Transcripts you can search
Source-language transcripts and translations into your working language make hours of footage searchable by name, place and phrase.
For OSINT & research analysts
Open-source research runs on video and audio in languages the team does not speak, and linguist time is the scarcest resource it has. Octavia transcribes and translates at volume so analysts can search, triage and decide what deserves a linguist — with the source kept alongside every line, and nothing treated as verified until a person has checked it.
The specific reasons this content usually stays in one language.
Every minute of foreign-language footage queues for the same few people, so most of it is never reviewed at all.
A name or a place mentioned forty minutes into a recording is invisible until someone sits through the forty minutes.
Regional dialect, code-switching, slang, phone microphones and street noise — the opposite of what translation is usually tested on.
The parts of the platform that matter for this work, rather than the full feature list.
Source-language transcripts and translations into your working language make hours of footage searchable by name, place and phrase.
You get the original-language transcript as well as the translation, both timecoded, so any line can be checked against the audio and quoted in the original.
A locked glossary fixes how people, places and organisations are transliterated across languages and scripts, so one name does not become four search terms.
Diarization separates voices in interviews, panels and street recordings, so a translated line stays with the speaker who said it.
Submit large sets of files programmatically and route the results into your own search and case tools, rather than uploading one file at a time.
One workflow across 60+ languages, so the decision about what needs a specialist linguist is made on content rather than guesswork.
Four steps from the file you already have to every language you need.
Confirm the platform’s terms and the privacy law that applies before bulk-downloading anyone’s content.
Batch or API submission across the set, with names and places pinned in the glossary.
Analysts search the translated text, flag relevant passages and send those, with the original audio, to a linguist.
Anything a finding, publication or decision rests on is checked by a human linguist against the source.
One upload in, every format you need to publish out.
Including the ones where the honest answer is a limitation.
Not on its own. Machine translation of unscripted open-source material is a triage tool: it tells you where to look and what to send to a linguist. Anything a published finding, an attribution or a decision rests on should be verified by a human linguist against the original audio, and quoted with the original-language text alongside. Treat the translation as a lead, not as the evidence.
Unevenly, and you should plan for that. Accuracy is strongest on widely spoken standard varieties and clear audio, and falls with regional dialect, code-switching, in-group slang, background noise and languages with less training data. A fluent-looking translation can still be wrong. Sample each new source type with a linguist before trusting it at volume, and weight your triage accordingly.
Yes. Use it on open-source material you are entitled to process, respecting platform terms, copyright and the privacy and data-protection law that covers the people in the footage — which often still applies when content is public. It is for researching public information, not for covert surveillance of individuals. Do not clone or synthesise the voice of anyone in the material; transcripts and subtitles are the right output.
Customer content is not used to train models, is encrypted in transit and at rest, and retention is configurable, so working copies need not outlive the project. Role-based workspaces keep one investigation’s material away from another team, and a defined processing region is available where you require one. If your work involves sensitive sources or legally restricted material, raise your requirements with us before uploading.
Practical guides from the Octavia editorial team.
The same engine, pointed at a different kind of work.
Start with a backlog you already hold, transcribe it in one pass, and see how much of it your linguists actually needed to hear.