Audio Is The Master Clock

The brief is 1 vertical 9:16 night-street interview outside an upscale club: about 30 seconds, 2 actors in English, word-accurate, over a continuous city-night bed. Before a single candidate was opened, 2 requirements were written down and fixed. Accurate English lip sync. A vertical frame. 5 engines were tabled against them and 4 were rejected. The first carried native synced speech and was still rejected, for 1 reason: the frame is fixed at 16:9 with a 768P ceiling, so it can produce neither a true portrait plate nor the 9:16 the card needs. It is the right model for a landscape dialogue beat and the wrong 1 for this. 2 video-only models went next and both fail the same way, because there is no speech pass anywhere in them, so the mouth moves to nothing in particular. A third, a text-to-video avatar, was rejected on that identical ground. 5 candidates, 4 rejections, and every 1 decided against a requirement written down before the engine was opened rather than an opinion formed after the first render came back.

The engine that survived is an audio-driven talking head. The still supplies the identity and a WAV we generate ourselves supplies the words. That single property is what makes it correct for this card, because word accuracy stops being a prompt hope and becomes structural: if the WAV says the line, the mouth covers the line. The same model answers 720p and 1080p, accepts a portrait input, carries no length cap, and runs on a fast pipeline with no GPU reservation, so a 30-second vertical card sits comfortably inside it. A second audio-driven model is held in reserve, also uncapped, in case the first fails a gate. Nothing about that ordering is accidental. A model that generates video and then hopes speech lands on top of it is making a bet on the audio. A model that takes a WAV as an input is making the audio a precondition.

Because the engine is audio-driven, the voice is generated and measured before the face is rendered. That inverts the ordinary text-to-video order, and it is the 1 decision the whole rest of the build hangs off. The audio pass runs first: 4 lines, 2 characters, 1 locked voice each, recorded to a measured line plan rather than an estimate. The 4 lines measure 3.48 seconds, 9.94 seconds, 2.09 seconds and 2.04 seconds. Spoken total: 17.55 seconds, inside a 30-second card with handles either side of every line. Only once those numbers exist does the shot list get built, because the shots exist to cover the words rather than the words existing to fill the shots. Audio is the master clock. Every 1 of the 4 takes is cut to it.

2 characters carry this scene and each 1 has a single locked voice, chosen for the register rather than the gender. The interviewer gets a warm, quick host energy. The woman gets a playful, bright, confident read. Neither voice is re-rolled between lines, because 2 reads of the same character sound like 2 different people. The dialogue is written first and its length is measured in seconds, 17.55 of them across 4 lines, and the 4 shots are then ordered to cover that length: a medium two-shot at the club doorway, a close-up on her, back to the two-shot, then a final close-up. The longest line by far is hers at 9.94 seconds, which is why that 1 gets the close-up while the two-shot carries the short questions. The frame is fixed at 9:16 for all 4 shots. The rule is the whole point of the ordering: the words are written and measured before the face is rendered, so no render is ever spent on a mouth with nothing accurate to say.

5 engines were tabled. 4 were rejected before a single frame was rendered.

4 shots that do not share a world will not cut, so the world block is pasted byte-identical into every 1 of the 4 shot prompts: warm amber light spilling from the club doorway onto wet pavement, a cold white LED tube above it, neon spilling red and blue, soft blurred city lights receding down the street, handheld framing with the horizon a couple of degrees off, available light only, skin unretouched with pores and texture visible, deep focus with the street legible kerb to shutters. Identity is locked the same way. The two-shot is generated first, and her close-up is generated from that plate as a reference-guided edit, so her face is the same face rather than a re-roll of the same adjectives. 1 upload round-trip buys the identity across the cut. The negative stack is deliberately not the shared global 1. That block carries a de-identifying stack written for a different register, and on an interview the faces must stay readable, so this project uses the anti-AI block only: gloss, plastic, cartoon, text, anatomy.

The night bed under the scene is constant and level-matched, never sidechain-ducked. That rule was carried in from the last build, where keying a duck off the finished master collapsed the floor wherever the master carried a transient and shipped 10 near-silent seconds while every container check passed. Each bed stem is trimmed to a common -37 dB before the crossfade, the finished film goes through two-pass loudness normalisation to -14 LUFS and -1.0 dBTP, and then it is re-measured per second: the build fails if any 1-second window falls below -42 dB. That is the shape of the pipeline, and it is the shape EX Venture runs: 35 staff and 200 plus interns a year, work built across 12 countries, 393 pages in the book, 450 plus generated articles and films behind the pipeline, and every order passing a named check before it ships. The probability that 2 different crews shoot the same 30 seconds the same way rises the moment the requirements are fixed before the engine is opened, and it rises again when the audio is measured before the face is rendered. Write the words. Measure the words. Then render the face that covers them.