Dead air, and why the words set a film's length

Every render already recorded, scene by scene, how long the picture held and how long the narration lasted. Nothing added the two together. When we did, two ads stood out: 13.0 seconds of silence in 31.1, and 8.0 in 28.5. One scene held its picture for 8.6 seconds on 3.4 seconds of words.

A viewer feels that instantly. The voice has finished its thought, the picture stays, and attention leaves. So the first job was simple: measure it on every film and say so when a film goes quiet.

The first warning fired on our own English home-page film, which is lively from start to finish. Before touching anything, we measured the gaps on every film our site shows. About one second per scene, almost everywhere. Narration time counted only runs of speech, so the breath between two sentences looked like silence when it was simply talking.

So the warning changed target. It now names the one scene that holds far past its words, with its silent seconds, because that is the real defect: a picture left on screen after the voice has moved on. A whole-film check remains as a backstop at a much lower threshold, worded as a film that drags, and the two never fire together. A warning should be rare enough that you act on it every time.

The next lesson came from animations, the films you write yourself. They had no target length, so an 87-second ad could pass comfortably inside the product's limits when the brief wanted thirty. We added a target, and saw at once that it could not be met the way a website film's is. The writer chose the words, and the words are the length. Stretching the picture would only create the silence we had just learned to detect.

So a target the script cannot meet is now declined before anything is spent, with the arithmetic:

• the word budget for that length, in that language and tone;
• the script's actual word count;
• how many words to cut or add.

The same idea guides storyboards you write by hand. Our length fitter steps aside when a person wrote the words, as it should, and the report now tells you exactly how many words to cut to reach your length.

Two reports looked like an unstable voice. Five renders of the same voice measured between 131 and 153 words a minute, which suggested the synthesiser drifted. It did not. Our rate divided the words by the voice track minus an estimated pause at each scene join, and on real films the real pauses were four to five times longer than estimated. The number moved when the punctuation moved; the voice did not. The steady rate, the sum of the measured speech, was already in the report, so we now return it and say which one to plan with.

The second report described sound slipping away from the picture. It was really the picture stopping. A generated clip has a fixed length, and a scene that held longer froze on its last frame while the voice went on. The clip's ending now plays out naturally rather than freezing, and the whole clip is never slowed down, which would have put people in visible slow motion.

Put together, a narrated film is now honest about its own time:

• it tells you when one of its scenes holds too long, and by how much;
• it tells you how many words to cut when a length is out of reach;
• it never charges for a length it cannot deliver.

The habit behind all of it is to measure the gap before choosing a threshold, and to ask which quantity really causes the effect. Here it was always the words. More of this thinking is behind how films are made.