Most of what lands here is not in English
That is not a guess, and it is not somebody else's survey – it is our own count. Of the transcripts this site produced in the two weeks to 31 August, 63% were in a language other than English – Russian, Portuguese, Spanish and Mandarin lead, then Polish, German, Norwegian and French. People are not bringing us English videos and asking for English text. They are bringing whatever they found.
Which raises the obvious question: if you cannot follow the language, what use is a transcript of it? That is the gap this page is about.
How it works, and why that matters for the timings
The audio is transcribed first, in its own language, the way it always is. Then each segment of that text is translated in place, one line at a time, and the English is put back into the same segment it came from.
The order is the whole trick. Because nothing is merged or re-split, the English lines keep the start times of the speech they came from – so you can still scrub to a moment, still quote with a timecode, and still export subtitles that line up with the picture. A translation that lost the timings would be a wall of text, and a subtitle file made from it would drift further out of sync with every line.
We check the count on every batch. If the translation comes back with a different number of lines than went in, the job is refused rather than patched up – because a transcript that has slipped by one line reads as perfectly correct and is wrong about when every remaining word was said.
What it is good for, and what it is not
- Following a talk you cannot otherwise follow. A conference session, an interview, a lecture in a language you do not read. This is the case it does best.
- Searching and quoting. English text you can grep, paste into notes, and cite with a timecode pointing back at the original audio.
- Not a substitute for a translator. Dense idiom, wordplay, overlapping speakers and unfamiliar proper nouns are where it thins out. If a sentence carries weight, the timestamp tells you exactly where to go and listen.
- English into other languages – no. The direction is one-way by design: any of the hundred languages into English, not the reverse.
One direction, and uneven across languages
Two limits matter before you lean on this. The first is direction. Whisper's translation task runs into English and nowhere else – the 2022 paper evaluates it as X→en speech translation (Radford et al., 2022, Table 4), and the model card says the same thing in one line: for translation, "the target text is in English". So French audio becomes English text. English audio does not become French.
The second limit is that quality is not flat across those hundred languages, and OpenAI says so itself: the models "perform unevenly across languages", with lower accuracy on low-resource ones. The training mix explains why. Of the 680,000 hours behind the 2022 model, 117,000 cover 96 other languages between them, with a further 125,000 hours of translation data. A Norwegian podcast and a Swahili one do not land in the same place, and we would rather you heard that from us than discovered it in a transcript.
Why we do not do it the obvious way
Whisper has a translate mode built in, and using it would have been the short path. We measured it before trusting it, and it does not hold up: on the same Swedish recording, the checkpoint we run returned the original Swedish with the flag on and the flag off – the same text, twice. A larger checkpoint translated some segments and left others in Swedish, while labelling the result English either way. Passing the source language, dropping the temperature to zero and steering with a prompt changed none of it.
A feature that works two times in three, and cannot tell you which time you got, is worse than no feature. So the translation step moved off the audio and onto the text, where it is repeatable. If it fails, it says so.
What it costs you
Nothing, and there is no account to make. Translation sits inside the same free tier as everything else here: fifteen transcripts a day from one address, up to 75 minutes of audio each. We sell speed, not features – the paid lane returns the same transcript sooner, and everything on this page works without it.
If your source needs different handling, those have their own pages: YouTube videos, podcast episodes, Instagram reels, and direct audio links.
Frequently asked
How do I get the English version?
Paste the link, open the options row under the box, and tick "translate to English". You get one transcript back, in English, with the original timings intact. Leave the tick off and you get the words as they were spoken.
Which languages can it translate from?
The 100 the engine recognises, which is every language in the picker – from Spanish and Mandarin down to Welsh, Nepali and Cantonese. The direction is one-way: any of them into English, not English into them.
Do the timestamps survive translation?
Yes, and that is the part worth explaining. The audio is transcribed first, then each segment of text is translated in place. Segment boundaries never move, so the English lines carry the same start times as the original speech – which is why the .srt and .vtt exports still work.
How good is the translation?
Good enough to read, follow and quote from; not good enough to publish as a translator would. It handles ordinary speech well and gets less reliable with dense idiom, overlapping speakers and proper nouns it has never seen. If a sentence matters, check it against the original – the timings tell you exactly where to listen.
Can I get both the original and the English?
Run it twice: once with the tick off, once on. The second pass reuses the audio we already fetched, so it costs you only the wait, not another download.
What if it cannot translate something?
You get the original transcript with a note saying it was not translated. What you will never get is the source language quietly presented as English – that failure is worse than an honest one, because it looks finished.
Does it cost anything?
No. Translation is part of the free tier, on the same terms as everything else: fifteen transcripts a day per address, up to 75 minutes each, no account.
Sources
This page mixes two kinds of evidence, and they should not be read as the same thing.
Published research and vendor documentation:
- Radford, Kim, Xu, Brockman, McLeavey, Sutskever. Robust Speech Recognition via Large-Scale Weak Supervision (2022, arXiv:2212.04356) – X→en speech translation in Table 4, and the training split: of 680,000 hours, 117,000 cover 96 other languages, plus 125,000 hours of translation data.
- openai/whisper-large-v3-turbo model card – English as the only translation target, and the "perform unevenly across languages" caveat, quoted rather than paraphrased.
Our own measurements, which nobody else has published:
- The 63% figure is a count of transcripts this site produced over the two weeks to 31 August 2026. It describes what people bring us, not language use in general.
- The Swedish result above – the same recording returning Swedish with the translate flag on and off – is our own test. We ran it before we shipped this page, and it is why translation runs on the text instead of the audio.
You already have the link.
Paste it above, tick "translate to English", and read it in a language you actually speak.