Transcribe Video

No. The speech model is downloaded to your browser and the transcription happens on your own machine — the recording is never sent to us or to anyone else. You can disconnect from the internet after the model has loaded and it will still work.

Yes, and it asks the speech model itself — the same model that transcribes also identifies languages. On the Dutch recording we test with, it named Dutch with 96% confidence. The language it settled on appears in the dropdown as soon as it knows, so you can see it was right, or stop within seconds if it was not. If you already know the language, naming it yourself is still the surest route.

Less time than the recording lasts, on every machine we measured. A ten-minute video took about 100 seconds on a laptop with a graphics chip and about six and a half minutes on one without. The tool shows an estimate while it runs, and it gets more accurate as it goes because it measures your machine rather than guessing.

Over twenty, including English, Danish, Dutch, German, French, Spanish, Italian, Polish, Finnish, Swedish, Norwegian, Portuguese, Russian, Ukrainian, Turkish, Arabic, Hindi, Japanese, Korean and Chinese. Choose one before you start, or let it work the language out from the audio — see below.

Into English, yes — there is an option for it. The model can write down what it hears, or translate it into English, and it cannot do any other pair. Danish to German is not something it can do, so we do not offer it rather than offering it badly.

Anything your browser can play: MP4, MOV, WebM, MKV and AVI for video, and MP3, M4A, WAV, FLAC, OGG and Opus for audio. It reads the sound out of a video, so you do not need to extract the audio first.

Plain text, Word (.docx), PDF, Markdown, CSV for a spreadsheet, JSON, and SRT or VTT subtitle files. The document formats can carry a timestamp against each paragraph, and the subtitle files carry the timings a video player needs.

Good, and honest about not being perfect. On a nine-minute Dutch interview it matched a desktop transcription tool nearly word for word; what it gets wrong is mostly proper nouns and names, which is what speech recognition is generally worst at. Read it through before you rely on it.

Twenty minutes per file. That is a memory limit rather than a rule — decoded audio is held in your browser's memory, and past twenty minutes a tab is gambling with your machine. Split a long recording with our audio trimmer first.

It downloads the speech model. On a computer with a graphics chip that is about 920 MB, because only full precision transcribes correctly — we measured the smaller versions and one was ten times slower while the other quietly produced nonsense. Without a graphics chip it is about 240 MB. Your browser keeps it either way, so the second file starts immediately, and “Faster, rougher” uses a model around a sixth of the size.

It is downloaded from Hugging Face, the public repository the model is published on. That is the one request this page makes to anybody else, and it goes the other way: files come down, nothing goes up. Your recording is never part of it — it is handed to the speech model inside your own browser and never touches the network. Everything else the tool needs, including the code that runs the model, is served from helpyself.com.

No cookies, and nothing from your recording or your transcript. The site keeps a few general preferences — your theme, whether you are signed in — and this tool adds one thing: the speech model itself, held in your browser's storage so the second file starts immediately. That is the 920 MB (or 240 MB without a graphics chip) described above, and it stays until you clear this site's data in your browser, which removes it completely.

Because something in the recording was not speech. Music, applause, a held note or a room with a fan in it will send any speech model into a repeating loop — it writes the same phrase over and over, fluently and in the right language, and nothing about the output says it is invented. When that happens the passage is decoded again a few times with more randomness, and if it still loops it is left out and the summary says so. A gap you can see beats a paragraph of dialogue that was never spoken.