Make an SRT from a video, in the browser
Drop an MP4, MOV or WebM in, and TapeText pulls out the audio, transcribes it with word-level timings and writes a subtitle file shaped to broadcast conventions: two lines, 42 characters, never longer than seven seconds. Export SRT or VTT. There is no watermark, no minute allowance and no account.
Subtitle generators are the clearest example of a free tier that is not one. Three videos a month capped at ninety seconds. Ten minutes of captioning, then a watermark burned into the export. Thirty minutes a month at 720p unless you pay eighteen dollars.
Those limits are not greed so much as arithmetic: every one of those services runs the speech model on their own hardware, and every minute you transcribe costs them money. The limit is where the cost stops being bearable.
Running the model in your browser moves that cost to a machine you already own, which is why there is no cap here to design. It is also why the output needs your eye: see the section below on what a generated subtitle file still gets wrong.
How it goes
- 1
Open the video file
MP4, MOV, WebM and MKV are read directly. Only the audio track is decoded, and the picture is never touched, so a 4K file is no slower than a 720p one.
- 2
Read it through once
Proper nouns and numbers are where speech models fail, and both tend to matter. Click a word to hear it.
- 3
Export SRT or VTT
Cues are built from the word timings, not by dividing a sentence evenly, so they stay in sync inside a long line. The file keeps your video's name, which is what makes players pick it up automatically.
What makes a subtitle file good, and what this one does about it
A subtitle file is not a transcript with timestamps. It is text shaped for someone reading at a glance while also watching, and the conventions are old, specific, and mostly ignored by automatic generators.
Line length and count. Two lines maximum, forty-two characters each. Past that a viewer stops watching the picture and starts reading, which is the failure the whole format exists to avoid. Where a cue has to be broken into two lines, the break is placed at the word boundary nearest the middle rather than greedily at the end, because a full line sitting over a two-word orphan is harder to read than two balanced ones.
Duration. Nothing shorter than one second, because a cue that flashes is a cue nobody reads; nothing longer than seven, because by then the picture has moved on. A short cue here is extended to reach a readable minimum, but never past the start of the next one, because overlapping cues make a player show both at once, which is the most common defect in machine-made subtitle files.
Timing inside a sentence. This is where word-level timestamps earn their place. A tool that only knows when a *sentence* started and ended has to divide it evenly between its cues, so a sentence with a pause in the middle drifts out of sync with the speaker and snaps back at the end. Each cue here is dated by the first and last word actually in it.
What it still does not do: it does not place cues around shot changes, it does not burn subtitles into the picture, and it does not know that the thing on screen is a name it should be capitalising. The first is an editing decision, the second is out of scope, and the third is why you read it through.
Questions
- Is there a watermark or a length limit?
- Neither. The model runs on your machine, so there is no per-minute cost to recover and nothing to cap. Videos of any length work; long ones simply take longer.
- Does it burn the subtitles into the video?
- No. It produces a subtitle file, SRT or VTT, that you add in your editor or alongside the video. Burning text into the picture is a video encoding job and is not something this tool does.
- Why is my exported file named after my video?
- Because that is what makes it work. Players and editors match a subtitle file to a video by filename, so `interview-01.mov` gets `interview-01.srt` and is picked up automatically.
- Can I use the SRT on YouTube?
- Yes. SRT is the format YouTube's subtitle upload expects, and the cue shaping here follows the same conventions its own editor enforces.