To turn foreign-language audio into English text, you need a tool that translates as it transcribes, not one that transcribes and then translates. Whisper does this in a single pass with its translate task, and produces noticeably better results than transcribing first and running the text through a translator afterwards.
Most people try the two-step route first: transcribe the Russian, then paste the Russian into Google Translate. It works, but it compounds errors. A word the transcriber got slightly wrong becomes a sentence the translator gets completely wrong.
This guide covers the one-pass approach, when the two-step route is still the right call, and what to expect for accuracy.
Why one pass beats two
When a model transcribes and translates together, it has the audio available while deciding what the English should say. Tone, emphasis and context are all still there. When you translate a transcript afterwards, all of that is gone. The translator sees only text, including whatever the transcriber got wrong.
The difference shows up most with homophone errors. If a transcriber mishears a Russian word as a similar-sounding one, a one-pass model still has the audio and the surrounding context to catch it. A separate translator does not. It faithfully translates the wrong word and the sentence quietly changes meaning.
Two-step is still the right choice in one case: when you need both the original-language transcript and the English. Then you want the source text as a deliverable, not just as an intermediate step.
How to transcribe foreign-language video into English
- Get the audio or video file onto your device.
- Install Whisper, or open a transcription tool that offers English output for non-English audio.
- Run the transcription with the translate task rather than the default transcribe task.
- Review the output for names, numbers and any culturally specific terms.
Method 1: Whisper's translate task
OpenAI's Whisper is free, open source, runs on your own machine, and handles around 99 languages. Most guides cover the basic command and stop there, missing the flag that actually matters here.
The default behaviour transcribes into the source language:
whisper interview.mp3 --model mediumRussian audio in, Russian text out. Adding the translate task changes that:
whisper interview.mp3 --model medium --task translateRussian audio in, English text out, in one pass. You can add --output_format srt to get timed subtitles instead of plain text.
Three things worth knowing before you rely on this:
- Translate only goes to English. This is a real limitation, not a setting you have missed. Whisper's translate task targets English and nothing else. For Russian to Spanish you still need two steps.
- Model size matters much more here than for same-language work. The
tinyandbasemodels are usable for clear English speech but noticeably weaker at translation. Usemediumat minimum, andlargeif you can wait. - It is slow without a GPU. The
mediummodel on CPU can run several times slower than realtime. A one-hour interview may take a few hours. Start it and walk away.
Method 2: An online tool with language output settings
If you do not want to install anything, some transcription tools let you choose an output language separately from the source audio. Paste a link or upload the file, set the output language, and the transcript comes back in that language.
Check before committing to a long file: run a two-minute clip first and confirm the output is actually in the language you selected. Some tools label this feature ambiguously and will hand you the source language back.
OmniveraLabs' transcription tool has an output language setting, and accepts uploads as well as YouTube links.
Method 3: Transcribe, then translate
The two-step route, for when you need both versions:
- Transcribe in the source language, using Whisper's default task or any transcription tool.
- Review the source transcript and fix obvious errors before translating. This step is what makes two-step viable.
- Translate the corrected text with DeepL, Google Translate, or a language model.
Skipping step two is why two-step gets its bad reputation. Feeding an unreviewed transcript into a translator guarantees the errors survive into the English.
What to expect by language
Accuracy varies more than most tools admit, largely by how much training data existed for a language.
| Language group | Examples | What to expect |
|---|---|---|
| Well resourced | Spanish, French, German, Russian, Mandarin, Japanese | Strong. Usable with light editing |
| Moderately resourced | Turkish, Hindi, Urdu, Arabic, Vietnamese, Tagalog | Good on clear speech. Expect to fix names and idioms |
| Lower resourced | Nepali, Pashto, Yoruba, Tibetan | Variable. Test a short clip before committing |
| Dialect-heavy | Regional Arabic, Swiss German, Cantonese vs Mandarin | Weaker than the standard form of the same language |
Two things affect quality more than the language itself: background noise, and whether the speaker uses the standard register or a strong regional dialect. A clean recording in a lower-resourced language often beats a noisy one in a well-resourced language.
Reviewing translated output
Translated transcripts fail differently from same-language ones, and the errors are harder to spot because the text reads fluently.
- Names get anglicised or invented. A model unsure of a proper noun will produce something plausible-sounding. Check every name against a source you trust.
- Idioms come through literally. A phrase that is ordinary in the source language can read as strange or nonsensical in English. If a sentence seems oddly specific or surreal, it is usually an idiom.
- Numbers and units shift. Watch for dates in the other convention, and for units that were converted when they should not have been, or not converted when they should.
- Formality flattens. Languages with formal and informal registers lose that distinction in English. If the register matters to your use, note it separately.
- Fluent does not mean accurate. This is the important one. Automatic translation produces confident, natural-reading English even when it is wrong. Do not use readability as a proxy for correctness.
If the material is going to be published, quoted, or used in any legal or medical context, have a human speaker of the source language check it. Automatic translation is a strong first draft, not a final one.
Subtitles in English for a foreign-language video
If the goal is subtitles rather than a document, the same one-pass approach applies. Whisper's translate task with --output_format srt produces an English SRT timed to the original audio.
A note on line length: translated text is often longer than the original. English translations of German or Russian frequently run longer than the source, which can push subtitle lines past comfortable reading length. Check for lines over about 42 characters and split them.
Our guide to converting audio to SRT covers the file format and how to edit timings.
Working with a video instead of a local file? OmniveraLabs' free transcription tool accepts uploads and YouTube links, with an output language setting. No software install.