How to judge a free subtitle generator before you upload anything

You can judge a free subtitle generator in about twenty minutes without reading a single review. Make one two-minute clip that contains the things subtitles usually get wrong, run it through each candidate, and compare the results on a fixed list: the file you get out, how the text is split into cues, whether the timing holds, the caps, what the vendor says happens to your upload, and whether the free tier marks the output. The clip does the arguing for you.

We do not publish a ranking of these tools. We make one of them, and ranking our own category would be marketing, not a comparison.

The clip that finds the cracks

Thirty seconds of one person reading cleanly separates nothing. The differences show up on the awkward parts, so build a clip that is nothing but awkward parts. Two minutes is enough, and short enough that you will not mind running it five times.

Put four things in it:

• A proper name, ideally one that is not in the top hundred. Names are where speech models quietly substitute a word that sounds close.
• A number said out loud, such as a date, a price or a phone number. Tools differ on whether you get “fourteen ninety nine” or “14.99”, and that decision is yours to live with in every cue.
• Two people talking, including one quick handover where the second person starts before the first has finished. Cue splitting has to cope with a speaker change.
• Five to ten seconds of silence or room tone, ideally in the middle. Silence is where some models emit a line of text that nobody said, and where others leave a gap that throws the timing.

Use material you already own. If a candidate works by uploading, your test file goes with it, so keep client recordings and anything covered by an agreement out of this. If you have nothing suitable, record yourself and a colleague reading two minutes of anything.

Keep one longer file in reserve, thirty minutes or more, for the tools that survive the short clip. It exists for one check, and that check matters most.

The questions to ask a free subtitle generator

Each item below is a question you put to every candidate, with an answer you can find in about two minutes. Run the clip through all of them first, then fill the table in one sitting, so you are comparing evidence rather than memory.

Cue quality matters more than word accuracy

Transcripts are judged on words. Subtitles are judged on what fits on a screen for the length of time it is there, which is a different problem. A tool can get every word right and still hand you something unusable: a hundred and eighty characters sitting in one cue for twenty-five seconds, with the first line long gone by the time the speaker reaches the point.

A readable cue has a fairly narrow shape. Around forty-two characters a line, two lines at most, each cue on screen for roughly one to six seconds, a small gap between cues, and breaks that fall at the end of a clause rather than in the middle of a noun phrase. Those are the values we build to, and how to make an SRT file has the reasoning behind each number.

Weight this heavily, because of what the repair costs. A wrong name is a find and replace, thirty seconds of work. Badly split cues mean opening a subtitle editor and re-timing by hand, which on a thirty-minute video is most of an afternoon. A tool that is slightly worse on words and much better on cues is the better subtitle tool.

Counting cues is the fastest proxy. At one to six seconds each, two minutes of continuous speech produces somewhere between twenty and a hundred and twenty cues. If your export contains six, the tool is writing paragraphs and calling them subtitles.

Drift shows up at the end, not the beginning

Your two-minute clip cannot tell you whether timing drifts, because two minutes is not long enough for it to become visible. That is why the finalists get the reserve file.

Two different faults hide here and they need different fixes. A constant offset means every cue is late by the same amount, and one adjustment in any subtitle editor corrects the entire file in a few seconds. Drift means the error grows as the file goes on: the first minute is perfect, the last minute is four seconds out, and no single offset fixes it. Drift usually means the tool processed the audio in chunks and lost track of where each chunk sat in the original recording.

The check itself takes a minute. Load the video into any player that accepts an external subtitle file, add the subtitles, skip to five minutes before the end, and watch. If the cues land there, they land everywhere.

Caps, accounts and what the vendor says happens to your file

Two numbers decide whether a free tier is usable for the work you have: minutes per file and minutes per month. A third catches people out, a limit on the number of files for the lifetime of the account. Read all three off the pricing page, and note whether each resets or is a one-time allowance. There is a structural reason caps exist wherever processing happens on someone else’s server, and the arithmetic behind the thirty-minute cap sets it out.

For the upload question, the only honest source is the vendor’s own privacy policy and terms. Look for how long files are kept, whether they are used to train models, and who else gets to process them. Be clear about what this part of the exercise is: you are reading a promise, not verifying behaviour. The only way to stop relying on a promise is to not send the file anywhere.

Marks on the output are the last thing to check

Check the export, never the preview. A preview can look clean while the file you download carries something extra. Three things to look for: a watermark burned into any video the tool produces, a credit line added as an extra cue at the end of the subtitle file, and a quality or length restriction that applies only to the download. Reading the final cue in a text editor catches the second one in seconds, and it is the one people miss until a client points at it.

Where our own tool lands on this list

Fair is fair, here are our answers to the same eight questions. FreeTranscribe exports SRT, VTT and plain text, and does not produce burned-in video at all, so if you need captions baked into the picture you will need a separate step in a video editor. Cues are built to forty-two characters, two lines, one to six seconds, and an eighty millisecond gap. There is no cap per file or per month, no account, and no watermark or credit cue. Nothing is uploaded, because the model runs on your own graphics card inside the browser tab, which you can confirm by watching the network panel stay quiet while it works.

The limits worth knowing before you bother testing it: desktop Chrome or Edge with a working WebGPU adapter, English only, and a roughly 200 MB model download the first time you use it. It runs the base size of OpenAI’s open-source Whisper model, which is solid on clear speech, with unusual names, strong accents and noisy rooms as its weak spots. Speed was about 1.5x real time on a desktop with a graphics card in our test, and a thin laptop takes roughly as long again as the recording. Run your own clip through it rather than taking our word for it. Try it with your two-minute clip.

Frequently asked questions

How long should the test clip be? Two minutes, with the hard parts packed in. Longer wastes your time on the first pass, and shorter will not contain enough awkward material to separate the candidates. The one thing two minutes cannot show you is timing drift, which is why the finalists get a thirty-minute file.

Should I measure word accuracy properly? Only if you have a verified transcript to compare against, which most people do not. Reading the five cues around the name and the number you planted tells you most of what a word error rate would, and it takes a minute instead of an hour.

What if a tool only offers burned-in captions? Treat it as a real limitation rather than an automatic dealbreaker. Burned-in text cannot be corrected later, switched off by the viewer, or reused as a transcript. If the tool also exports a file, use the file.

Is a bigger cap always better? No, because the cap interacts with the account requirement. A generous monthly allowance behind a signup and a file-count limit can be worth less in practice than a smaller per-file cap you can use immediately and repeatedly. Compare what you can get done in the next hour, not the headline number.

Sources, checked 15 September 2026

• Our own cue builder in the transcriber’s exporter: forty-two characters per line, two lines, one to six seconds, eighty millisecond gap.
• Our own tool’s supported browsers, language, model size, export formats and measured speed, as stated elsewhere on this site.
• No third-party sources were needed. This post describes a method and makes no claim about any other product.