Synthesia Workflow for Text-Led Video That People Finish

Most organizations do not run out of things to say. They run out of things to show. A policy update, a research summary, a certification module, a quarterly results briefing, a technical explanation of a process that happens inside a database: none of these arrive with natural footage attached. The substance is text. The screen wants pictures. Avatar-led tools such as Synthesia solve the presenting problem by putting a lip-synced presenter on screen in minutes, but they do not solve the translation problem. A document read aloud is still a document, only slower and harder to skim.

This is a production workflow for the harder half: turning text-led material into video that a real person will finish. It covers the pre-production decisions that prevent rework, script conversion for the ear, visual choices for abstract ideas, pacing, quality control, accessibility, and the mistakes that make an informational video feel like a narrated PDF.

The bottleneck is translation, not rendering

Written text is built for the eye. It is hierarchical, skimmable, and dense. A reader can jump to section four, reread a difficult sentence, ignore a footnote, and scan bold labels for orientation. Speech is linear and time-bound. When a document becomes a video, every structural affordance disappears unless the producer rebuilds it deliberately.

There is also a speed mismatch. Comfortable listening pace sits around 140 to 160 words per minute. A 1,500-word internal memo read aloud becomes a ten-minute video. Ten minutes of a single presenter on a static frame is a large request for an audience that never opted into a lecture. The result is usually a video that is technically accurate, professionally lit, and practically unwatchable.

Three failure modes appear again and again.

• The recital. The avatar reads the document in order, including the definitions section, the disclaimer, and the bullet list of contact details. Nothing has been rewritten for the ear.
• The deck with a face. A presenter sits in a corner while full slides slide past. The viewer reads and listens to the same words at the same time. Attention splits and both channels lose.
• Wallpaper footage. Generic stock clips cycle behind narration about risk, growth, alignment, or compliance without any relationship to the sentence being spoken. The visuals become noise the viewer learns to ignore.

Each failure is fixable, and the fixes happen before generation. Rendering is the fast part. Translation is where the project succeeds or stalls.

What an avatar presenter is actually good at

Avatar video is not a downgrade from live action. It is a different medium with a specific strength: an authoritative person delivering structured information, consistently, in many versions, at low marginal cost. That strength maps well to onboarding, compliance, product training, explainers, internal announcements, and any content that needs a stable human face without a studio booking.

It maps badly to anything that needs hands on an object, spatial demonstration, emotional storytelling, comic timing, or a live reaction. A presenter can describe how a valve is replaced. A camera should show it.

Use a simple decision test.

Message type	Better format	Why
Here is what changed and what you must do	Avatar presenter	Authority, clarity, easy re-versioning
Watch how this assembly comes apart	Live or generated footage	Requires physical demonstration
Follow along in this dashboard	Screen recording	The interface is the subject
Here is the idea behind our strategy	Avatar plus generated visuals	Needs a metaphor layer
We need to explain a sensitive policy	Avatar with careful scripting	Tone control and consistency

Most real projects are hybrids. A seven-minute compliance module might use an avatar for 80 percent of the runtime and cut to a screen recording for the two minutes where an actual system is being shown. That is the correct answer far more often than a purist one.

A pre-production brief that prevents most rework

Changing a presenter after the visuals are built is a reshooting event, not a tweak. Replacing a voice after captions are timed is a re-editing event. The cheapest hour in the workflow is the one spent deciding the production rules before anything is generated.

Write a one-page brief that answers eight questions.

1. Who is watching? A new hire, a regulated audience, a customer, or a busy executive needs different levels of context and different proof.
2. What should they do after watching? Submit a form, approve a change, pass a quiz, understand a trade-off, or simply know that a policy exists.
3. What is the maximum runtime? Pick a number before scripting. A five-minute limit forces editorial judgment; an open limit invites a recital.
4. Which presenter represents this message? Match the presenter to the context, not to personal taste. A lab coat in a finance explainer breaks trust instantly; a hoodie in a regulated-industry compliance module can do the same.
5. Which voice belongs to this series? One voice per series, written down, with pronunciation overrides.
6. What visual system holds the series together? Palette, type treatment, diagram style, and whether generated footage is allowed.
7. Where will the video live? An intranet page, a learning platform, a public site, and a sales deck impose different caption, aspect ratio, and length requirements.
8. Who owns updates? Text-led content goes stale faster than almost any other kind. Name the person who will refresh it when the policy or product changes.

For a training series, the brief may say: audience is new support staff, action is to resolve a refund request correctly, maximum runtime is six minutes, presenter is a friendly operations lead, voice is warm and measured, visual system is white diagrams on a deep blue field, distribution is the learning platform, and the update owner is the support enablement lead. That page is enough to keep five episodes consistent.

Keep a style sheet next to the brief. It should list presenter name, voice name, frame, background, caption style, pronunciation rules, and banned visual clichés. Every future episode starts from that page, not from a blank project.

Script surgery: rewrite for the ear, then chunk into scenes

The single highest-leverage task in the entire workflow is rewriting the document before it reaches the avatar tool. Most videos that feel robotic are not failing because of the voice model. They are failing because the script was never written for speech.

Rules that pay for themselves immediately:

• One idea per sentence. Split anything with a semicolon, a long dash, or the word whereas.
• Target 15 to 22 words per sentence. Long sentences force unnatural breath points and make the viewer hold too much in working memory.
• Use active voice and second person. You will submit the form by Friday beats submission of the form is required.
• Expand abbreviations on first use, then define pronunciation. Say the letters or the spoken form deliberately, once, and lock it in the tool's pronunciation settings.
• Convert lists into spoken enumerations. First, second, third gives the ear the structure bold gave the eye.
• Cut parentheticals entirely. If the aside matters, make it a sentence. If it does not, delete it.
• Put the action before the explanation. Viewers decide whether to keep watching in the first fifteen seconds. Tell them what they will be able to do, then explain why it matters.

A before and after shows the difference.

Document: Employees who have not completed the mandatory data-handling module, available via the intranet portal under Compliance then Training, by the end of the quarter will be flagged in the audit review process.

Script: You have until the end of the quarter to finish the data-handling module. You will find it on the intranet, under Compliance, then Training. If it is not done by the deadline, your name appears in the audit review. It takes about twenty minutes.

The second version takes a few seconds longer to say and is far more likely to be acted on. It also gives the visual team something concrete to work with: a calendar, a path through the intranet, a clock, and a checklist.

Once the script is clean, chunk it into scene units of two to four sentences, roughly 20 to 40 seconds each. Those chunks become production units: one avatar shot, one visual idea, one caption block. Keep the chunked script in a single master document with numbered scenes, and treat every downstream asset as an attachment to a scene number. When the policy changes later, scene seven is easy to find.

Presenter, voice, and frame: lock them before generation

Avatar choice is a trust decision. The presenter should feel like the person this audience would accept delivering this message. A healthcare explainer may call for a calm clinician in plain clothes. A software onboarding series may call for a product specialist who looks like a real colleague. A public safety announcement may call for a uniform only if the uniform is accurate and appropriate.

For a series, pick one presenter and keep them. Recognition is worth more than variety. A second avatar should represent a distinct show, not a mood change in the middle of a module. If the topic needs a different authority, such as a legal expert appearing for one section, introduce that person clearly and return to the main presenter.

Voice needs the same discipline. Choose one voice per series and document it. Set pronunciation overrides for brand names, acronyms, product codenames, and any number that could be read two ways. Slower than feels natural is usually correct, because the presenter is not in the room and cannot read confusion on a viewer's face.

Frame and layout should remain identical across the series unless there is a narrative reason to change. A medium close-up holds attention for straight narration. A side-panel layout works when the visuals genuinely carry information. Full-body framing rarely earns its screen space in a text-led video. Layout jumps read as production errors, not emphasis.

A useful test is to watch three episodes with the sound off. If the presenter, background, caption position, and visual palette look like they belong together, the style system is working. If each episode looks like a different vendor made it, the style sheet is either missing or being ignored.

Building a visual grammar for abstract material

The hardest translation is abstract nouns. Risk, alignment, efficiency, culture, scalability, resilience, and governance have no natural picture. If you leave them to generic stock footage, the viewer gets wallpaper.

Use a four-second test. For each scene, ask what a viewer should be able to point at four seconds in. If the answer is the presenter, the scene needs work. If the answer is a specific object, diagram, number, or place, the scene has a visual anchor.

Four visual types cover almost everything in text-led video.

• Process diagram. Use for anything sequential: approvals, onboarding steps, escalation paths, procurement cycles. Build it once, then animate it in stages that match the narration. A diagram that fills in as the presenter speaks is easier to follow than a finished diagram that appears all at once.
• Data callout. Put one number or one comparison on screen long enough to read. A single animated figure landing on the beat does more than a chart with twelve series. If the number needs context, show the comparison as a simple before-and-after rather than a dense dashboard.
• Concrete metaphor. For abstract ideas, generate a specific physical image rather than a generic one. Risk becomes a slow push-in on stacked paper under a single lamp. Scalability becomes identical blocks multiplying across a surface. A bottleneck becomes traffic compressing into one lane. Specificity is what makes the metaphor land.
• Real footage or screen recording. When the subject is a system, a place, or a physical action, show it. Nothing generated beats the actual interface when the viewer needs to recognize the screen they will use.

When you generate metaphor shots or b-roll, write prompts like a shot list, not like a wish. Subject, action, setting, lighting, lens, and mood, in that order. Keep a fixed palette and grain across the series so scenes cut together without a visible seam. The prompt library has reusable structures for this, and the AI video generator is a fast way to test a metaphor before committing to it across eight scenes. For style frames that need to match an existing diagram system, the AI image generator can produce reference stills before motion enters the pipeline.

Slide hygiene matters as much as b-roll. Five words on screen, maximum. The slide should not repeat the sentence being spoken. It should carry the part speech cannot: the number, the shape, the sequence, or the contrast.

Pacing: how to stop a talking-head video from becoming a metronome

Avatar video does not fail because it is boring. It fails because it is uniform. One presenter, one frame, one vocal cadence for nine minutes becomes a metronome. Pacing is how you break the metronome without turning the video into a distraction.

Length targets that hold up in practice:

• 60 to 90 seconds for announcements and single-action requests.
• 3 to 5 minutes for explainers with three or four beats.
• 6 to 8 minutes for training, but only with visible chapter markers and a stated roadmap.

Rhythm rules:

• Change something every 4 to 7 seconds. A cut, a diagram step, a caption, a frame-size change, or a new object on screen. Not every change needs to be a new scene. Small changes do the work.
• Trim narration gaps. Long silences feel like buffering. Keep pauses at 0.3 to 0.6 seconds, and use a slightly longer pause only after a key claim.
• Use a repeatable arc. Hook in ten seconds, roadmap in fifteen, three content chunks, a fifteen-second recap, and a ten-second action. The arc is invisible when it works.
• Read the script aloud. If you run out of breath, the avatar will too, in a way viewers can hear.

A practical 90-second structure looks like this: a direct statement of the change, the one action required, the deadline, the consequence of missing it, where to get help, and a final recap. That is enough for most internal announcements. A six-minute training module needs a different shape: a realistic scenario, the rule, a worked example, a common mistake, a check for understanding, and a short summary.

One useful trick is to build the shortest version first. Cut the script to 60 seconds, produce it, then decide what the extra three minutes actually add. Most projects discover that the long version was never necessary. The short version also becomes the clip people share in chat, which is often how the full module gets discovered.

Assembly, versions, QA, and accessibility

Assembly is where sloppy projects lose their gains. Name every asset by scene number, such as s04-risk-broll-v2. Keep one master project per episode and store the chunked script next to it. When the policy changes in six weeks, you should be able to find scene seven in under a minute.

Plan three versions from the start: the master, a short cut for internal feeds, and any language variants. If you need multiple languages, generate each language separately rather than relying on dubbing. Lip sync and mouth shapes are noticeably better, and text expansion in German or Japanese can break a tight layout. Keep the layout flexible enough for longer captions and wider words.

Quality control should be a checklist, not a feeling.

• Pronunciation of every name, acronym, and product term.
• Numbers, dates, and currency read the way a human would say them.
• Disclaimers on screen long enough to read, not just long enough to exist.
• Captions accurate, including technical terms, with sensible line breaks.
• Contrast checked against the background, and color never the only signal for meaning.
• Consistent presenter, voice, and layout across the series.
• Mobile check: captions and on-screen text readable on a phone without zooming.

Accessibility is not optional for the content type most likely to need an avatar presenter. Add captions and a transcript. Describe visuals in the transcript where they carry information the narration does not. Avoid rapid flashing and unnecessary motion. Keep on-screen text large enough to read on a small screen, and do not rely on red versus green alone to communicate status. If the video is part of a public or regulated communication, check it against the accessibility standard your organization uses and keep a record of the review.

Mistakes that flatten text-led video

• Uploading the document as a script. The tool will faithfully read your worst writing. Rewrite first.
• One avatar shot for eight minutes. Add diagrams, cutaways, frame variation, and captions that carry structure.
• Decorative stock footage. If the visual does not illustrate the sentence, delete it. A clean diagram beats irrelevant footage.
• Reading slides aloud verbatim. Viewers process both channels and retain neither. Let the slide carry the number or shape while the voice carries the meaning.
• No pronunciation pass. One mispronounced product name undermines the whole video, especially in a customer-facing explainer.
• Skipping the short version. A 60-second cut often outperforms the full module for awareness and sharing.
• Rebuilding from scratch each time. A style sheet, scene-numbered assets, and reusable video templates cut production time by more than half after the second episode.
• Treating delivery as the finish line. Plan the update cycle. Text-led content goes stale faster than any other kind, and a stale training video teaches the wrong process with confidence.

FAQ: questions that show up after the first episode

Can we produce several videos per week without losing quality?

Yes, if the bottleneck stays in scripting rather than rendering. Keep a style sheet, a scene-numbered asset library, and a fixed visual palette. Teams that standardize those three things routinely ship weekly episodes with the same person-hours a single video used to take. The risk is not speed. The risk is skipping the editorial pass because the tool makes generation feel easy.

Should the avatar appear in every scene?

No. Use the presenter for framing and emphasis, and cut away for evidence. A rough ratio of 60 percent presenter and 40 percent visual carries most informational content without feeling either static or fragmented. For a highly emotional message, the presenter may deserve more screen time. For a process walkthrough, the diagram or screen recording should dominate.

How do we handle a document that is genuinely long?

Split it into a series with a shared intro and a short recap per episode. Ten minutes of continuous narration is a hard sell. Four three-minute episodes with numbered titles get watched, rewatched, and linked individually. If the material must remain one video, add chapter markers and state the roadmap in the first twenty seconds so viewers know what they are committing to.

What is the minimum viable quality bar?

A stranger can summarize the main point after one viewing. Every name is pronounced correctly. Captions match the audio. The visuals never contradict the narration. Those four checks catch most embarrassing errors. A fifth check is useful for regulated content: a subject expert confirms that the simplified script did not accidentally change the rule.

Where does generated footage fit if we already have an avatar?

Exactly where stock would have gone, but with more specificity. Use it for metaphors, transitions, abstract concepts, and any scene where the archive has nothing honest to offer. Keep one palette across generated and real footage so the result looks deliberate. Do not use generated footage to fake a real product interface or a real location when accuracy matters.

How often should avatar video be refreshed?

Whenever the underlying policy, product, or numbers change, and at least on a fixed annual review cycle. Because updates are cheaper than reshoots, schedule them rather than waiting for a complaint. Keep a change log tied to scene numbers so the next update is a targeted edit, not a rebuild.

Do we need a different presenter for every audience?

Usually not. One presenter can serve several audiences if the script changes. A separate presenter makes sense when the series brand changes, when the authority type changes, or when the language version needs a different cultural context. Too many presenters make a training library feel disjointed.

Where Orelon fits in the workflow

Avatar video gets your words on screen. The visual layer is what makes people stay. Once the script is rewritten for the ear and chunked into scenes, the remaining work is imagery: metaphor shots, abstract concept b-roll, diagram backgrounds, and stylized transitions that hold a series together.

That is where Orelon earns its place in the pipeline. It is an AI video generator built for cinematic ideas in motion, useful for the scenes a camera cannot easily capture and a stock library cannot honestly represent. Start with the hardest abstract noun in your next document. Generate three visual interpretations in the AI video generator, compare them against the narration, and keep the one a viewer could point at four seconds in. Build the rest of the episode around that standard. Browse the Orelon blog for more production workflows, and turn the driest document on your roadmap into something people actually finish watching.