Synthetic media production

What this actually solves

The reason is almost always scale or maintenance, and usually both. A company has a training library that has grown to dozens of videos. Every year policy shifts and a portion of them go out of date. Bringing them back means a studio booking, a presenter's day, and an editor for each one — and because that is expensive and slow, the sensible commercial decision is to leave them stale for another year. Meanwhile the accurate version of the content exists only as a slide deck nobody opens.

The second case is language. A product demo exists in English and needs to work in German, French, Spanish, Arabic and Japanese. Done conventionally that is a separate presenter, a separate shoot and a separate edit for every market, which is why the temptation is to decide quietly that English is enough and accept the drop-off in the regions where it is not. Synthetic production changes the unit of work: one recording, one script, and a render per language with the mouth movement matched to the new audio.

What is genuinely different here is not the rendering — the tools are commodity and anyone can buy them. The variable is the script, and it is the script that decides whether the finished video works. A good script survives a mediocre avatar. A weak script is not rescued by the best avatar on the market, and the failure is quiet: the video is technically fine, publishes cleanly, and nobody watches it through.

What we do, and in what order

1. Script first, and it takes about half the project's hours. Written to be spoken rather than read — short sentences, ordinary words, no clause a person would not say out loud. It opens on a situation the audience will recognise from their own week, marks a visual change roughly every forty-five seconds so no avatar shot runs too long, annotates each section with whether the avatar is on screen or the screen is showing something else, and closes on three things to remember. Two to five days.
2. Avatar and voice, with consent handled before anything is generated. Either a stock presenter chosen for the audience, or a custom one built from three to five minutes of clean footage of your own person. Voice is chosen or cloned. Written consent comes from the individual whose face and voice these are, not only from their employer, and it is captured before any generation starts rather than retrospectively. One to two days.
3. Generate section by section and review the seams. Not one long render. Building in sections means a paragraph can be reworked without redoing the whole piece, and it exposes the two things that most often read as artificial — mispronounced acronyms and numbers, and a delivery pace that never varies. Both are fixed by editing the script, which is why the script is not closed until the render has been heard. One to three days.
4. Edit into something that looks like your organisation made it. B-roll, screen recordings and demonstrations at each marked visual change, captions burned in for the substantial share of viewers watching without sound, brand colours, logo and consistent lower thirds so eight videos look like one series rather than eight separate purchases. One to three days.
5. Localisation, with a human between the translation and the avatar. The script is translated, then checked by a person who speaks the language before a single frame is rendered — because the failure mode here is not gibberish, it is phrasing that is grammatically correct and unmistakably foreign to the one audience the version exists for. One to two days per batch of languages.
6. Quality pass and disclosure. A fixed checklist before anything publishes: pronunciation of every proper noun and acronym, caption timing, audio levels across sections, the disclosure line where the material is customer-facing, and a full watch-through at normal speed rather than a scrub. One day.

The stack is unremarkable and worth naming so you know what you are buying: HeyGen or Synthesia for avatar video, ElevenLabs for voice and cloning, Descript for transcript-driven editing, CapCut or Premiere for the final cut, and Runway or Kling where b-roll or an effect shot is needed.

What moves the price

The range is $500–$6,000 across 1–4 weeks, and three things push a project along it.

How many videos and how long each runs. A single three-minute explainer and a series of eight five-minute training modules are different orders of work, though the series gets cheaper per video after the first because the template, the presenter and the styling are already decided.

Whether we write the script or you supply one. Scripting is around half the effort, so this is the single largest swing in the quote. A supplied script still gets a spoken-language pass — most documents written to be read are unusable as narration without one.

Languages, and whether a custom avatar is needed. Each additional language adds a translation, a human check and a render. A custom avatar built from your own presenter adds a capture and consent step that a stock presenter does not.

What usually goes wrong

All the effort goes into the tool and the script comes out of a chatbot in one pass. It is the most common failure in this service by a wide margin, because the tools are the novel part and the script is the boring part. The result renders perfectly, ticks the box on the training matrix and is abandoned partway through by the people it was made for. If half the hours on a project are not in the script, the project is being built the wrong way round.

Consent is taken from the company rather than the person. A manager approves the use of an employee's face and voice, everyone is comfortable, and eighteen months later that employee has left — possibly not on good terms — and their likeness is still presenting your onboarding. Consent must come from the individual, in writing, and it should state what the material can be used for and what happens when they leave.

Translation is machine-generated and goes straight to render. The lip sync is flawless, the grammar is defensible, and every native speaker in the target market hears within a sentence that nobody who speaks their language was involved. The credibility cost lands precisely on the audience the localisation was bought for, and it is prevented by one review step that costs a fraction of the render.