Translation

Translation takes a timed artefact — an SRT, a WebVTT file, or an alignment.json — and produces the same file in another language. Every start_time, end_time and segment index is copied byte-for-byte from your source. Only the text changes.

Subtitle cues are where a line ran out of width, not where a thought ended. A cue reading “called family” only makes sense as the tail of the cue before it — “…vultures I once”. Translated on its own it becomes a verb of summoning, which is not what the line says.

So VocaSync never translates cues one at a time. Your transcript goes to the model as continuous writing, and the finished translation is laid back across your original cues — with wording free to move across cue boundaries where the target language needs it. A phrase that recurs in your source gets the same rendering every time it appears.

You can translate into any of 57 languages — the same set the rest of VocaSync speaks.

The language you translate from is a smaller set: 13 languages. That is not an arbitrary limit. A timed artefact has to be produced by alignment in the first place, so a source language has to be one alignment supports. A target needs nothing but text.

A subtitle has to be readable in the time it is on screen. That is measured in characters per second (CPS), and translations often run longer than their source — so a cue that was comfortable in English can be too dense once translated.

Cues that exceed the limit are condensed automatically rather than shipped with a warning attached. The limit itself depends on the writing system: a Chinese character carries far more information than a Latin one, so the same characters-per-second figure is a very different amount of reading. You can set your own limit per project if you need a different threshold.

Register sets the voice of the translation:

• Formal — documentaries, news, business content, broadcast subtitling.
• Casual — vlogs, interviews, podcasts, contemporary fiction.
• Anime-faithful — keeps honorifics, name order and character speech patterns rather than domesticating them.
• Drama-faithful — keeps the formality shifts and terms of address that carry status between characters.

Some languages mark the speaker’s gender grammatically where English does not — Slavic past-tense verbs, for one. It cannot be inferred from a genderless source, so speaker gender is a setting you supply. Left unset, the translation prefers constructions that avoid the marking and stays consistent where it cannot.

You can also add a free-text context note describing the production, and both register and speaker gender are included at no extra cost.

Translated artefacts carry segment-level timings only. There is no per-word timing on a translated track, and that is deliberate: words do not correspond one to one between languages, and a translated word has no audio to measure against. Any per-word timing would be invented rather than measured.

£0.08 per 1,000 source characters. One rate — register and speaker gender are included, and there is no upgrade tier.

Translate from your own code with an API key. Endpoints and source-artifact handling are documented in the API reference.