Can you translate Zulu or Xhosa audio to English?

Yes, but plan for weaker output and more checking than you would need for Japanese or German. The practical route is two steps. Get a transcript in Zulu or Xhosa from a tool that supports that language, check it against the recording, then translate the text into English. Whisper's own translate task is a weak choice for Xhosa.

Whisper's translate task is a weak choice for Xhosa

Whisper was not trained on isiXhosa at all. The authors of a 2025 study on isiXhosa child speech say that Whisper's training data includes no isiXhosa and no closely related languages ¹. In their tests, a Whisper medium model fine-tuned on the same amount of transcribed child speech reached an 80.4 percent word error rate for isiXhosa, against 47.4 percent for Afrikaans ².

That result is for transcription, on child speech, so your recording may differ. Recognition that poor is still a warning sign for any output made from the same audio, including a translation. The question on whether Whisper can translate audio explains what the task does for other languages.

That study covers isiXhosa and Afrikaans, so it says nothing about Zulu. Test Zulu recordings yourself before you trust the output.

Zulu and Xhosa have weak support for translation into English

SeamlessM4T's language table marks Zulu as a low-resource language for speech translation into English ³. It marks Xhosa as zero-shot for the same task ⁴. In that paper, a language is zero-shot when it lacks the primary training data for the task ⁵.

Both labels describe the training data, not the quality of the output. They tell you the model had little to learn from, so expect more errors than you would get for Spanish or German. The table gives no translation score for either language, and in the sources we checked we did not find one for Zulu or Xhosa speech.

Transcribe in Zulu or Xhosa first, then translate the text

Google Cloud's Speech-to-Text supports both languages for recognition. Its supported-languages table lists Zulu (South Africa) as zu-ZA ⁶ and Xhosa (South Africa) as xh-ZA ⁷, both for the chirp_2 model. That gives you a transcript in the original language, which you can check.

For the translation, Google's Cloud Translation page lists Xhosa and Zulu in its Translation LLM languages. It says translations from any language to any language in that list are supported ⁸. The page lists which languages are supported. It does not measure how good the output is.

Work through the steps in this order:

1. Set the spoken language to Zulu or Xhosa, not English, in the tool you use for the transcript. A wrong language setting gives output that looks like a translation and is not one.
2. Check the transcript against the audio, and mark every word you cannot confirm. Start with names, places, numbers and dates, since they are the easiest to check and the costliest to get wrong.
3. Translate only the checked text, and compare the English with the transcript for meaning, not just for smooth wording.
4. Have a Zulu or Xhosa speaker read the English against the original for anything that matters.

It depends on the speaker, the audio and the stakes

Several things change the answer for a given recording.

• The speaker. Accent and dialect change accuracy within one language. The guide to speech recognition and accents covers how that works.
• Mixed languages. If the speaker switches between Zulu or Xhosa and English, expect errors at each switch. The glossary entry on code-switching explains the term.
• The stakes. For evidence, a medical history, a contract or anything legal, use a qualified human translator. Machine output alone is not enough.
• Privacy. Both steps above can run on a hosted service, and the audio then leaves your device. Read the provider's policy before you send a confidential recording, and see our guide to Whisper and privacy for how hosted services handle audio.

Where to go next

For the general method, see our guide to translating audio and video to English. Japanese works differently, so the question on translating Japanese audio covers that case. A phone recording has its own steps in the question on translating a voice note to text, and a video with subtitles is covered in the question on translating a video into English subtitles. The overview of languages and accents puts all of these in one place.