How to Set Video Length in Synthesia for Platform Limits

Why Duration Control Is a Strategy Problem, Not a Settings Slider

Most underperforming AI videos fail for a mundane reason: the runtime does not match the place it was published. The avatar looks fine, the voice is natural, the script is competent — and the video still gets scrolled past, skipped, or clipped mid-sentence by a platform that has its own idea of how long a video should be.

Duration is the first constraint a viewer feels, before they process a single word of the script. On short-form feeds, completion rate is one of the strongest signals the recommendation system reads. If your video is 78 seconds long and the format's natural loop is closer to 30, a large share of viewers leave before the payoff, and the algorithm reads that exit as a quality signal. On the opposite end, a 90-second training explainer that should have been four minutes leaves learners without the context they came for, which shows up later as support tickets rather than view counts.

This is why "set the video length" is rarely a single click. In a realistic AI content pipeline, you are balancing four forces at once:

Platform rules, which are hard ceilings and soft cultural norms — a Shorts feed tolerates three minutes technically but rewards far less.
Narration pacing, which determines how many words actually fit into a given runtime.
Model behavior, which decides whether a generated clip lands at the length you asked for or drifts by a few seconds.
Assembly overhead, the title cards, transitions, and end frames that add time you never scripted.

The practical goal is not to find one universal runtime. It is to build a repeatable method that turns a target duration into generation parameters, then verifies the result before you publish. Synthesia is a good anchor for that method because it is script-driven: if you can control the words, the pauses, and the scene boundaries, you can control the runtime with unusual precision.

How Synthesia Handles Video Length Under the Hood

Synthesia is built around a different mental model than a timeline editor. You do not drag a playhead to define the end of the video. You write narration, assign it to scenes, and the runtime emerges from those decisions. Understanding where length is actually determined prevents most runtime surprises.

The three places runtime is decided

1. The narration script. This is the dominant factor. Avatar narration plays at a fairly consistent words-per-minute rate, so the script length is the single largest contributor to total runtime. Any sentence you add or cut moves the finish line.

2. Scene boundaries and transitions. Each scene break introduces a pause — sometimes a fraction of a second, sometimes more if you add a transition animation or a beat of silence. A fifteen-scene video with generous transitions can carry several extra seconds of pure overhead compared with the same script in eight scenes.

3. Silence and pacing controls. Adding pause after a sentence, or a short lead-in before narration starts, is a deliberate pacing tool. It also changes the runtime. Two seconds of breathing room after a key claim is often worth it, but you should count it rather than discover it.

Why the editing timeline is the wrong place to fix a runtime problem

A common workflow mistake is to generate the video, watch the export, notice it runs 12 seconds long, and start trimming in a separate editor. That approach works, but it usually damages pacing: you cut the sentence that carried the call to action, or you speed-ramp the avatar's speech until it sounds unnatural.

A better rule: fix runtime in the script and scene structure, not in post-production. Trimming should be a last two-percent adjustment, not the primary tool. When you treat the script as the runtime control surface, a re-export becomes predictable instead of risky.

One more structural note: very long single-scene videos become fragile. If a scene carries several minutes of narration, any change to the script forces a full re-render of that entire block. Splitting long sections into shorter scenes gives you finer control when you need to shave eight seconds off the final export.

Translating Platform Constraints Into Generation Parameters

The useful question is not "how long should a video be?" but "what runtime does this specific surface reward, and what does that mean for my word count and scene count?"

Start with a constraint map. Treat the numbers below as working defaults and re-verify them periodically, because platforms change their limits without much ceremony.

Surface	Hard ceiling	Practical sweet spot	Aspect ratio
Vertical short-form feed	60–90 seconds typical	21–45 seconds	9:16
Extended vertical feed	up to several minutes	45–90 seconds	9:16
Professional feed / social timeline	~10 minutes	60–120 seconds	1:1 or 16:9
Standard video platform	long-form allowed	4–10 minutes	16:9
Course or onboarding module	no meaningful limit	3–7 minutes	16:9
Paid ad placement	varies by placement	15 or 30 seconds	mixed

Turning a ceiling into a target

A hard ceiling is not a target. A 60-second limit does not mean a 60-second video. For vertical feeds, a good default is to aim for roughly 70 percent of the ceiling: if the ceiling is 60 seconds, target 40–45 seconds of content plus a short end frame. That leaves room for a hook that lands in the first two seconds and a payoff that arrives before attention thins.

For training and internal content, the math flips. Learners tolerate and often prefer longer runtimes, but only when the video is chunked. A 12-minute module split into three 4-minute videos performs better in completion data than one 12-minute block, because each chunk gives the viewer a natural stopping point and a visible sense of progress.

Aspect ratio changes perceived pace

Vertical video feels faster. Cuts appear more abrupt, narration sounds more clipped, and viewers expect a new visual idea every few seconds. A script that reads comfortably in a 16:9 explainer can feel rushed in 9:16. When you convert a horizontal project to vertical, plan to shorten rather than letterbox — and plan to shorten the sentences, not just the scenes.

Subtitles are a duration decision

If your target surface is watched with sound off, which is common on social feeds, then captions carry the script. Long compound sentences become unreadable in the two-line caption window. Shorter sentences reduce runtime and improve legibility at the same time, which is one of the rare cases where two goals align perfectly.

Building a Duration Budget Before You Generate

The most reliable way to hit a runtime is to plan it arithmetically before you open the editor. This takes about five minutes and saves multiple re-renders.

Script math that actually holds up

Natural narration lands somewhere between 130 and 160 words per minute depending on language, voice, and content density. Use 150 words per minute as your baseline, which is 2.5 words per second.

30 seconds ≈ 75 words
45 seconds ≈ 110 words
60 seconds ≈ 150 words
3 minutes ≈ 450 words
5 minutes ≈ 750 words

These are content words only. Then subtract a pause allowance: roughly 1.5 to 3 seconds per scene transition, plus any deliberate beat you add after a key point. A ten-scene 60-second video with one-second transitions realistically supports closer to 130 words, not 150.

A worked example

Suppose you need a 45-second product teaser for a vertical feed.

Word budget: 45 × 2.5 = 112 words of narration.
Pause allowance: 6 scenes with short transitions ≈ 6 seconds.
Opening and closing frames: 2 seconds.
Adjusted narration budget: 45 − 6 − 2 = 37 seconds of speech, or about 92 words.

Suddenly the script brief is specific: 92 words, 6 scenes, no sentence longer than about 14 words. That is a far better starting point than "write something punchy" and then discovering the export runs 58 seconds.

The two-second rule for hooks

Whatever your runtime, the first sentence should be short enough to complete inside two seconds — roughly five words. Anything longer and the viewer has not yet learned what the video is about when the scroll impulse arrives.

Choosing Models When Exact Runtimes Matter

Synthesia handles the avatar-and-narration portion of a pipeline well, but most teams also generate supporting footage: b-roll, stylized inserts, animated backgrounds, or transitions created by other generative video models. Those models have their own duration behavior, and it is worth planning around it.

Long-context models and the short-clip mismatch

Large video generation models are often optimized for extended, cinematic sequences. That is excellent when you want a sweeping establishing shot, and awkward when you need a precise four-second insert. Text-to-video models commonly produce clips in fixed increments — often four, five, or ten seconds — and may not respect an arbitrary request like "exactly 3.4 seconds."

Practical approach: generate at the model's native increment, then trim to the exact length you need in the assembly step. Always generate slightly longer than required, never shorter. A five-second clip trimmed to four is easy; a four-second clip that needs to become five is a re-generation.

Models optimized for predictable output

Some models are tuned for speed and consistency rather than maximum fidelity. They tend to produce shorter clips with fewer artifacts and more predictable framing, which makes them useful for repetitive inserts like lower-third textures, product rotation loops, or background motion. When your runtime budget is tight, choose the model that reliably gives you the duration you asked for, even if the image quality is a notch lower. A slightly softer four-second clip is better than a beautiful six-second clip you have to cut in half and lose the composition.

Video-to-video, references, and length stability

When you supply a reference clip or a first frame, duration behavior often becomes more stable because the model is extending or transforming something with a known length. This is the most reliable way to control runtime for generated footage: produce a simple reference with the exact timing you want, then let the model restyle it. The timing survives; the visuals change.

A practical division of labor looks like this:

Narration and on-screen presenter: script-driven avatar tools, where runtime is fully deterministic from the text.
Precise-length inserts: video-to-video or reference-guided generation, where you control timing before the model runs.
Atmospheric background footage: free-running text-to-video at native increments, trimmed during assembly.

Mixing these deliberately removes most of the runtime uncertainty from your pipeline.

Pre-Generation Scripting and Pacing Analysis

Before rendering anything, run a pacing pass on the script. This is the cheapest quality gate in the whole workflow.

Read it aloud, timed

Read the script at your intended delivery pace with a stopwatch. Do not read it at presentation speed — read it the way the avatar will, which is even and unhurried. If the read lands more than five percent above your target, cut content now. Cutting text before generation is free; cutting it after means a re-render.

Mark the pause points explicitly

Add markers where the avatar should breathe: after a statistic, before a call to action, between a problem statement and its solution. Pauses serve two purposes — they improve comprehension and they let you fine-tune the runtime in small increments later without rewriting anything.

Verify sentence length distribution

Scan for any sentence over 20 words. In generated narration these are the sentences that sound rushed, and they are also the ones most likely to be read faster than your estimate, throwing off the runtime math. Break them.

Check scene-to-word ratio

If a scene carries only one short sentence, you are paying a transition cost for very little content. If a scene carries eight sentences, you have created a block that cannot be adjusted without a heavy re-render. Aim for a balance where most scenes carry three to five sentences.

Preview the voice before committing

A voice preview costs almost nothing and reveals whether the selected voice runs fast or slow. Some voices read noticeably quicker than others; a voice that runs 8 percent fast can shave five seconds off a minute-long video, which is enough to change your ending.

Queue Discipline: Rendering in Time-Bound Batches

Runtime planning interacts with render operations more than most people expect, especially when a single project contains multiple language versions or variants.

Sort renders by runtime, longest first. Long renders dominate the total processing window. Starting the five-minute video last means the whole batch waits on it. Starting it first hides the long tail behind the short renders.

Group by runtime tier, not by project. Keeping a 15-second tier, a 60-second tier, and a multi-minute tier lets you predict throughput and spot anomalies — a 40-second clip in the 15-second tier is a signal that something went wrong upstream.

Keep a re-render buffer in your schedule. If you plan to publish three videos and each needs one correction, budget for six renders' worth of time. Teams that assume a perfect first pass end up publishing late, which is worse than publishing a slightly imperfect video.

Version naming that encodes runtime. Name exports with the surface and duration, for example product-teaser-vertical-45s. When a platform's limit changes, you can find and update affected assets immediately instead of opening every file.

Post-Generation Trimming, Padding, and Re-Rendering

Once you have an export, you have three levers. Use them in this order.

1. Trim silence and transitions first

Before cutting content, cut overhead. Tighten the lead-in before narration begins, reduce a transition's duration, and remove any trailing frames after the last word. This often recovers two to five seconds without touching a single sentence.

2. Trim content from the middle, never the end

If you still need to cut, remove a supporting example rather than the closing call to action. Viewers who reach the end are your most valuable audience segment, and the ending is what they act on. A common failure is trimming the last sentence to hit a time target — that turns a working video into a dead end.

3. Pad deliberately, not accidentally

If the video comes in short and needs to reach a specific ad length, add a value-carrying frame rather than dead air: a summary line, a logo with a URL, a single statistic. Three seconds of a useful end card reads as intentional; three seconds of frozen avatar reads as a rendering error.

Re-render decision criteria

Ask three questions before hitting re-render:

Does the error change meaning? A mispronounced product name: re-render. A slightly late pause: it depends.
Is the delta larger than two seconds? Smaller deltas are almost always cheaper to fix in assembly.
Will this asset be reused? A template used for fifty videos justifies a clean re-render. A one-off does not.

Common Mistakes That Inflate or Shrink Runtimes

Counting words and forgetting pauses. The single largest source of runtime overruns. Always subtract transition and beat allowances from your word budget.

Assuming the timeline length equals the export length. Encoding, hold frames, and end cards can add time. Verify with the actual file, not the editor's readout.

Re-rendering to fix two seconds. Pull the export into an editor and trim. Save the re-render for content changes.

Using one master for every surface. A 90-second horizontal explainer pushed into a vertical feed will be clipped or replaced by a preview. Produce a genuinely shorter vertical cut, not a cropped version of the long one.

Ignoring how captions change pacing. Captions force sentence breaks. Write for them from the start rather than retrofitting them.

Letting the video run to the platform maximum. "Allowed up to 90 seconds" is a technical fact. "Rewards 40 seconds" is the one that affects performance.

Skipping the timed read. Nearly every runtime surprise traces back to a script that was never read aloud with a stopwatch.

Over-segmenting scenes. Twenty micro-scenes multiply transition overhead. Fewer, richer scenes are easier to control and usually look more coherent.

QA Checklist and FAQ

Pre-render checklist

Target runtime defined per surface, not per project.
Word count within budget after subtracting pause allowances.
Longest sentence under 20 words.
Every scene carries 3–5 sentences.
Voice preview timed against your estimate.
Transition and end-frame overhead counted.
Re-render buffer scheduled.

Post-render checklist

Measured runtime from the exported file, not the editor.
Hook completes within the first two seconds.
Call to action lands before the final frame.
Captions legible with sound off.
Aspect ratio matches the destination surface.
Export filename encodes surface and duration.

Frequently asked questions

Can I set an exact runtime in Synthesia? You control it indirectly through script length, scene count, and pause settings. An exact target is achievable, but you hit it through the narration and structure rather than a single duration field.

How many words fit in 60 seconds? Roughly 130–150 words of narration, minus allowances for transitions, opening frames, and any deliberate pauses. Plan for 130 to be safe.

What if my video must be exactly 30 seconds for an ad placement? Write to 65–70 words, keep pauses minimal, and use the assembly step for the final adjustment. Ad placements are usually strict about duration, so build in a small margin and trim rather than pad.

Do longer AI videos perform worse? Not inherently. Long runtimes fail when the content density is low — the viewer senses padding. A dense ten-minute training video can hold attention; a thin 90-second promo cannot.

Should I generate one long video and cut it into shorts? It works only if you write the long version with cut points in mind. Record the long form in clearly separable sections, each with its own hook, so a vertical cut stands alone. Chopping a continuous narrative usually produces fragments that make no sense out of context.

How do I keep runtime consistent across languages? Dubbed or regenerated narration runs at different speeds by language. Recalculate the word budget per language using that language's words-per-minute rate, and accept that some versions will need light re-writes to land in the same window.

What is the best default target for a vertical feed? Around 30–45 seconds. It is long enough to deliver one complete idea and short enough to survive an impatient scroll without losing the ending.

How often should I re-verify platform limits? Every few months, plus any time a format suddenly underperforms. Limits drift quietly, and a video that used to fit a feed can start getting truncated without any warning in your dashboard.