Speech translation in one step
The usual way to translate a video is a two-stage pipeline: transcribe the speech in its own language, then pass the text through a machine translator. Every stage adds its own errors, and the second stage sees only the words, not how they were said.
This tool takes the shorter path. Whisper, the speech model behind it, was trained not only to transcribe but also to produce English directly from foreign-language audio; about 125,000 hours of its training data were exactly that pairing of non-English speech with English text. In translation mode it listens to the recording and writes English, with no intermediate transcript in the original language.
That gives a few practical advantages:
- Timings stay attached to the English text. Each translated segment keeps the start and end time of the speech it came from, so SRT and VTT output lines up with the video.
- No compounding errors. In a two-stage pipeline a misheard word becomes a confidently wrong translation; here the model works from the sound itself, so there is no flawed transcript for a translator to build on.
- One job, one wait. You get English text or subtitles back from a single upload.
The trade-off is that the target language is fixed: Whisper translates into English only.
Good uses and weak spots
It works well for understanding and light publishing:
- Following a conference talk, lecture or interview recorded in a language you do not speak.
- Adding English subtitles to a short clip for an international audience.
- Getting the gist of a voice message, a news segment or a customer video before deciding whether it needs a professional translator.
Know where it struggles:
- Proper names and terms. Names of people, places and products can be translated literally or misspelled. Check them before you publish.
- Idioms and humour. Fixed expressions are sometimes rendered word for word.
- Mixed-language speech. When speakers switch languages mid-sentence, parts may come out untranslated or skipped.
- Occasional untranslated stretches. Now and then a segment stays in the original language. Choosing the spoken language explicitly instead of leaving it on Detect makes this rarer.
- Recordings already in English. These are simply transcribed, so there is no point running them here; use transcription instead.
Choosing the output
Three formats are offered, and the English text also appears on the page when the job is done, ready to copy:
- TXT puts each translated phrase on its own line, without timings. Best for reading, notes, quotes and pasting into a document.
- SRT is the safe choice for subtitles: video editors, desktop players and most upload forms accept it.
- VTT is for HTML5 video on a website, where the
<track>element accepts only VTT.
To show the English subtitles, add the SRT as a subtitle track in your editor or player, or keep it next to the video with the same file name. If the subtitles turn out to be a second or two off, or you need another format later, the subtitle converter can shift the timings and switch between SRT and VTT.
Preparing the file
Video and audio are both accepted: MP4, MOV, MKV and WebM, plus MP3, WAV and M4A. Only the sound is used, so a large video can be replaced by its audio track to make the upload faster. A 300 MB screen recording usually becomes 10 to 15 MB as an MP3, and the translation comes out the same.
Recognition takes roughly as long as the recording itself, and the free limit is measured in minutes of audio rather than megabytes. If a recording is longer than your limit, split it at natural pauses and translate the parts one at a time, so no sentence is cut in half at a boundary.
Processing happens on our servers because the model is far too large to run in a browser tab. The upload is deleted once the result is ready, and the result itself expires automatically within about an hour, so download or copy it promptly.