Speech to Text Translation: How to Transcribe and Translate Audio

By whispertranscribe.ai Editorial Team · Last updated: November 2026 · 12 min read

Speech to text translation means taking a recording in one language and producing text in a different one. There are two ways to do it: transcribe the recording in its original language and then translate that text, or use a model that writes the target language directly from the audio. The first is easier to check. The second is faster.

Below: both pipelines and where each breaks, four practical methods including multilingual SRT and VTT subtitles, a quality-factor table, a post-translation checklist, and when a professional translator is the safer choice.

Quick answer: how to do speech to text translation

For most recordings, transcribe in the original language first, fix the transcript, then translate the corrected text. If you only need English and want speed, a direct speech-to-English tool is fine for a first draft. Either way, someone who knows both languages should check anything you publish or rely on.

  1. Transcribe in the spoken language. Pick the language yourself rather than relying on auto-detect.
  2. Correct the source transcript. Names, numbers, and terms first.
  3. Translate and review. Translate the text, then check it against the source and the audio.

What is speech to text translation?

Speech to text translation covers any process that starts with speech in language A and ends with text in language B. Plain transcription stays in one language. Plain translation starts from text. This task combines both, and how the two are combined decides how easy the result is to check.

Two pipelines: cascade vs direct

Researchers call the first approach a cascade: speech recognition, then text translation. The second is direct (or end-to-end) speech translation, where one model maps audio straight to target-language text. OpenAI's Whisper can do the direct version, but only into English.

Transcribe, then translate (cascade) Direct speech translation
Steps you see Two: source transcript, then translation One: audio in, target text out
Target languages Any language your text translator supports Often limited; Whisper's translate task outputs English only
Source-language text You keep it Usually not produced
Where you can catch errors Between the two steps, in the source text Only by comparing the output with the audio
Re-translating later Reuse the corrected transcript Run the audio again, or translate the English
Best for Research, publishing, subtitles in several languages, anything quoted Quick English understanding of a recording

What can go wrong in each pipeline

In a cascade, errors compound. If the transcript says "fifty" where the speaker said "fifteen," the translation will carry the wrong number in perfect grammar. The fix is cheap: correct the transcript before you translate it.

In a direct pipeline, there's no source text to inspect. A misheard word and a mistranslated word look the same in the output, so a reviewer has to go back to the audio. Direct output can also read more fluently than it deserves, which makes errors harder to spot.

Which method should you use?

Choose by what the text is for and how many languages you need.

Your situation Recommended method Why
You need to understand a foreign-language recording quickly Method 2: transcribe straight into English One step, readable draft
You'll quote the speaker in an article or report Method 1, plus Method 4 for key quotes You can check the original words
You need subtitles in two or more languages Method 3: source SRT or VTT, then translate the cue text One timing file serves every language
Research interviews to analyze and cite Method 1 Coding and citations stay tied to the original transcript
Contracts, court material, medical or immigration records Method 4 with a qualified professional Errors carry real consequences

Method 1: Transcribe in the original language, then translate the text

This is the most flexible route. The corrected transcript becomes your master copy, and every translation starts from it.

Step 1: Transcribe in the spoken language

Upload the file to a transcription tool and set the language before you start. In whispertranscribe.ai, a browser-based transcription tool powered by OpenAI's Whisper model, you choose from 130+ languages, with regional options and Noise / Low-Quality Audio/Video modes for rough recordings. There are dedicated pages for common sources such as Spanish audio and Chinese audio. For speakers who switch languages, the Multilingual Mix option covers English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch.

If the recording sits in Google Drive, Dropbox, Vimeo, or on YouTube, transcribe it from the URL instead of downloading it. Turn on speaker detection for interviews and meetings.

Step 2: Correct the source transcript

Fix the transcript before translating, because every uncorrected error gets copied into each language. Search for names, product terms, and acronyms first, then replay every number you plan to keep. Accuracy is up to 99% on clear audio. Background noise, strong accents, and crosstalk lower it, so plan for this pass.

Step 3: Translate the corrected text

Export the transcript as DOCX or TXT and run it through a text translator. Check the format limits first. As of November 2026:

  • Google Translate translates documents up to 10 MB in .docx, .pdf, .pptx, and .xlsx, and PDFs must be 300 pages or less, according to Google Translate Help. Subtitle files aren't on that list.
  • DeepL's file formats page lists DOCX, PDF, and TXT. It also lists SRT and VTT, including Pro Individual, Pro Team, Pro Business, and API plans, with a 150 KB or 200 KB cap per subtitle file. DeepL's free account doesn't cover TXT.

Give the translator context where the tool allows it: who is speaking, the audience, and how formal the result should be.

Step 4: Keep both versions side by side

Save the source and the translation as a pair with matching names, such as interview-04_es_v2.docx and interview-04_en_v2.docx, so any disputed line can be traced to the original words.

Method 2: Transcribe straight into English

If English is your target and you mainly need to understand or summarize the recording, a direct tool saves a step.

Using the translate audio to English page

whispertranscribe.ai's translate audio to English page turns speech in other languages into English text. The page describes it this way; it doesn't list a set of target languages, because the output is English.

  1. Choose the spoken language. Step 1 on the page asks for the language of the recording, not the target. The page says choosing it correctly gives more accurate English, and it suggests trying another option if the result isn't satisfactory.
  2. Upload the file. Step 2 takes the usual audio and video formats, including MP3, WAV, M4A, MP4, MOV, and MKV.
  3. Turn on speaker detection if more than one person talks. The page lists speaker detection for translation.
  4. Select Translate audio to English, then review the English text and export it.

Treat the output as a working draft; we don't promise translation quality. If you also need the original-language transcript, run normal transcription too.

Using open-source Whisper on your computer

OpenAI's Whisper repository documents a translate task that outputs English:

whisper japanese.wav --model medium --language Japanese --task translate

Two cautions from the README: the default turbo model isn't trained for translation and returns the original language even with --task translate, so use medium or large; and the output is English only. German audio into French text still needs a text translation step.

How good is it? The Whisper paper says the training data included 125,000 hours of X→English translation data. Without training on CoVoST2, the benchmark it was tested on, Whisper scored 29.1 BLEU, which the authors called a new state of the art. BLEU scores the overlap between machine output and reference translations. It is not a percentage of correct sentences. The same paper notes that Whisper still trails earlier systems trained directly for high-resource languages, and that more training data predicts translation quality far less reliably than it predicts transcription quality.

Method 3: Create multilingual subtitles (SRT and VTT)

For video, produce one subtitle file in the original language, then translate only the text of each cue. The timing stays identical across languages.

Step 1: Generate the source subtitle file

Use the subtitle generator or the MP4 to SRT converter and export SRT or VTT in the spoken language. Fix the text and timing in this file first; it's the template for every language.

Step 2: Translate the cue text, not the structure

An SRT cue has a number, a timecode line, and text. Only the text should change. Here's a made-up example:

12
00:01:04,200 --> 00:01:07,900
La reunión empieza el 3 de mayo a las 15:00.
12
00:01:04,200 --> 00:01:07,900
The meeting starts on May 3 at 3 p.m.

The cue number and timecode stay identical. The date and time are adapted to the target audience. If your translator only accepts documents, copy the text into a table and paste it back cue by cue. DeepL's Pro plans can take SRT and VTT directly, as listed above.

Step 3: Check reading length

A translated line can run longer than the original, and a cue that read comfortably in 3.7 seconds may no longer fit. Split long cues or shorten wording, but don't shift the timecodes unless you recheck them against the audio.

Step 4: Upload each language as its own track

Name files with language codes, for example talk.en.srt and talk.es.srt, and upload each as a separate subtitle track on the platform. Soft subtitle tracks let viewers switch languages.

Method 4: Add a human review step

For anything published, quoted, or used in a decision, have a bilingual reviewer compare the translation with the source. Machine output often reads fluently even when it's wrong.

A simple review sheet keeps this fast. Here's a made-up example of the format:

Cue Source line Translation Flag Reviewer note
#37 Vamos a cerrar el trato el lunes We'll close the deal on Monday OK —
#81 Eso no es moco de pavo That's not turkey snot Idiom Use "That's no small thing"
#126 Facturamos 1,5 millones We invoiced 1.5 million Number Confirm currency with speaker

Point the reviewer at names, numbers, quotes, legal or medical terms, and anything flagged unclear. whispertranscribe.ai doesn't offer human transcription or translation, so this step is yours or a professional's.

What affects speech to text translation quality?

Quality is set twice: once when the speech becomes text and again when the text changes language. Some factors hurt both.

Factor Effect on transcription Effect on translation What to do
Audio quality (noise, echo, distance) Missed or misheard words Mistakes carried into every language Record close to the mic; use the cleanest original file
Accents and dialects More errors on regional speech Regional words may be translated literally Pick the regional language option if one exists
Technical terms Terms spelled by sound Wrong or inconsistent equivalents Prepare a glossary and fix terms in the source
Language pair Depends on the spoken language Common pairs usually beat rare ones Budget more review for less common languages
Proper nouns Names turned into common words Names "translated" or changed List names before you start; search and fix
Code-switching Wrong language guessed mid-sentence Mixed text confuses the translator Use a mixed-language mode; tag each language
Overlapping speech Lines merged or dropped Speaker meaning lost Separate mics; check labels at interruptions

Google's speech data best practices say lossy codecs such as MP3 and M4A "may reduce accuracy," and noise and echo hurt more when a lossy codec is also used. OpenAI's README adds that Whisper's performance varies widely by language. Both are good reasons to test a short sample before a long batch.

Post-translation checklist

Run through this list after every translation, before anyone outside your team sees it:

  • Numbers. Every figure matches the audio. Watch decimal commas: "1,5" in many European languages is 1.5 in English.
  • Currency and units. Currency stays the speaker's unless you state a conversion; kilometers aren't silently changed to miles.
  • Names and titles. People, companies, and places are spelled the way they spell them, not translated.
  • Dates and times. Day-month order differs by country, so 03/04 can mean 3 April or March 4. Write the month as a word when it matters.
  • Quotes. Direct quotes are checked against the audio, not just the transcript.
  • Idioms and humor. Literal translations of idioms are replaced with natural equivalents, or flagged.
  • Cultural context. References that don't carry over get a short note rather than a silent rewrite.
  • Register. Formal and informal address (tú/usted, du/Sie) fit the relationship between speakers.
  • Speaker labels. Each line is still attributed to the right person.
  • Unclear passages. Anything uncertain is marked [unclear] in both languages instead of guessed.

When do you need a professional translator?

Use a qualified human translator when an error could cost money, health, legal standing, or trust. This is not legal advice, and requirements vary by country, court, and agency, so check with the body receiving the document.

Situation Why machine output isn't enough
Immigration filings US rules (8 CFR 103.2(b)(3)) require foreign-language documents sent to USCIS to come with a full English translation, certified by a translator who states it is complete and accurate and that they're competent to translate
Court evidence and depositions Courts often set their own rules for who may translate or certify
Contracts and agreements One wrong word can change an obligation
Medical consultations and records Dosages, symptoms, and consent must be exact
Published quotes and press material Misquoting someone damages trust and can create liability

For lower-stakes work like internal notes, research drafts, or understanding a video, machine translation plus your own review is usually enough.

Which export format do you need?

Pick the format for the next step. Transcripts in whispertranscribe.ai export as DOCX, PDF, TXT, CSV, SRT, and VTT.

Format Use it for
DOCX Translating in Word, DeepL, or Google Translate; editing with comments
TXT Pasting into a translator or another app
PDF A fixed copy of the approved translation
CSV Side-by-side review sheets and research coding
SRT Subtitle tracks for YouTube, video editors, and players
VTT Captions for HTML5 web video players

FAQs

What is the difference between speech to text and speech to text translation?

Speech to text keeps the original language: Spanish audio becomes Spanish text. Speech to text translation ends in a different language, so Spanish audio becomes English text. It can happen in two steps, by transcribing and then translating the transcript, or in one step with a model that writes the target language directly. The two-step route is easier to check because you can correct the source text before it is translated, and you can reuse it for other languages later.

Can Whisper translate audio into languages other than English?

No. Whisper's built-in translate task outputs English only. You can give it speech in many languages, but it writes English text. For any other target language, transcribe the audio in its original language and translate the transcript with a text translator. OpenAI's README also notes that the default turbo model isn't trained for translation and returns the original language even when you ask it to translate. Use the medium or large model for translation.

Is it better to transcribe first and then translate?

Usually, yes, when the result will be published, quoted, or translated into more than one language. Transcribing first gives you source text you can correct, so a misheard name or number gets fixed once instead of in every translation. Direct speech-to-English is quicker and works well when you just need to understand a recording. For subtitles in several languages, a corrected source SRT or VTT file is the most practical starting point.

How accurate is speech to text translation?

It depends on the audio, the language pair, and the subject matter, and there's no single accuracy figure that applies to all of them. Transcription on clear audio can be very accurate. whispertranscribe.ai reaches up to 99% on clear audio, while noise and accents lower that. Translation adds its own errors, especially with idioms, names, and technical terms. Common language pairs generally do better than rare ones. Check numbers, names, and quotes yourself, and get a bilingual review for anything important.

Can I translate subtitles and keep the timestamps?

Yes. Translate only the text of each cue and leave the cue numbers and timecodes unchanged. You can do this by hand, by pasting cue text into a translator and back, or with a tool that accepts subtitle files. DeepL lists SRT and VTT on its Pro Individual, Pro Team, Pro Business, and API plans. Google Translate's document feature doesn't list subtitle formats. After translating, check that long lines are still readable in the time each cue is on screen.

You can use it for a first look, but not as the final version. Legal, medical, contract, and immigration material often needs a qualified human translator, and some agencies require a certified translation. US immigration rules, for example, require a translator's certification that the English translation is complete and accurate. Requirements vary by country and institution, so check with whoever will receive the document. This is general information, not legal advice.


Ready to start? Upload a recording to translate audio to English, or transcribe it in the original language first for subtitles and other target languages. You can try it free with no account. Paid plans start from $9.99/month billed annually; see pricing. whispertranscribe.ai doesn't offer an API, human transcription, or a meeting bot.

References

Ready to turn your recordings into text?

Transcribe audio and video with whispertranscribe.ai.

Start transcribing ↗