Priya runs learning and development for a 400-person logistics company. Her onboarding video was two years old, referenced a warehouse layout that no longer existed, and only existed in English. Rebuilding it the traditional way meant booking a studio, hiring a presenter, and chasing a budget approval she wasn’t going to get.
So she opened Synthesia on a Tuesday morning, pasted in a rewritten script, and had English, German and Japanese versions rendered before lunch.
That’s the upside. The part nobody warns you about is that your first render will look slightly off in ways you can’t immediately name. Here’s the process in the order you’ll actually do it, plus the fixes that save the most time.
Gather four things before you open the editor
Skipping this step is the single biggest reason first projects drag into a second week.
- A locked script. Not a draft. Every line change means re-recording the avatar’s speech for that scene, and it’s fiddly to track which scenes you’ve already re-rendered.
- Brand assets in the right formats. A transparent PNG logo, your hex colour codes, and a font you actually have a licence for. Synthesia can’t use the font sitting on your designer’s laptop.
- Three to five reference videos. Anything your team already considers “good” is more useful than a written brief. Screenshot the layouts you like.
- One person with approval authority. A committee reviewing a 90-second video is how projects die.
Step 1: Decide between a template and a blank canvas
Synthesia ships with templates built around common jobs: training modules, product explainers, sales outreach, internal announcements. Each comes pre-loaded with a handful of scene layouts, a suggested avatar and a colour scheme.
The rule of thumb that works: if more than 60% of the template’s structure matches what you’re making, start there and edit it. If less, start blank. Fighting a template you’ve outgrown takes longer than building from scratch, because you’ll spend twenty minutes deleting elements you never wanted.
It’s worth understanding why so many teams arrive at this tool in the first place. The shift toward AI-generated corporate video has happened quietly and fast, and it’s changed what an L&D or comms team can realistically produce in-house.
Step 2: Write the script for a synthetic presenter, not a document
An avatar reading a paragraph written for the eye sounds robotic. Three adjustments fix most of it.
Keep sentences under 20 words
Long sentences with nested clauses are where AI delivery falls apart. “The system, which was updated last quarter and now includes automated routing, will notify you” becomes “The system was updated last quarter. It now routes automatically. You’ll get a notification.”
Write numbers and acronyms the way they should be spoken
Type “thirty” rather than “30” if you want it read as a word. For internal jargon like SKU, ERP or KPI, spell it out phonetically in the script (“S-K-U”) if the voice reads it as a word instead of letters.
Break long paragraphs into separate scenes
Rather than inserting commas to create pauses, cut the paragraph in two and put each half in its own scene. You get a clean visual change and a natural breath at the same time.
Step 3: Cast an avatar that matches the job
Stock avatars are ready immediately and cover a wide range of ages, ethnicities and wardrobe choices. A custom avatar, built from footage you supply, takes a few days and a consent form but puts a real person on screen. For a CEO’s quarterly update, that’s usually worth it. For a compliance refresher that gets watched twice a year, it isn’t.
Match the avatar’s energy to the content. A cheerful presenter reading a redundancy announcement reads as tone-deaf. The teams that get this right tend to think about avatar selection the same way they’d think about casting, and there’s a useful breakdown of what these AI avatars genuinely can and can’t pull off in this look at AI avatar limits if you want to sanity-check your plan.
Step 4: Choose the voice, then slow it down
Stock voices are fine for most projects. Cloning a voice takes about five minutes of clean audio, recorded in a quiet room, roughly 40cm from the microphone, with no background hum from a fridge or an air conditioning unit.
Here’s the tip that improves output more than any other setting: drop the speed to 0.95x. Synthesia’s default pace is built for marketing videos, and it’s noticeably quick for training or technical content where people are trying to follow along.
Step 5: Build scenes and layer the visuals
Work scene by scene and resist the urge to perfect scene one before moving on. Get the skeleton built, then refine.
Useful visual elements to rotate through:
- Short screen recordings of the actual software, which beat static screenshots every time
- Simple icon-and-caption layouts for processes and lists
- Full-bleed brand photography with the avatar in a corner insert
- Text-on-background title cards between sections
One warning: don’t put a full sentence on screen while the avatar reads it aloud. Nobody watches a video to read. Cut the on-screen text to a label or a number, and let the voice carry the detail. If you’re sourcing b-roll and want a sense of what other AI video tools handle well, a closer look at prompt-to-video generation is a reasonable comparison point.
Step 6: The last 10% that eats 40% of your time
This is the polish pass, and it’s where amateur-looking output becomes credible.
Pronunciation. Proper nouns and product names will get mangled. Override them in the script or use the per-word pronunciation field, which lets you type out how the word should sound.
Eye-line. Avatars look straight down the lens by default. For side-by-side layouts where the avatar sits next to a diagram, that can feel intense. Switching to a different avatar pose or layout often fixes it faster than fiddling with settings.
Gesture density. Avatars with big, frequent gestures look great in full frame and distracting in a small corner box. Pick a calmer presenter for insert layouts.
Before you build all twelve scenes, render a 20-second test containing your hardest words and your smallest on-screen text. Catching a problem here costs two minutes. Catching it after a full build costs an afternoon.
Step 7: Localise, then check for overflow
Translating a video into eight languages used to be a five-figure project. In Synthesia it’s a dropdown.
The catch is text length. German expands by roughly 30% compared to English. Japanese wraps differently. Text that fits neatly in your English scene will overflow its box in translation, so check each localised version scene by scene rather than trusting the preview. Different accents and delivery styles also land differently with different audiences, and this comparison of AI presenter approaches is worth reading before you standardise on one voice across every market.
A worked example: 12 slides into four minutes
Take a typical safety briefing built from a 12-slide deck.
The slides merge into nine scenes. The script lands at 260 words, which renders to roughly 4 minutes 10 seconds at 0.95x speed. Writing and revising the script took 90 minutes, since it was really a rewriting job rather than a writing one. Building the scenes took 35 minutes. The polish pass, pronunciation fixes and localised overflow checks, took another 25. Total: just under three hours for a video that would previously have been a two-week supplier project.
That ratio is the whole point.
When the render comes out wrong
Most first-time problems fall into four buckets, and each has a quick fix.
Audio sounds clipped or rushed. Slow the speed to 0.9x and split the scene. If a single sentence is longer than about 25 words, no setting will save it.
The avatar blinks or moves oddly. Usually a sign that the scene is very short, under three seconds. Merge it with the neighbouring scene.
On-screen text is cut off. Shrink the font one step and shorten the copy. Never rely on the safe-area guideline alone.
A name is still mispronounced. Replace the word in the script with a phonetic spelling, render just that scene, and re-export. You don’t need to rebuild the whole video, which is the feature that keeps teams using the platform after the novelty wears off. For anyone still weighing up whether synthetic video belongs in their workflow at all, the story of how these tools spread through corporate teams explains the pattern better than a feature list can.
Start with something small and low-stakes this week. A three-scene internal update is enough to learn the interface, hit the pronunciation snags, and see how the final render actually feels to watch. The second video takes a third of the time.