Preprint
Speech Recognition for Automatically Assessing Afrikaans and isiXhosa Preschool Oral Narratives
Christiaan Jacobs et al., arXiv (Stellenbosch University and University of Edinburgh), 2025
Read the originalRead on October 17, 2026
- What it is
- A study that fine-tuned OpenAI's Whisper on five minutes of transcribed preschool child speech in Afrikaans and isiXhosa, and measured the word error rate in each language. It reports that Whisper's training data has no isiXhosa.
- How much weight it carries
- A preprint by university researchers, not yet peer-reviewed in the version we read. The results are for young children's speech after fine-tuning on a few minutes of audio, so they do not describe Whisper on adult isiXhosa.
Facts taken from this source
- After fine-tuning Whisper medium on the same amount of transcribed child speech, the word error rate was 80.4 percent for isiXhosa, against 47.4 percent for Afrikaans.
“the WER for Afrikaans is significantly better than for isiXhosa: 47.4% compared to 80.4% for Whisper medium” (Results section, first row of Table II)
- The authors state that Whisper's training data includes no isiXhosa and no closely related languages, so they chose Swahili as the language token when they fine-tuned on isiXhosa.
“On the other hand, no isiXhosa or any closely related languages are included in Whisper.” (Section III, base speech recognition model)
Pages that rest on this source
- Can AI transcribe Zulu, Xhosa and other African languages?Question
- Can you translate Zulu or Xhosa audio to English?Question
- Dictation Languages: What Each App and Device SupportsGuide
- How to Translate Audio and Video to English TextGuide
- Is speech recognition less accurate in languages other than English?Question