Translate subtitles without losing the timecodes

Timecodes survive translation

Most subtitle translators re-run the video through a model and generate a new set of cues from scratch — the new file rarely lines up with the old one, and you retime it by hand. This does something different: it takes the cues that were already generated during transcription and sends only their text to Claude for translation, one cue at a time. The start and end of every cue is copied through untouched.

1
00:00:01,120 --> 00:00:03,860
Wir haben den Umsatz im dritten Quartal verdoppelt.

2
00:00:04,400 --> 00:00:07,900
Das lag hauptsächlich am neuen Vertriebsteam.

1
00:00:01,120 --> 00:00:03,860
Nous avons doublé notre chiffre d'affaires au troisième trimestre.

2
00:00:04,400 --> 00:00:07,900
C'était surtout grâce à la nouvelle équipe commerciale.

00:00:01,120 --> 00:00:03,860 and 00:00:04,400 --> 00:00:07,900 appear on both sides, unchanged. Only the sentence in between is different.

103 translation languages, separate from transcription

Transcription and translation use different models and different language lists. The recording only needs to be in one of the languages the transcription model recognizes — translation afterward can target any of 103 languages, because that step is handled by Claude working from text, not by re-listening to the audio.

Frequently asked questions

Do I need to already have subtitles to use this?

No — you need the audio or video. Upload the recording, get an English (or auto-detected) transcript with SRT and VTT files, then request a translation from the finished job. There is currently no path to upload an existing subtitle file and translate only that, without transcribing the source recording first — if that is what you need, the format converters at /tools can at least switch an existing file between SRT and VTT while you decide.

Why does translation keep the same timecodes as the original?

Because it never touches the audio again. The transcription pass already produced timestamped segments — start and end time for each spoken chunk. Translation sends the text of those segments to Claude and gets back the same number of segments in the target language, then rebuilds the SRT or VTT from the original start and end times. Nothing is re-aligned, because nothing needs to be.

How many languages can I translate into?

The translation dropdown on a finished job lists 103 languages, separate from the 66 languages the transcription model can recognize in the source audio. A recording only needs to be in a language the model understands; the translation target list is much larger because Claude, not the transcription model, produces the output text.

What if a cue in the original subtitles is wrong?

The translation is only as good as the segment it started from — this pipeline translates, it does not re-transcribe or correct. Fix a cue in the source language first (or re-run transcription with a different model) if the wording is off, then request the translation again.

Can I get a translated PDF or plain-text transcript too, not just subtitles?

Yes. The same postprocess step accepts the plain transcript, the AI summary, or the action list as its source instead of SRT or VTT — those don't carry timecodes to preserve, so they translate as flowing text rather than cue by cue.

Does translating cost extra on top of transcription?

No separate per-translation charge — it runs against the same job, billed only for the audio minutes already transcribed. 10 minutes are free every day.