Convert a video to text

A transcript is not a caption file

Both come from the same speech recognition, and that is where the similarity ends. A caption file is cut into short timed fragments sized to be read off a screen while something else is happening. A transcript is meant to be read on its own, so it wants sentences and paragraphs.

Handed an .srt when you wanted a transcript, you have a document broken every few words with a number and two timecodes between each piece. It is the right file for a video player and close to unreadable as prose.

So the output here is grouped into paragraphs with a timestamp on each, capped in length so a fluent speaker does not produce one unbroken wall. Markdown or plain text, and the same transcription can be exported as .srt as well if you later want both.

How it works, and what it costs you

Your browser decodes the video's audio with its own decoder through WebCodecs, resamples it to 16kHz mono, and runs a Whisper model in a Web Worker. Nothing is sent anywhere, because there is no server side to this product and no /api directory in it.

Transcription measured 8.6x faster than realtime on a 31-minute recorded lecture, Chrome, 16GB Windows laptop, using the WebGPU backend. Without WebGPU the same work falls back to WebAssembly on the CPU, which is much slower and has not been measured, so treat the figure as the good case rather than a promise.

The real cost is the first run. The model is a 45MB or 80MB download from Hugging Face before anything can be transcribed, after which your browser caches it and later runs start immediately. That download is the only network traffic involved, and it goes to you.

What it gets wrong, in advance

Accuracy is decent, not perfect, and the failures are specific rather than random. Knowing them in advance is the difference between a useful document and a confidently wrong one.

• Proper nouns and names, which come out as a plausible common word spelled with total confidence
• Acronyms and technical vocabulary, which become the nearest everyday phrase
• Numbers, which are transcribed by sound and can land an order of magnitude out
• Overlapping speech, which produces one voice or an average of two that neither said
• Long stretches with no speech, where the decoder can get stuck repeating a phrase until the quiet ends
• Speakers, which are not labelled at all. There is one continuous transcript, not a dialogue

Before you paste it into an assistant

The most common next step is handing the transcript to an AI to summarise, and a transcript pasted in raw is treated as a reliable document. A misheard name becomes a confident claim about a person and a mistranscribed figure becomes a number in your summary.

The copy-for-an-assistant export puts a short preamble in front of the text saying it was machine-transcribed and naming the places such models go wrong. That makes the model hedge the right parts rather than everything, and it is the whole of why that export exists as a separate button.

Worth noticing what that step costs, though. Transcribing here uploads nothing, and pasting the result into a hosted assistant uploads the text. If the video was one you could not upload, the transcript is usually one you cannot paste either.