Synthesia: Source Becomes Video, Accuracy Still Needs an Owner
Synthesia is easy to describe as an AI avatar video tool. That description became too narrow once I used it.
In this Product Value Evaluation, I tested whether a first-time user could take existing written business information and turn it into a short Japanese instructional video without becoming a scriptwriter, presenter, instructional designer, or video editor.
This evaluation was conducted on September 1, 2026, in a paid Starter account configured for Japan, using Chrome from Brazil.
Synthesia did transform the source into a finished video.
The more difficult question was how much of the judgment disappeared with the production work.
Key Findings
- The source-to-video transformation was real. Synthesia turned written support information into a structured Japanese video with scenes, narration, visuals, an avatar, and a finished render.
- The instructional result was not independently reliable. Source retrieval failed, required information was omitted, unsupported information was added, and later automated corrections created new problems.
- AI Dubbing produced the strongest finished result in the evaluation. A 1:23 English video became a fluent Japanese lip-synced version that I would have been comfortable using without another translation or localization review in this test case.
- Synthesia took on more interpretation than HeyGen did in the earlier PVE. That increased the amount of work the product could remove, but also increased the number of decisions that still needed human review.
The Task
The test used a short section of Zoom’s public support material explaining how to schedule a first meeting.
The goal was not to hand Synthesia a finished Japanese script and ask it to animate it. I wanted to see whether the product could take existing information and do some of the work normally sitting between source material and finished training content.
That included deciding how to structure the information, dividing it into scenes, writing Japanese narration, selecting visuals, assigning a presenter and voice, and assembling the video.
Once the exact source text was available, Synthesia did all of that.
This is meaningful product value. The user did not need to film anyone, record narration, construct scenes manually, or move between separate writing, design, voice, and editing tools.
But the first important failure happened before that value could be trusted.
When the Source Failed, Generation Continued
The original Zoom URL returned Page Not Found inside Synthesia.
The product continued anyway.
It generated an English outline containing information that was not reliably grounded in the designated source.
That is a more consequential failure than a broken import button. The system did not simply fail to retrieve the source. It moved directly from failed retrieval to plausible content generation.
Supplying the relevant Zoom text directly allowed the experiment to continue, but source fidelity still required review. The Japanese outline omitted the final Save action and introduced unsupported information.
Later, a broad automated correction created additional problems, including a voice change and another unsupported instruction.
Narrow corrections worked better. When I already knew exactly what was wrong and specified the element that needed to change, Synthesia could execute the correction.
That produced an important distinction:
The product was better at applying a known correction than at discovering that the correction was necessary.
The production work had moved into the product. Editorial responsibility had not.
A Finished Video Was Not the Same as a Finished Instructional Result
The final Zoom video contained all six required source actions and played as a complete 34–35 second video.
But some issues only became apparent after final generation.
The pacing was somewhat breathless between steps, and one section began narrating Step 3 after the visual had already moved ahead.
For entertainment video, that might be a minor timing flaw. In instructional content, visual and verbal sequencing are part of the information.
This matters because the draft preview did not expose the complete final speaking-avatar performance. The final render consumed credits and took more than seven minutes in this run.
Synthesia therefore compressed the production workflow substantially, but the review loop remained real and some of it happened late.
AI Dubbing Was Different
The most polished automated result in this evaluation came from a capability outside the main P01 workflow.
I used a public 1 minute 23 second English-language video and asked Synthesia to dub it into Japanese with lip sync and adaptive duration.
Synthesia quoted 329 credits before generation, about 27% of the Starter plan’s 1,200-credit monthly allowance. The result took approximately 24 minutes to generate.
It was remarkably good.
The Japanese was fluent, the lip sync was visible, and the voice remained close enough to the original speaker that the integrated impression was not of an ordinary dubbed track. It looked and sounded more like the speaker herself had suddenly become fluent in Japanese.
Get IC Eight’s stories in your inbox
Join Medium for free to get updates from this writer.
Some source-language texture remained. One expression, そこで、皆さんの出番です (“This is where you come in”), was understandable but sounded more like English discourse logic carried into Japanese than something a Japanese speaker would naturally originate in that context.
I would not necessarily classify that as a dubbing defect.
There is a difference between translating what a speaker said and rewriting the performance as though the thought itself had originated in Japanese. The latter is adaptation.
For this particular video, I would have been comfortable using Synthesia’s result without an additional human translation or localization pass.
That is a substantially stronger level of automation than the instructional-video workflow reached.
Fluent Japanese Did Not Erase the Speaker
The dubbing result also produced an unexpected perceptual effect.
The speaker’s mouth had been synchronized to Japanese, but her facial articulation and body language remained recognizably her own.
The result did not look like a Japanese performer.
It looked like a foreign or bicultural person who happened to speak extremely fluent Japanese.
That coherence matters. Localization does not necessarily improve when every visible behavior is pushed toward the target culture. Replacing a recognizable person’s movement with stereotypically Japanese nonverbal behavior could create an entirely different kind of mismatch.
For dubbing, preserving the person while changing the language may be more important than trying to localize the person.
Compared with HeyGen: More Transformation, More to Verify
The earlier HeyGen PVE started from a finished Japanese script.
Its first-use value was relatively compact: choose a presenter and voice, provide the script, and generate the video. The production path was easy to understand because relatively little interpretation was being delegated.
Synthesia took on more.
In this evaluation, it was asked to move from existing information toward the structure of the video itself. That potentially removes more work from the user, but every additional interpretive step creates another place where source meaning, instructional logic, language, layout, or timing can drift.
The comparison is therefore not that one product automated more successfully than the other.
It is that they exposed different boundaries.
HeyGen made the presenter-production layer easy to remove. Synthesia showed how much further source-to-video transformation can go, and why review becomes more important as the product takes on more of the thinking between source and output.
AI Dubbing was the exception. There, Synthesia came much closer to making the transformation itself feel finished.
One Capability I Still Want to Test
Another Synthesia capability did not appear in the Starter self-serve account I tested: Roleplay Sessions.
It is part of Synthesia’s broader platform, but its public feature page used a sales-led path, and Japanese was not among the listed languages.
I would very much like that to change.
Japanese Roleplay would be an unusually interesting Product Value Evaluation because the problem would not end with producing fluent Japanese.
A simulated sales conversation, management conversation, or difficult feedback session also has to get the interaction right.
The sequence of acknowledgment and disagreement matters. Hierarchy matters. Indirectness matters. The way an objection is introduced and answered matters. The expectations attached to the roles matter. And if the system evaluates the learner, the scoring criteria themselves need to make sense in that conversational environment.
So the question would not simply be:
Can the avatar speak Japanese?
It would be:
Does the conversation behave like a plausible Japanese professional interaction?
Synthesia, give me Japanese Roleplay and I will happily find out.
The Product Value
Synthesia clearly reduced the amount of specialist production work required to create business video.
What it did not consistently remove was the judgment required to decide whether the resulting video was accurate, well structured, appropriately localized, and ready for its intended use.
That boundary differed by capability.
For source-to-instructional-video creation, the strongest observed workflow was still:
AI transforms → human reviews → AI corrects → human verifies
For AI Dubbing, the human role became much smaller.
That difference may be more useful than asking whether Synthesia can generate video.
It can.
The more interesting question is which transformations still need an owner after the generation is done.
Related: Rask AI: The Translation Was Serviceable. The Dubbing Was Not. → — the same source video, dubbed by a different product.
Finding the gap is only half the work. The other half is deciding what to do about it.
Originally published at https://ic-eight.com.
