Este artículo aún no está traducido — estás leyendo el original en inglés.
Why I Put a Review Step Before the Dub
Users didn't ask for better AI voices first. They asked for control. Here is why I added multi-speaker voices, line-by-line review, a glossary and SRT export to AI Video Translator.
The updates came from my inbox, not a roadmap
When I launched AI Video Translator, I assumed the first complaints would be about voice quality. Does the dub sound natural? Does the accent hold up? That is what every AI dubbing demo competes on.
That is not what people wrote to me about. I run peters.lab as a one-person studio, so every support email lands in my own inbox. The emails about video translation kept circling one theme: control. People did not want a better black box. They wanted to see what the box was about to do, and fix it before it did it.
This October I shipped four changes that all come from that theme. Here is the story behind each one, and what building them taught me about localization as a solo builder.
One voice for everyone breaks the moment two people talk
The first version handled a single speaker well. Upload a tutorial, get it back in another language, in a voice that sounds like you. Then people started uploading interviews, podcasts and panel recordings.
In a two-person interview, a single dubbed voice is worse than a bad translation. The viewer can no longer tell who is asking and who is answering. The meaning of the conversation is in the back and forth, and one voice flattens it.
So the tool now detects each speaker and keeps their own voice in the translated version. The host still sounds like the host, the guest like the guest. I built this because users asked for it directly, and it was the clearest sign that people were using the tool for real conversations, not just demos.
Why the review step sits before the dub, not after
This is the change I am most opinionated about. You can now read every translated line, and edit it, before any audio is generated.
The order matters because of how video translation fails. In an earlier post I wrote that manga translation is constrained by space and video translation is constrained by time. A translated line has to fit the length of the original speech. When a line is wrong, fixing it in text costs seconds. Fixing it after dubbing means regenerating audio, re-checking timing, and listening to the whole segment again.
Most errors are not dramatic, either. A product name translated literally. A joke that lands as nonsense. A technical term rendered three different ways. A creator who knows their own content spots these in a minute of skimming text. The same creator finds them painful to hunt down in finished audio.
So the workflow is now: transcribe, translate, let the human read and correct, then dub. The AI does the heavy lifting, and the person who owns the video gets the last word before anything is locked in.
A glossary, because creators publish series, not single videos
The review step fixes one video. But creators rarely publish one video. They publish a course with twenty lessons, or a channel with a weekly show. If episode 3 calls your product one thing and episode 7 calls it something else, viewers notice.
The new glossary lets you fix how specific terms are translated: product names, character names, brand terms, jargon. You can import and export it as CSV, so the same list follows you across a whole series instead of being retyped each time.
This one was easy to justify, because I had already built the same idea for AI Manga Translator, where character names have to stay identical from the first chapter to the last. Consistency turned out to be the same problem in both products, just with a different container.
Subtitles as a separate file, burn-in optional
The smallest change, and maybe the most practical. Every dub now comes with a separate SRT subtitle file. Burning subtitles into the video is still available, but it is now a choice, not the default.
The reason is simple: platforms handle subtitles differently. Many creators want to upload captions as their own track so viewers can turn them on or off, and so the platform can index the text. Others want them baked in for short-form clips. A separate file serves both, and it means a subtitle typo can be fixed without re-exporting the whole video.
What I took away from this round
Looking back, none of these four features made the AI smarter. They all made the human more powerful. That is a different product philosophy from the one I started with, and I think it is the right one for localization.
When I wrote about why I built AI Video Translator, the pitch was a simple flow: upload, choose a language, get a translated video. That flow is still the core. What changed is that I now see the middle of it as the place where trust is earned. Fast output gets someone to try a tool. Being able to check and correct that output is what makes them use it on videos they actually care about.
It also confirmed something from my 6-month indie stack retrospective: running two translation products side by side is less work than it sounds, because so many lessons transfer. The glossary moved from manga to video. The review mindset will probably move back the other way.
If you translate or dub videos, you can try the new workflow on AI Video Translator. And as always, if something breaks or you want a feature, email me at hello@peterslab.co. As this round shows, I really do read those.