Transcribe audio to text

Why this one not uploading matters more than usual

Every other tool on this site processes a file. This one processes a recording of somebody talking: an interview, a lecture, a consultation, a meeting, a voice note. The people in it did not choose your transcription service.

Hosted transcription means a copy of that recording on somebody's server, subject to their retention policy, their staff, and their breaches. Here the audio is decoded in the page and handed to a model running in the same tab. Open the Network tab while it works: the model comes down, and nothing goes up.

Which model to use

Tiny is quick and fine for clear speech, one speaker, no background. Base is the default and handles accents and a noisy room. Small is noticeably better on difficult audio and is three times the download.

The first run of a model downloads it; after that the browser keeps it and the tool works offline. Transcription itself runs at roughly real time on a laptop, an hour of audio takes about an hour, because it is your processor doing it rather than a rented one.

Subtitles are not just a transcript with times on

They are constrained by how fast people read, and the constraints are why some subtitle files are comfortable and others are exhausting.

A line runs to about 42 characters and a cue to two lines, because more than that cannot be taken in at a glance and on a phone it covers the picture. A cue stays up for at least a second, anything shorter flashes, and no more than about seven. And the real limit is reading speed: past roughly seventeen characters a second, most people cannot finish the line before it goes.

The cues here are split at sentence ends where there are any, merged when they are too short to read, and wrapped on words. Any cue still faster than a person can read is marked rather than silently left, because the fix is usually to cut a word rather than to change the timing.

SRT or WebVTT

SRT is the one nearly everything accepts: video players, editing software, social platforms. WebVTT is the web standard, is what an HTML video element wants, and can carry styling and positioning that SRT cannot.

The only difference in the timestamps is that SRT writes a comma before the milliseconds and WebVTT writes a full stop. That single character is the commonest reason a subtitle file is rejected as invalid.

When the subtitles are out of sync

Usually because the audio was trimmed after transcription, or the video has a few seconds of countdown at the front. Shift moves every cue by the same amount, which fixes a constant offset. It cannot fix drift: subtitles that start right and gradually slide are a frame-rate mismatch, and that needs the file re-timed rather than shifted.

What it will not do

It will not reliably tell speakers apart, that is a different model and a harder problem. It will not punctuate perfectly, and it invents punctuation confidently where the speech was ambiguous. It struggles with heavy crosstalk, poor microphones and specialist vocabulary. And it is not a legal record: for anything that matters, read the transcript against the audio.