How do you translate an audio file into another language?

To translate an audio file into another language, transcribe the recording in the language it was spoken, check that transcript against the audio, and then translate the text into your target language. Some speech tools translate the sound directly, which saves a step. Many of them produce only English, and the ones that offer other targets have their own language lists and prices. For English output, our guide to translating audio and video to English covers each route in more depth.

Transcribe first, then translate, when the words matter

The two-step route is the most reliable one, because you can check the middle of the process. First, transcribe the recording in its own language, and set that language yourself rather than relying on detection. Second, read the transcript against the audio, correct it, and only then send the text to a translator. A translator cannot fix a word it was never given correctly.

Skipping that check has a measurable cost. In a 2026 study of Vietnamese speech translated into English, translating the recognizer's output instead of clean transcripts cost between 6.79 and 10.64 BLEU points ¹. BLEU is a score for how closely a machine translation matches a human reference, so a drop of this size is a real loss of quality. The study covers one language pair and read speech, so it shows how large the effect can be, not how accurate any particular app is. The question on speech translation accuracy explains how these errors carry through to the translation.

The steps are simpler on a phone. Samsung's Voice Recorder transcribes a memo, and after that you can translate the memo into other languages, share it or move it to Samsung Notes ². That is one app's feature, so check that your recorder offers it, and which languages it supports, before you rely on it for anything important.

Some speech models translate directly, within a fixed list of languages

A direct route skips the separate transcript. Meta's SeamlessM4T is a single model that supports speech-to-text translation and speech recognition for up to 100 languages ³. The words "up to" matter. Coverage varies by task, and translation quality differs between language pairs, so a language on the list can still give weak output into your target.

Smaller apps cover less. The F-Droid listing for Seamless says that, as a standalone app, it translates between five languages: English, Spanish, Portuguese, Hindi and Russian ⁴. If your pair is not on a short list like that, the app will not help, however well it hears the speaker.

Direct models avoid one hand-off between two systems, and that is their appeal. The SeamlessM4T authors list error propagation from speech recognition into translation as a known weakness of cascaded systems, which run recognition and translation as separate models ⁵. Avoiding the hand-off does not remove the need to check the output against the audio, so the same check still applies.

Direct translation also changes the bill. Microsoft says you are charged for the speech to text transcription and the text translation for each target language ⁶. Asking for three target languages costs more than asking for one, so request only the languages you need.

Many tools translate only into English

Not every translator accepts a target language. OpenAI's translations endpoint "supports translation into English only" ⁷. Whisper's paper describes its translation training data as speech from other languages into English, with 125,000 hours of that data ⁸. If you need English, either route works. For any other target, transcribe with Whisper and translate the transcript with a separate translator.

For speech with more than one language in it, the transcript needs extra care. If the speaker switches between languages, expect errors at each switch, and see the glossary entry on code-switching.

It depends on the language pair, the price and where the audio goes

The right route depends on three things you can check before you start. The table compares the routes by what they output.

Three checks settle most choices. First, confirm the target language is on the route's list, for your exact pair and not just the language family. Second, look at the price per target language, since each extra target can add to the bill. Third, decide whether the audio may leave your device. A cloud service sends the file to a server, and the provider's policy then governs it. An on-device route keeps it with you.

What to do

1. Pick the target language and confirm the route you want covers that exact pair.
2. Set the spoken language and transcribe the file, rather than letting the tool guess.
3. Read the transcript against the recording, and correct names, numbers and dates first, since they are the costliest to get wrong.
4. Translate the checked transcript into your target language.
5. Have a speaker of the target language read the result for anything you will quote, sign or act on.

Two related questions go further. For a recording in Zulu or Xhosa, see the question on translating Zulu or Xhosa audio. For apps and APIs that return a translation from one call, see the question on which APIs combine speech to text and translation. The broader idea of turning speech into text in one language and text in another is covered by the glossary entry on speech translation.