Working With Footage in a Language You Don't Speak: A Reporter's Workflow

Two jobs: understanding and publishing

Footage in a language you don't speak creates two different tasks, and confusing them is where errors get into print.

Understanding

What is this clip about? Is there anything newsworthy? Which minutes should a translator look at? Speed matters more than polish, and an imperfect machine translation is usually good enough.

Publishing

What exactly did this person say, and what did they mean? Accuracy is everything, the words appear under your byline, and a machine translation alone is not good enough.

AI translation is a strong tool for the first job and a draft for the second. The rest of this workflow keeps that line clear.

Triage: what is in the footage

A translated transcript turns an hour of unfamiliar speech into something you can skim in minutes. With a timestamped transcript in the original language and a translation in yours, you can:

• Skim the translation for names, places, numbers and claims relevant to your story.
• Mark the timestamps of passages worth a professional's time, instead of paying for a full translation of every clip.
• Spot where several clips mention the same event or person, by searching the translations.
• Decide quickly that a clip is irrelevant, which is often the most valuable result.

Treat everything at this stage as a lead. Write "machine translation, unverified" on any notes you share with editors, so a paraphrase from triage never migrates into a draft as a quote.

What machine translation gets wrong in news footage

News footage is close to the worst case for automatic recognition and translation:

• Street audio, crowds, sirens and wind bury speech and cause skipped or invented words.
• Several people talk at once, and machine transcripts typically do not say who is speaking. Attribution, the core of a quote, is exactly what the transcript lacks.
• Names of people, places and organizations are misheard, then mistranslated or transliterated inconsistently.
• Dialects, regional slang, coded language and slurs may be flattened into neutral words or rendered literally, losing the meaning that made them newsworthy.
• Sarcasm, irony and rhetorical questions can be translated as sincere statements.
• Footage that switches between languages can be recognized inconsistently, as explained in translating mixed-language videos.
• Numbers, dates and negations are short words with large consequences, and are easy to mishear.

Errors compound: a recognition mistake in the original-language transcript becomes a confident-sounding error in the translation. That is why the original transcript matters as much as the translation during verification.

Verifying quotes with a fluent speaker

For every quote you intend to publish, broadcast or rely on for a factual claim:

1. Cut or note the exact clip with start and end timestamps, plus enough context before and after to judge meaning.
2. Find a fluent translator, ideally a native speaker familiar with the region and dialect. How to choose and brief one is covered in finding a reviewer for AI translation.
3. Ask them to translate from the audio first, without seeing the machine translation, so the draft cannot anchor their reading. Show them the machine version afterward only to compare.
4. Ask specifically about tone, register, dialect and anything that does not translate cleanly, and whether the speaker's identity or affiliation is evident from their speech.
5. Agree on the published wording, including any bracketed clarifications.
6. Record the translator's name, the clip, the timestamps, the date and the final wording in your verification notes.

For quotes that carry legal risk or could endanger someone, a second independent translation is worth the cost. Citing the time of the quote in your notes or published work follows the patterns in quoting videos with timestamps.

Protecting sources and handling footage

Before running footage through any tool, decide whether it may be processed by a third party at all. A voice can identify a person even when their face is hidden, so audio from a vulnerable source is sensitive in itself. Questions to settle first:

• Does your newsroom or organization have a policy on cloud processing of source material? Follow it.
• What leaves your device, and for how long is it kept? With mydubly, the video file stays on your device and only compressed audio is sent over HTTPS, then deleted within 30 minutes of completion. The details and their limits are discussed in on-device video privacy.
• Would running speech recognition on your own computer be more appropriate? Whisper is open source and can run locally for transcription, at the cost of setup time and slower processing on ordinary hardware.
• Are you working on a copy? Keep the original file untouched, with its metadata, for verification and for any later dispute about authenticity.

Translation is also separate from verifying the footage itself. Who filmed it, when and where are provenance questions; language can help, through accents or place names, but on-screen text, signs and captions in the picture are not part of a speech transcript and need a fluent reader.

Rules that keep errors out of print

• Never publish a machine translation of a quote, even with a disclaimer.
• Never present a paraphrase from triage notes as a direct quote.
• Do not publish an AI-generated voice speaking a real person's translated words as if it were them; viewers may take it as the person's own voice. Subtitles or a clearly labeled human voice-over are the established formats.
• Keep the original-language wording of every published quote in your notes.
• Disclose your translation method if your outlet's standards require it.

For broadcast or online video, translated subtitles over the original audio keep the speaker's real voice and tone, which matters for credibility. Dubbing replaces the voice entirely; the differences between formats are covered in dubbing vs voice-over.

Limits of AI translation for reporting

• Accuracy cannot be judged by someone who does not speak the language; fluent text can be wrong.
• No speaker labels means attribution always needs a human check.
• Short clips give recognition and language detection little context, so errors are more likely.
• Only supported languages can be processed. Many languages that matter in international reporting are not covered by any given tool.
• Machine translation gives one reading. Ambiguous phrases, where a human translator would flag two possible meanings, are silently resolved.

Where mydubly fits in a newsroom workflow

mydubly is suited to the triage stage. You add a video or audio file from your device, the spoken language is detected automatically, and a transcript job returns the original-language transcript with timestamps, SRT and VTT subtitles and, if you choose a target language, a translated transcript, all at 1 credit per minute with a 5-credit minimum per file. Files can be up to 2 hours long.

It supports 21 languages, including Arabic, Turkish, Russian, Hindi, Chinese, Spanish and French, but many languages that matter in international reporting, such as Ukrainian, Bengali, Tamil and Indonesian, are not supported, so check the list before relying on it. It cannot fetch footage from a link, so you need the file itself; downloading from social platforms is subject to their terms. It does not label speakers, read on-screen text, or certify translations, and its translated subtitles, like any machine output, need a fluent speaker's check before broadcast. Start from the video translator, or the interview transcription page if you are working with your own recorded interviews.

Next step

Set up the verification log before the next story: a simple sheet with clip, timestamps, original wording, machine translation, human translation, translator and date. Then triage the footage you already have and send only the moments that matter to a fluent colleague.