> ## Content Index > Fetch the complete content index at: https://www.livain.com/llms.txt > Use this file to discover other available public pages before exploring further. # Why talking-head video looks scripted: finding gaze and repetition mechanically in the edit - URL: https://www.livain.com/blog/talking-head-video-gaze-repetition-mechanically-edit/ - Published: 2026-10-09T07:00:00.000Z - Updated: 2026-10-09T07:00:00.000Z - Description: Two notes on a dental-hygiene course reel turned into a repeatable method: eye-region contact sheets to locate every glance off-lens, and per-segment transcription so captions land on the take you actually kept. - Author: Remco Livain - Tags: Video, AI, Content Production, Dental, Brand At four frames a second, across two and a half minutes of footage, you can see every single time a person's eyes leave the lens. I know this because I built a contact sheet of nothing but the eye region, timestamped, and went through it looking for the moments that were going to make a video feel like someone reading off a screen. There were three. All three got cut. The video is fifty-three seconds long and the difference is the whole difference. ### Two rules, handed to me as a correction The piece was a short promotional reel for a dental hygiene association's regional section, advertising a one-day course for practitioners. I cut a first version from a multi-take recording, it was technically clean, and it was not right. The note that came back was two sentences, and they now apply to every talking-head edit I touch. Never say the same thing twice. And keep eye contact — no glances down, no glances at the screen, because the moment the eyes go sideways it reads as a teleprompter. What interested me is that both of those are findable mechanically. They sound like taste. They are not. ### You can find a glance without having an eye for it The repetition is the easy one. You have a transcript, you read it, duplicated ideas are obvious on the page in a way they are not in the ear. In this recording the speaker greeted the audience twice, said a near-identical closing line twice, and said "I'm looking forward to it" in two separate places. On screen those feel like warmth. Back to back in a one-minute cut they feel like padding. Gaze is the one people assume needs a human. It does not. Crop the eye region out of every frame at a few frames per second, tile the crops into a sheet with timestamps burned in, and look at the sheet rather than the video. A sideways glance that is impossible to catch at normal speed is unmissable as a grid, because the frames either side of it are looking straight at you and that one is not. Three passages went for gaze alone in this edit. One where she looked at the screen for most of a sentence about online reviews. One mid-sentence glance while listing course content. One clear screen-check right before a clause I liked. The last one hurt, and it went anyway, joined to the clause before it so the sentence still parses. Then the same pass over the finished render, which found one remaining head turn of about a quarter of a second — while smiling, mid-anecdote, and genuinely fine. Knowing it is there and choosing to keep it is a different thing from not knowing. > "Authentic" on camera is mostly two measurable things. The judgement is only in which take you keep. ### The research agrees, for once This is not just received wisdom from video people. There is a decent literature on it, and the finding is stronger than I expected: in a 2016 PLOS ONE study, Kreysa, Kessler and Schweinberger found that participants were [more likely to believe truth-ambiguous statements when the speaker's gaze was direct](https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0162291&ref=livain.com), and took measurably longer to reject the ones they disagreed with. Direct gaze does not just look better. It slows down the viewer's scepticism. Which is exactly why I would not use that as a persuasion trick and would use it as a quality bar. If a sideways glance costs you credibility on a true claim, the glance is the defect. ### The caption trap nobody warns you about Here is the technical bit, and it bit me properly on the first version. If you transcribe the whole source file and then use those word timings to place captions, a multi-take recording will destroy you. The speaker says a line, stumbles, says it again. The transcript contains both. The captions get placed using timings from take one while the picture is showing take two, and the words drift off the mouth by a second or more in a way that looks like a sync bug rather than an editing mistake. The fix is to transcribe each kept segment separately, after the cut decisions are made, and offset the resulting token times into the output timeline. More calls to the model, far less debugging. Word-level timing in Whisper is fiddly at the best of times — the whisper.cpp community has [an unresolved thread on getting dynamic time warping to behave](https://github.com/ggml-org/whisper.cpp/discussions/2307?ref=livain.com), and the [CrisperWhisper paper](https://www.isca-archive.org/interspeech%5F2024/zusag24%5Finterspeech.pdf?ref=livain.com) exists precisely because verbatim timestamps are a known weak point. Do not build a caption pipeline that assumes they are exact. German made it worse, as German does. The model split words into sub-word tokens, so captions broke mid-word in the middle of compounds, and a normalisation step helpfully removed the hyphen from a hyphenated product name. Both were fixed by merging tokens that carry no leading space before normalising, which is the sort of detail that is obvious in hindsight and invisible in advance. ### The rule that was not in my style guide One more, because it is the most transferable thing here and it has nothing to do with video. After delivery, the association came back with a correction: the three-letter abbreviation of their own name must never be used in public material. Central board rule. It appeared exactly once in my edit, on a date card, and it took five minutes to fix and re-upload in place so the link already shared kept working. I had no way of knowing that rule and no business assuming it did not exist. Client brand rules live with the client, not in your style guide, and the time to ask for them is before the first render rather than after the handover. I have [turned up to work with a brand that had no brand book at all](https://www.livain.com/blog/the-brand-book-didnt-exist/), which is its own problem, but the opposite case is sneakier: the rules exist, they are firmly held, and nobody thinks to tell you because internally everyone already knows. ### Why this matters outside a reel Dental and medical practices are sitting on the most persuasive marketing asset they have — clinicians who can explain things with authority — and mostly not using it, because the output looks stiff. It usually is not a confidence problem. It is two mechanical faults: the eyes leave the lens, and the same point gets made twice. Both are fixable in the edit, cheaply, by anyone willing to look at frames instead of vibes. And both are far better fixed by not doing them in the first place, which is the note I now give before the camera goes on rather than after. Look at the lens. Say it once. ### Sources & further reading **External** Kreysa, Kessler & Schweinberger (2016), [Direct Speaker Gaze Promotes Trust in Truth-Ambiguous Statements](https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0162291&ref=livain.com), PLOS ONE [CrisperWhisper: Accurate Timestamps on Verbatim Speech Transcriptions](https://www.isca-archive.org/interspeech%5F2024/zusag24%5Finterspeech.pdf?ref=livain.com) (Interspeech 2024) [Word-level timestamps with whisper.cpp and DTW](https://github.com/ggml-org/whisper.cpp/discussions/2307?ref=livain.com) (whisper.cpp discussion) **Related posts** [I replaced Loom with a single HTML file](https://www.livain.com/blog/replaced-loom-with-a-single-html-file-claude-code-buy-vs-build/) [Clinic advertising against empty appointment slots](https://www.livain.com/blog/clinic-advertising-empty-appointment-slots-paid-media-scheduling/) [The brand book didn't exist](https://www.livain.com/blog/the-brand-book-didnt-exist/) [Count patients, not bookings](https://www.livain.com/blog/count-patients-not-bookings/)