarter of it is read again on a different seed rather than shipped.

The room is not silent

Takes are joined with two hundred milliseconds of breath and a bed of room tone underneath, because silence between takes is a hole you can hear. It is the oldest trick in radio and it is the difference between an episode and a file.

Levelled for a phone speaker

Two-pass EBU R128 to −16 LUFS with a −1.5 dBTP ceiling, which is what Apple and Spotify actually want for speech, and then measured again to check it landed. A mastering pass that does not check what it produced is a hope.

The negative is kept

Alongside every published file there is a lossless master at the original rate, unlevelled — so a re-master years from now does not start from a compressed copy of a decision made today.

What the people are

The nineteen at the table are written characters, not real people and not portraits of anyone. Each has a fixed age, place, work, gift and one thing they never do, and none of that changes when it would be convenient. Their voices are built from timbre references that carry no identity — several of them from the speech voices built into macOS — so no living person's voice has been taken to make this show.

Everything published here is CC BY-NC 4.0: share it, quote it, teach with it. Nobody resells it.