Translation

Whisper can translate non-English speech directly into English text. The same model that transcribes also translates — translation is selected by the <|translate|> task token in the start-of-transcript prefix, instead of <|transcribe|>.

Surfaces

CLI:

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

Python:

result = model.transcribe("japanese.wav", language="ja", task="translate")

The output result["text"] is in English; result["language"] still reports the detected source language ("ja").

What models support translation

• All multilingual models (tiny, base, small, medium, large-v1, large-v2, large-v3).
• Not the turbo model. The README spells this out: "the turbo model is not trained for translation tasks". Even when called with --task translate, turbo returns the original language. For best translation use medium or large.

The English-only models (tiny.en, base.en, …) cannot translate at all.

How translation is wired in

The token prefix Whisper feeds the decoder is built by Tokenizer.__post_init__ in whisper/tokenizer.py:

sot_sequence = [<|startoftranscript|>]
if language is not None:
    sot_sequence.append(<|<lang>|>)
if task is not None:
    sot_sequence.append(<|transcribe|> if task == "transcribe" else <|translate|>)

get_tokenizer(multilingual=True, language="ja", task="translate") returns a Tokenizer whose sot_sequence ends with the <|translate|> token. DecodingTask uses that prefix to seed the decoder, so every subsequent autoregressive step is conditioned on "the user wants English from this language."

Caveats

• Word timestamps for translations are explicitly flagged as unreliable by the code: whisper/transcribe.py emits warnings.warn("Word-level timestamps on translations may not be reliable.") when both --word_timestamps and --task translate are used. Because the alignment between source-language audio and English target text isn't 1-1, the DTW alignment heuristic won't be as tight as for transcription.
• The translation target is always English; there is no path to translate into other target languages with the published models.

See also

• Tokenizer — how the SOT sequence is assembled.
• Decoding — how the prefix conditions the decoder.
• Reference: models — model-by-model capabilities table.