Can Whisper translate audio?

Yes. Whisper can take speech in another language and output English text. It does not translate into other languages: the translate task is an English-only output. For the translation step, use a multilingual model rather than turbo.

Whisper translates speech into English, not into other languages

The model's training data includes translation pairs. The Whisper paper reports 125,000 hours of speech translated into English, which it describes as X to English translation data ¹.

The Whisper README shows the translate task with a Japanese recording. It sets the model, the spoken language and the task together: whisper japanese.wav --model medium --language Japanese --task translate ( README on GitHub) ².

So the answer to "can it translate?" depends on the target. Into English, yes. Into Spanish, German or Chinese, the documented translate task does not apply, because its output is English. The guide to Whisper's languages and accuracy covers the full set of languages the model handles, and the guide to translating audio to English explains which apps and services offer translation on a phone or computer.

Use a multilingual model for translation, not turbo

Whisper has several sizes, and the choice matters for translation. The repository says its turbo model is an optimized version of large-v3 that transcribes faster ³, and that it is not trained for translation tasks ⁴. The README also makes turbo the default: the default setting selects it, and it works well for transcribing English ⁵.

If you ask for English and get the original language back, check which model you ran. The README warns that turbo returns the original language even when --task translate is specified ⁶. It recommends medium or large for the best translation results ⁷.

In practice, pick one of the multilingual models for any recording you want in English. Turbo is a good default for transcribing English, and it is the wrong choice for translating other languages.

Get the original transcript and the English version in two steps

A translated transcript carries the errors of the original transcript. If the first step misreads a word, the translation of that word is wrong too. So run the two steps separately and check the original first:

1. Transcribe the recording with the language set, for example --language Spanish, and read the result against the audio.
2. Run the translate task on the same file, with a multilingual model, to get the English version.
3. Compare the English output with the original for names, numbers and technical terms, which translation often changes.

Keeping both outputs also helps you see which step introduced an error. The question on why Whisper transcribes in the wrong language covers the case where the language setting is the problem rather than the translation.

Where translation fits in the wider picture

Translation is a separate job from recognition. The recognizer has to be right first, and the accuracy of the original language, which varies widely, sets the ceiling for the English output. The guide to Whisper's languages and accuracy shows how much that varies, from low word error rates in Spanish to much higher rates in languages with little training data.

Two sibling questions cover the next choices. The question on how good Whisper is at Hindi shows how accuracy changes for one language, and the question on which Whisper model is best explains how to choose the size you run.

For a broader view of tools that translate speech, see the overview of languages and accents, and the guide to speech to text with translation. The term for the job is speech translation.