How do you translate Japanese audio to English?

Run Whisper's translate task on the recording, set the spoken language to Japanese, and use the medium or large model. The output is English text, and Whisper runs on your own computer as a command-line program ¹. On one published test, Japanese scored lower than German for translation ^{2 3}, so read the English against the audio before you rely on it.

Whisper's translate task turns Japanese speech into English

Whisper can translate Japanese speech into English in one step, without a separate Japanese transcript. The Whisper README gives the command as whisper japanese.wav --model medium --language Japanese --task translate ⁴. The --language flag names the spoken language, and --task translate asks for English output. Leave out the task flag and the same command writes Japanese text instead ¹. The question on whether Whisper can translate audio explains why the translate task only produces English.

Install steps are in our guide to translating audio to English. Once Whisper is installed:

1. Pick the audio or video file you want to translate.
2. Run the command above, with your file name in place of japanese.wav.
3. Open the text output and read it against the recording, paying attention to names, numbers and dates.

Use medium or large, not turbo, for the translation

The model you pick decides whether you get English at all. The README warns that the turbo model returns the original language even when the translate task is set ⁵. It recommends medium or large for the best translation results ⁶. If your output is still Japanese, check the model name before you blame the recording.

Japanese translation scores below German

Whisper's translation quality depends on the language it hears. On the Fleurs read-speech test, the large-v2 model scored 18.9 BLEU when translating Japanese into English ². The same model scored 34.6 BLEU for German ³. BLEU compares a machine translation with a human one, and a higher score is closer. It counts shared words, so it is a rough guide to meaning, not a grade for your file.

Part of the gap may start in recognition. Japanese is written without spaces between words, so the Whisper paper scores its transcripts by character rather than by word ⁷. A character error rate counts wrong characters the way a word error rate counts wrong words. On the third evaluation set of the Corpus of Spontaneous Japanese, a 2025 comparison put Whisper large-v3 at 16.5 percent character error rate ⁸. That corpus holds spontaneous speech, with fillers and hesitations, which is closer to a real recording than read sentences are.

A recognition error becomes a translation error. If the Japanese transcript misreads a word, the English will carry that mistake forward, and speech translation inherits the errors of the speech recognizer that feeds it.

What to do with a Japanese recording that matters

1. Make a Japanese transcript first. Run the same command with --language Japanese and without --task translate, then check the transcript against the audio.
2. Translate the checked text, or run the translate task on the file, and compare the English with the Japanese for names, numbers and dates.
3. Have a Japanese speaker check any passage you will quote, sign or publish. Machine output can read smoothly and still be wrong, so smooth is not the same as correct.

If you want a hosted service to produce the transcript, Google Cloud's Speech-to-Text lists Japanese for its chirp_2 model ⁹. That page covers recognition only, so the translation is a separate step.

If the recording is private, keep it on your own computer. Our guide to running Whisper locally covers the setup, and the guide to Whisper and privacy explains what hosted services do with audio.

Where this answer does not apply

The steps above assume a file you can run on a computer. Some Japanese audio is better handled another way.

• Recordings on a phone. A voice memo can be transcribed on the phone, and the transcript can then be translated as text. The question on translating a voice note to text covers those steps.
• Videos for subtitles. A subtitle file has its own timing and line rules. The question on translating a video into English subtitles covers that workflow.
• Other languages. The numbers above describe Japanese only. Whisper's accuracy varies widely by language ¹⁰, so test any language on your own recordings. The question on translating Zulu or Xhosa audio covers a lower-resource case.

For more on the tools that handle Japanese, see our guide to Japanese, Chinese and Korean speech to text and the question on the best speech to text for Japanese. The overview of languages and accents puts this question in context.