The Read Is A Setting

The first version of the read failed, and it failed in a way that is hard to see from inside the edit. The synthetic voice was quick, pitched high and slightly wrong on every vowel. The script underneath it was fine. That is the expensive shape of this defect. A weak read is invisible in a waveform and obvious inside 1 second of playback, and it makes a written piece look worse than it is. The instinct is to rewrite the script. The rewrite is not the fix. What changed the result was a different engine and a different set of numbers inside it, and the numbers are the part worth writing down. A voice is a setting before it is a talent. 5 settings carry the outcome: the model, stability, the similarity boost, the style, and the speaker boost. Engine defaults are tuned for the widest possible case, which means they are tuned for nobody in particular. A default read is not neutral. It is somebody else's choices for an audience that is not this 1. Set the 5 by hand, once, and the same script in the same engine becomes a different product. Leave them alone, and every take gets argued about as if the words were the problem. The studio already runs 2 registers, 1 muted and editorial, 1 bright and quick. The settings are those same 2 registers written as numbers rather than adjectives, which is the reason a decision can be handed to anybody and reproduced on the next take. 1 page, 5 values, 2 temperatures.

The settings table has 5 rows, and 3 of the 5 are ranges rather than single values. That is the tell that they were chosen against real takes instead of copied from a default. The model first. The multilingual variant is the workhorse and the most natural of the 3. A turbo variant buys speed, and a flash variant buys the lowest latency of the 3. 2 of those 3 are trades rather than additions, and the trade is always the same 1: naturalness spent to buy milliseconds. Stability second, and this is the row that does the most work. The range is 0.30 to 0.45. Lower means more natural micro variation, the small unevenness that makes 1 line sound like speech rather than output. Anything from 0.70 upward reads flat and robotic, and 0.70 is not a subtle difference from 0.45. It is a different product. The similarity boost third, between 0.70 and 0.80. Its only job is keeping the same voice recognisable across takes, so 2 takes cut together sound like 1 person on 1 afternoon rather than 2 guests in a room. Style fourth, between 0.30 and 0.50. A little expression without going cartoon. Push past 0.50 and performance turns into impression, which is 1 step away from impersonation. Speaker boost fifth, and it is the only true toggle on the table. Switch it on for clearer articulation. 1 row, 1 value, no trade to weigh. So the usable summary is 5 rows, 3 ranges, 3 model options and 1 toggle. Everything else in the engine is decoration, and every 1 of these can be written next to the script before a single take is rendered.

The table produces 2 products and the difference between them is 2 numbers moving together. An energetic read for a short social piece nudges stability down to 0.30 and style up to 0.45. A contemplative editorial read holds stability at 0.45 and style at 0.25. Read those 2 lines slowly, because they carry the whole point. Same script. Same voice. 2 settings pairs. 2 different pieces with 2 different temperatures, and neither 1 of them is a compromise of the other. The mistake is to treat the 2 as moods of 1 take. They are not. A mood is discovered in the room and cannot be written down. A temperature is decided in writing, next to the line, before the render, and it survives being handed to a second person on a different day. There is 1 more consequence worth naming. A voice that drifts between the 2 temperatures inside a single piece reads as a mistake at 1 cut, even to a listener who could not name what changed. The drift is felt as restlessness long before it is diagnosed. So the temperature is chosen in the same place the hook is chosen, on the top of the same page, and it is circled before the first line is recorded. 2 pairs of numbers, 1 circle, 1 decision.

The engine reads punctuation rather than instructions, and 4 marks carry the timing. An ellipsis is a beat, and it lands the clauses in separate moments rather than 1. A comma is a short breath. A full stop is a full stop. A dash is a slightly longer dramatic pause. That is worth stating plainly because the previous engine accepted a pause written as a decimal number of seconds, 0.40 in the guide's own example. That syntax does not exist in this engine. A script still carrying the old markers gets them silently dropped or read out loud, and either 1 of those outcomes costs a take. The second half of the pacing is interjections, and the rule is to write the sound rather than the instruction. Stage directions in brackets are the wrong shape. A bracketed laugh is a request the engine cannot act on. Written as a real word, the same beat becomes a performance, and it lands with the timing a listener expects from a person who actually found the line funny. The guide carries 2 wrong lines and 2 right lines for exactly this. The wrong version brackets the beat. The right version spells it and then keeps the ellipsis that follows, which is the smallest possible change for the largest possible difference in the read. So the pacing toolkit is 4 marks, 1 dead syntax to remove, and 1 habit to acquire: anything the engine should perform has to be written as text.

A voice is a setting before it is a talent. Write the numbers down once and they reproduce on every take.

The guide tables 6 premade voices, each 1 behind a 20 character identifier. They are premade in the strict sense, so there is no cloning run, no consent step and no waiting for a model to learn anything. The 6 are distinctly different instruments rather than 6 shades of the same 1. 1 is warm and natural with an American read and works as the all-round default. 1 is soft and slightly younger. 1 is clear, young and energetic. 1 is bright and friendly. 1 is animated and carries the highest energy of the 6. 1 is warm, deep and confident. That is 6 voices, 0 cloning runs, and 0 permissions to negotiate before the first take. A cloned voice of a real person is better still if 1 is available, and getting it is 1 sentence of agreement from the owner plus 1 upload. The point of listing them is that the voice is a cast decision rather than a setting. Pick the instrument before the script is read, because a line written for the animated 1 and delivered by the soft 1 will sound like a mistake in 2 different ways at once. Cast, then tune. 6 instruments, 5 settings each, and the settings do not rescue a miscast read, which is the reason the order matters.

1 rule overrides the whole table. Energy never comes from speed. The read is not sped up to make a piece feel alive, because the rate is what makes it human in the first place. If a piece feels slow, the fix is in the edit: shorter holds, more cuts, a change of framing. That leaves the method in 1 line. Low stability plus real punctuation plus real interjections equals a human voice. 3 inputs, 1 output, and every 1 of them is a decision taken before the render rather than after it. There is 1 trap worth naming. A wobbly take invites the instinct to raise stability, which moves the read toward the flat and robotic end and destroys the variation that made it sound like speech. Raise the similarity boost instead and re-read the line. 30 years of building, in 12 countries, under 2 rules: work with what the constraint allows, and refuse the version that looks better on paper than it can look in the room. Built companies across twelve countries. Restructured an industrial group. Deployed 210 energy systems. Took a company through a Frankfurt IPO. Built 200 wood gasification machines at Entrade. Signed a personal guarantee on a company that folded on Christmas Day 2008. In the building for the 2003 blackout. 393 pages in paperback, published 8 April 2026. The probability that a studio which writes its 5 settings down before the first take ships a read that sounds like a person is high. The probability that a read left on engine defaults sounds like a person is close to 0, and it spends 3 takes and 1 re-record finding out. So the whole method fits on 1 index card, next to the script: model, 3 ranges, 1 toggle, 4 marks, 6 voices, 1 rule. Written down once, it stops being taste and becomes a setting that anybody on the team can reproduce.