Can AI transcribe Zulu, Xhosa and other African languages?

AI can attempt Zulu and Xhosa, but the tool decides whether it works at all, and the published evidence on how well it works is thin. Some speech services list both languages. The best-known open model, Whisper, does not. For the wider picture, see our overview of transcription and the guide to transcribing audio in other languages.

Google Cloud and Meta list Zulu and Xhosa, and Whisper does not

Google Cloud's Speech-to-Text supported-languages table lists Zulu (South Africa) and Xhosa (South Africa) for its chirp_2 model ^{1 2}. That is an API for developers, not an app you use directly.

Meta's Omnilingual ASR is a speech recognition model built to cover a very large set of languages. Its supported-language list includes the code for Zulu, zul_Latn ³, and the code for Xhosa, xho_Latn ⁴. Being on the list means the model accepts the language. It does not tell you how accurate the output is.

Whisper's language table includes Shona, Yoruba, Somali and Afrikaans ⁵, but Zulu and Xhosa do not appear in it. Researchers who fine-tuned Whisper on isiXhosa wrote that no isiXhosa or closely related languages are included in Whisper's training data ⁶.

Measured accuracy for African languages is uneven, and the gap is large

Whisper's measured results show how far accuracy can fall for an African language. On the FLEURS read-speech test, Whisper large-v2 had a word error rate of 39.3 percent for Swahili ⁷. Swahili is one of the African languages in Whisper's table, and it still had nearly two errors in every five words. For isiXhosa, the clearest measurement we found is from a 2025 study of preschool children's stories. The researchers fine-tuned Whisper medium on the same amount of transcribed child speech in each language. The word error rate was 80.4 percent for isiXhosa, against 47.4 percent for Afrikaans ⁸. This is child speech after fine-tuning, so it shows the size of the gap, not an accuracy figure for adult speech.

We did not find an independent, current published word error rate for adult Zulu or Xhosa speech, so this page gives none. Vendor pages may publish figures, but a vendor's own test is not an independent check.

A fluent-looking transcript can still be wrong, especially in a language the tool handles poorly. Read the whole output, not only the parts that look right. For the category, see low-resource language.

Test the tool on your own recordings before you rely on it

1. Collect three to five recordings that match your real audio, including the hardest ones: fast speech, background noise, and any switching between languages.
2. Transcribe each one with the language set to Zulu or Xhosa. In Google Cloud, choose zu-ZA or xh-ZA. In Omnilingual, use the codes zul_Latn or xho_Latn.
3. Correct one transcript by hand, word by word, to make a reference.
4. Count the wrong words against that reference to get the word error rate for your audio. For the score, see word error rate.
5. Decide what a wrong word would cost. If an error could change the meaning, plan a human transcriber for that work. Human and AI transcription compared sets out the trade-offs.

If the recording is a voice note on a phone, the guide to voice memos and voice notes covers the recording side. Whichever tool you use, the test above is the most direct check of its accuracy for your language.

For related languages, see how to transcribe Hindi audio and how to transcribe Japanese, Chinese or Korean audio.