Why we built a subtitle tool that shows you its mistakes

Earlier this summer I shipped TimedSubs, a tool that turns finished scripts into subtitle files. SubtitleGenerator started somewhere different: in the comment sections of the biggest AI subtitle tool's users.

I'm not a developer. I come from operations and content growth. My version of market research has always been the same: go where real users complain, and read everything they write.

For weeks I read public complaints about the category leader. Not the marketing pages, not the comparison blogs — the places where users describe what actually made them angry. A few themes kept showing up:

• Users couldn't trust the timeline. Clicking a timestamp to jump to that moment didn't behave the way they expected.
• Adding the very first subtitle to a clip was weirdly hard.
• Missed words. When the AI transcription dropped a word or got one wrong, nobody helped you find it. You scrubbed through hundreds of lines yourself.

Then I ran the same test videos through their product and ours side by side. What I found wasn't a secret: every tool in this category gets the transcription mostly right. And that "mostly" is the problem nobody was solving.

Every subtitle tool is built on the same bet: that the first draft will be good enough. The transcript comes back, maybe you tweak a style, maybe you fix the one wrong word you happened to notice — but the cost of fixing mistakes was always yours to pay. If the AI got 2% of your words wrong, finding that 2% was a manual job. A hundred-line transcript is a haystack, and every tool told you it was all hay.

So the design question we kept coming back to was: what if the tool stopped pretending the draft was right, and instead showed you exactly where to look?

That became the Fix experience — the core of SubtitleGenerator:

• Words the transcription model is unsure about are marked, not hidden.
• A visible review count tells you how many are left: 3… 2… 1…
• "All clear." A finish line. You know when you're done, because the tool says so.

Try the Fix loop on the free 20-second sample — no signup needed.

SubtitleGenerator is the full workflow around that core: drop in a video — no signup — and generate timed subtitles. Fix them with the review loop. Translate the whole track in place into 109 languages, powered by the same script-locked alignment engine behind TimedSubs. Then style, animate and export. The original video never leaves your browser.

But the build itself was a series of lessons. Three of them hurt.

Lesson one: your competitor research is a lie waiting for a ship date

In my research notes I had written down the category leader's weaknesses with total confidence. One note said they locked custom fonts behind their paid plan and didn't even have a fonts page. I nearly turned that note into marketing copy. Before I did, we re-tested.

They had a fonts page. Fully built, well structured. My note was from an earlier snapshot — they'd shipped it since.

The lesson: eighteen data points is not a conclusion. Competitors fix things. Any claim you write down about them expires the moment they deploy. If you're going to compare, compare on the same material, on the same day, with screenshots — and date your checks.

Lesson two: the gap between "it runs" and "it runs with a real payment network" is where the money leaks

We switched payment providers mid-build. Not because we wanted to. The one we'd chosen became the center of a story that made the rounds in indie circles this year: a fellow founder's account permanently closed and his balance frozen for 90 days, over a compliance escalation that had nothing to do with fraud. The platform was protecting its card-network relationships, and his revenue went to zero overnight.

That made us look hard at our own single point of failure. Every dollar of future revenue would sit behind one merchant-of-record decision we couldn't influence. So we switched before launch — and the new provider nearly broke us in ways no local test could catch.

Their webhooks carry a timestamp, and we validated signatures with a five-minute tolerance window. Reasonable, right? Except their retry system replays delayed notifications carrying the original timestamp. A payment that succeeded after a slow retry looked like a forgery and got rejected. We now accept 45 minutes of drift.

Then PAYG. A $3 top-up checkout generated a hosted URL with an expiry. Our code cached that URL and guessed the expiry locally. Users who came back a few hours later clicked a dead link. The provider knew the real expiry the whole time — we had just never asked. Now the server trusts the provider's authoritative expiry, and expired intents get a fresh session.

None of this showed up in tests. It showed up when a real payment network behaved the way real payment networks behave.

Lesson three: every privacy promise has a hidden engineering bill

Our promise is that the original video stays in the browser — no upload, no server copy; only the extracted audio is sent for transcription and deleted after processing. That privacy line is a feature users love. It's also an engineering bill nobody warns you about.

Real videos arrive as MKV, WebM, MOV — and the audio inside can be AAC, Opus, or something even stranger. A browser can play almost anything. But to process, remux and export that audio, we had to handle formats the browser barely documents: Opus transcoding, AAC remuxing, sample-rate edges that only appear with specific cameras and screen recorders.

Every "we never touch your data" claim hides a pile of format edge cases. We paid that bill because the promise matters. But we now read every privacy claim — including our own — with more respect and more skepticism.

What came out of it

SubtitleGenerator is live. Free to start, no signup. Pro and Max add whole-track translation, HD exports without a watermark, eight subtitle formats, custom fonts and saved brand styles.

The Fix loop came directly from those user complaints: when an AI transcript comes back 98% right, the product's job should be showing you exactly where the 2% is. Everything else — styles, translation, export — is depth around that promise.

Open the 20-second sample and watch the review count reach "All clear." If you'd rather start with your own footage, upload a video — it stays in your browser.