Pepys

Guide

How to translate a transcript

A working guide for researchers and journalists taking an interview across languages – accurate quotes, surviving timestamps, and when you need a certified human.

The short answer

To translate a transcript, first transcribe it in the source language, then translate segment by segment so timestamps and speaker labels survive. Machine translation reads fluently but still makes meaning-level errors, so have a fluent speaker check the passages you'll quote or publish. For immigration, court, or research use, a certified or back-translated version is often required.

Transcribe in the source language, then translate the transcript

Do the transcription in the original language before you translate anything. The source transcript is the thing you translate from, so if it's wrong, the translation multiplies the error. Start with a clean, speaker-labeled draft in the source language, then translate that text – not the audio directly.

Translate segment by segment so the structure survives. A transcript translator that works cue by cue keeps every timestamp and speaker label in place, which matters more than it sounds. Flatten the transcript into one block and you lose the ability to jump back to the audio and check a line. A translated SRT subtitle file will then no longer sync to the video.

Keep speaker labels through the translation. Who said what is part of the meaning, especially in a multi-speaker interview or focus group, so preserve diarization rather than merging turns. Translate what each speaker says, not the identity: 'Speaker 1' stays 'Speaker 1' while their words change language.

Why does machine translation still need a human pass?

Machine translation reads fluently but still makes meaning-level errors a human has to catch. In a pairwise ranking experiment, raters preferred human over machine translation more strongly when judging whole documents than isolated sentences, because errors invisible at the sentence level become decisive across a document (ACL / EMNLP 2018).

Don't trust surface smoothness. A confidently phrased mistranslation of a quote is exactly the kind of error fluency hides, and it's the sort of thing that becomes a correction later. Read the passages you plan to quote against the source, ideally with a fluent speaker, before any of it goes to print.

Spend your review time where meaning is load-bearing. Negations, idioms, numbers said in passing, and named entities are the spots machine translation most often flips. Those are also the words you're most likely to quote, so that's where a human pass earns its keep.

How do you check a translated transcript is accurate?

The established check is back-translation. Translate forward, then have a second person independently translate the text back into the original language, and compare the two for meaning. Brislin's cross-cultural method built its five equivalence criteria on exactly this comparison of meaning (Journal of Cross-Cultural Psychology, 1970).

Do the back-translation blind. The person translating back shouldn't see the original source text, or they'll unconsciously reproduce it. Where the back-translation and the original diverge, that gap points you straight at the phrases whose meaning didn't carry, and you reconcile those before finalizing.

Reserve this for the high-stakes material. Survey items, questionnaire wording, and quotes headed for publication are worth back-translating; a full three-hour transcript you're only reading for understanding usually isn't. Match the rigor to how the translation will be used.

Translation is interpretation, not word-swapping

In research, translation is an interpretive act with consequences for your findings, not a neutral swap. Temple and Young argue that whether you identify the act of translation, and who performs it, carries epistemological implications for how the data gets represented (Qualitative Research, SAGE, 2004).

The translator isn't a neutral conduit. In a cross-language study, the translator becomes a producer of research data whose identity and experiences shape the analysis (Squires, 2009). Treat them as a co-analyst, not a passthrough, and record the choices they made about ambiguous or culturally loaded phrasing.

Write the translation procedure into your methods section. Name who translated, the language pair, their fluency, and how you verified equivalence. Reviewers of cross-language qualitative work expect that trail, and it's what lets a reader judge whether a translated quote can be trusted.

When do you need a certified translation?

Some uses require a certified human translation, and machine output won't do. USCIS rules require any foreign-language document to carry a full English translation the translator certifies as complete and accurate, plus a certification that they're competent to translate from that language (8 CFR 103.2(b)(3)).

Courts hold a similar line. Under the Court Interpreters Act, the Administrative Office certifies interpreter qualifications, and the presiding officer must use certified interpreters in proceedings the United States brings (28 U.S.C. 1827). The certified human record is what counts; automated output isn't a substitute.

So split the job by purpose. Machine-translate to understand a source or draft your first pass. Then commission a certified human translation only for documents that legally need one: filings, court records, and official submissions. Paying twice beats a rejected filing.

The steps, in order

  1. 01

    Transcribe in the source language

    Get an accurate, speaker-labeled transcript in the original language first. You translate from the transcript, not straight from the audio, so its accuracy sets the ceiling.

  2. 02

    Translate segment by segment

    Run it through a translator that works cue by cue so every timestamp and speaker label stays aligned, instead of collapsing the transcript into one flattened block.

  3. 03

    Have a fluent speaker review the quotes

    Machine translation reads smoothly but hides meaning errors. Read the passages you'll publish against the source with someone fluent, watching negations, idioms, and numbers.

  4. 04

    Back-translate anything high-stakes

    For survey items or quotes going to print, have a second person independently translate the text back into the original language, then reconcile any difference in meaning.

  5. 05

    Get a certified translation for official use

    For immigration filings or court, commission a certified human translator. Machine output isn't an accepted certified record, so keep it for understanding and drafts only.

Tips from people who do this a lot

  • Keep the source transcript beside the translation so any quote can be traced back to the original wording and its timestamp.

  • Watch negations, idioms, and numbers – they flip meaning quietly and are where machine translation trips most often.

  • In your methods section, name who translated, the language pair and their fluency, and how you checked equivalence; cross-language reviewers expect it.

  • Translate the words, not the speaker labels – keep 'Speaker 1' stable so diarization survives and you can still tell the turns apart.

  • Machine-translate first for your own understanding, then pay for a certified human translation only for the documents that legally require one.

Try it now

Drop in your recording or paste a link and get a clean, speaker-labeled transcript in minutes. Your first 60 minutes are free.

or paste a link
InstagramTikTokYouTubeFacebookSpotifyApple Podcasts

60 min free · no card required · we never train on your audio

PodcasterJournalistContent creatorResearcherStudent
Trusted by 100,000+ creators, podcasters, journalists & researchers

How to translate a transcript – questions, answered

Can I just use machine translation for a transcript?

For understanding or a research first pass, yes. But machine translation still makes meaning-level errors that are hard to spot, and raters prefer human translation more strongly across whole documents. Have a fluent speaker check any passage you'll quote, and use certified human translation for legal filings.

Do timestamps and speaker labels survive translation?

They can, if you translate segment by segment instead of flattening the transcript into one block. Cue-by-cue translation keeps each timing and speaker label aligned. A translated SRT or VTT still syncs to the original video, and you can jump back to the audio to check any line.

How do I verify a translated transcript is accurate?

Use back-translation. Translate forward, then have a second person independently translate it back into the original language, and compare the two for meaning. This comparison-of-meaning method underpins the established cross-cultural equivalence criteria and catches errors a single pass hides, especially in quotes and survey items.

When do I need a certified translation, not machine output?

For official use. USCIS requires foreign-language documents to carry a full English translation the translator certifies as complete and accurate, and federal courts require certified interpreters in proceedings the United States brings. Machine output isn't an accepted certified record for those, so use a qualified human translator.

Does translation change my research findings?

It can. Translation is an interpretive act, not a neutral word-swap, and who translates shapes how the data is represented. The translator effectively becomes a producer of research data. Document who did it, their fluency, and how you checked equivalence, and report that in your methods.

References

  1. 1.Läubli, Sennrich & Volk (2018), Has Machine Translation Achieved Human Parity? A Case for Document-Level EvaluationACL Anthology (EMNLP 2018)
  2. 2.Temple & Young (2004), Qualitative Research and Translation DilemmasQualitative Research (SAGE)
  3. 3.Squires (2009), Methodological challenges in cross-language qualitative researchInternational Journal of Nursing Studies (via PMC)
  4. 4.Brislin (1970), Back-Translation for Cross-Cultural ResearchJournal of Cross-Cultural Psychology (SAGE)
  5. 5.8 CFR 103.2(b)(3) – translations of foreign-language documentsCornell Legal Information Institute (eCFR)
  6. 6.28 U.S.C. 1827 – Court Interpreters ActCornell Legal Information Institute

Keep reading

Don't just take our word for it.

Ask ChatGPT, Claude, or Perplexity what Pepys is and who it's for. One click, and your favorite AI does the homework.

Get your transcript – free to start

Pay as you go – credits never expire, nothing to cancel. Or start free with 60 minutes, no card.