Skip to content
Rescreen
Signup
Menu

Rescreen / Free tools / Video to text

Video to text

A free video to text converter that transcribes locally. A speech model downloads to your browser once and runs on your own computer, so the recording and every word in it stay on your device. Edit the transcript, then download it as text or SRT subtitles. No account, no upload.

Transcribe a file

Drop a video here

or choose one from your device

MP4, MOV, WebM, MKV or an audio file, up to 3 hours. Stays on your device.

How to convert a video to text

From a recording to an edited transcript, without uploading anything.

  1. Choose a video

    Choose or drop a video or audio file. The page reads its length and checks it can decode the sound before anything is downloaded.

  2. Pick the language and model

    Leave the language on Detect automatically or choose it. The first time, the speech model downloads (81 MB to 150 MB); your browser keeps it for next time.

  3. Watch it transcribe

    Lines appear as each half-minute is done, with a progress bar and the time left. Cancel at any point; nothing is kept.

  4. Review and download

    Play the video from any line, correct names and terms in place, then download a .txt file, .srt subtitles, or copy the text.

What it does, and what it does not

It turns speech into text: one line per phrase, each with the time it was said, in the language spoken. It works on screen recordings, talks, interviews, lectures and voice notes, and on audio files as well as video.

It does not tell speakers apart, translate, summarise, or add the words on the screen. It writes what it hears, so a mumbled word, an unusual name or a product term can come out wrong, and it may tidy up “um”s and false starts that were really said. Treat the result as a first draft to read through, not a verified record.

Worked example

The sample clip is sixteen seconds of narration over a test picture, the way a short tutorial sounds: four sentences with pauses between them, one of them almost two seconds long. With the language onDetect automatically, the tool hears English and writes three lines, each stamped with the second it starts; the last two sentences come out as one line because there is no long pause between them.

Click 0:09 and the video jumps to “Give the project a short name”. Change “dashboard” to your product’s own word and the change is in the text file, the SRT subtitles and the copied text. Turn on timestamps to keep the times in the text file, which helps when you send someone notes on a recording.

The speech model, and what it downloads

The transcriber is OpenAI’s Whisper, an open speech model released under the MIT license, in the ONNX versions published by the onnx-community project on Hugging Face. It runs through transformers.js and ONNX Runtime Web (both Apache 2.0). Whisper reads sound in 30-second windows; the page cuts each window at the quietest moment near its end, so a sentence is not split mid-word.

  • Standard is Whisper base: 150 MB to run on a graphics card, 81 MB to run on the processor.
  • Small is Whisper tiny: 80 MB or 45 MB. It is quicker to download and run, and makes more mistakes, especially outside English.

The page uses the graphics card through WebGPU where the browser offers it with 16-bit support, and WebAssembly on the processor everywhere else. The model files come from huggingface.co; the browser stores them, so the next file starts at once and the tool keeps working offline. Clearing the site’s data removes them.

Languages offered by name: English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Swedish, Russian, Ukrainian, Turkish, Indonesian, Japanese, Korean, Chinese. English is by far the most accurate. Detection hears about a hundred languages, but these small models are often unusable outside the well-resourced ones, so check a minute of the result before relying on it.

How accurate and how fast it is

We measured it on one machine: a 2020 MacBook Air (Apple M1, 16 GB) in a Chromium-based browser, using narration recorded with the Mac’s built-in voices over a faint background hiss. Word error rate is the share of words wrong, missing or added.

TestModel, engineSpeedWord error rate
English, 89 s videoStandard, graphics card7× real time1.1%
English, 89 s videoStandard, processor4× real time2.2%
English, 89 s videoSmall, graphics card10× real time1.9%
English, 89 s videoSmall, processor6× real time1.1%
English, 15 min videoStandard, graphics card8× real time0.9%
Spanish, 20 sStandard, graphics card7× real time1.8%
German, 17 sStandard, graphics card6× real time6.3%
French, 16 sStandard, graphics card6× real time14.6%

Synthetic voices are clear and evenly paced, so these are best cases. Real recordings, with accents, cross-talk, music, a distant microphone or specialist words, will have more errors, sometimes many more. Memory stayed flat at under 1 GB for the browser through the 15-minute file, because the sound is read and transcribed half a minute at a time.

Which browsers and devices it works in

Tested in a Chromium-based browser (the engine behind Chrome and Edge) on an M1 Mac, on both the graphics card and the processor, at desktop and phone widths, with MP4, WebM and M4A files and a silent video. Firefox and Safari 26 provide the same building blocks but have not been tested yet. Phones and tablets have not been tested; the model needs several hundred megabytes of memory and may not fit next to other apps, so choose the Small model on a phone.

If the browser lacks something the page needs, it says so before you choose a file. If it cannot decode a file’s sound, it says that before downloading the model.

Limitations

  • Files up to 4.0 GB and 3 hours. The file is read a piece at a time, so its size does not decide memory use.
  • The first run downloads the model; on a slow connection that takes longer than the transcription.
  • No speaker labels, no translation, no word-by-word timing: times are per phrase.
  • Stretches with no sound loud enough to be speech are skipped. Speech buried under loud music may be missed or misheard.
  • If the recording has no speech at all, the page says so rather than inventing text.
  • Mistakes are certain, more so for names, numbers, jargon and accents. Read the transcript before you publish it.

Is my video uploaded?

No. The sound is decoded and transcribed inside this page and never sent anywhere; there is no server that could do it instead. The only things downloaded are the model files from Hugging Face and the page’s own code. We count anonymous page views and whether a transcription started, finished or failed, with which model and engine, and why it failed from a fixed list of reasons; those events never include the video, its name, its length, its language or any of its words. See the privacy policy.

Frequently asked questions

Made subtitles? Convert the SRT for a web player with the SRT to VTT converter. Cutting “um”s out of the video itself? See the filler-word guide.

Is it really free?

Yes. There is no account, no limit on how many files and no watermark. It costs us nothing to run, because your device does the work.

Does it work offline?

After the first run, yes: the page and the model are stored by your browser, so a later file transcribes without a connection as long as the page is still open or cached.

Can I get subtitles?

Yes. Download .srt subtitles saves one cue per line, with your edits. Most video players and YouTube accept SRT; a web page’s video player needs WebVTT, which the SRT to VTT converter makes.

Why does it skip the start of my video?

Lines only appear where there is speech. Silence, and sound too quiet to be speech, is skipped, so the first line starts when the talking does.

Why is the language wrong?

Detection listens to the first half-minute of speech. If that is music or a different language from the rest, choose the language yourself and press Transcribe again.

Spend less time editing
your next recording.

Create an account and record your first video in your browser.

Signup