Does dictation work with Nigerian and Kenyan English accents?

Yes, but how well depends on the engine. In our October 2026 test, Parakeet TDT v3 had a word error rate of 13.44 percent on Nigerian English ¹ and 12.17 percent on Kenyan English ². Apple SpeechTranscriber had 20.20 percent ³ and 17.29 percent ⁴. Whisper large-v3-turbo had 22.03 percent ⁵ and 19.21 percent ⁶. Lower is better.

Parakeet barely changed; Apple and Whisper got clearly worse

The baseline is Mainstream US English: 12.32 percent for Parakeet ⁷, 15.45 percent for Apple ⁸ and 16.93 percent for Whisper ⁹.

• Parakeet: Kenyan English was marginally below US English, and Nigerian English was higher by a modest margin.
• Apple: both groups were higher than US English, and Nigerian English was far above it.
• Whisper: both groups were higher than US English, and Nigerian English was again far above it.

In every engine, Nigerian English had more errors than Kenyan English. Nigerian English also had the highest rate of any variety of English for Apple and for Whisper, ahead of Jamaican English (19.49 percent for Apple ¹⁰, 20.44 percent for Whisper ¹¹).

Choosing the engine mattered more than the accent

Parakeet's Nigerian English figure (13.44 percent ¹) was lower than the US English figures of both Apple (15.45 percent ⁸) and Whisper (16.93 percent ⁹). In other words, a Nigerian English speaker using Parakeet got fewer errors than a Mainstream US English speaker using either of the other two. The gap between engines was larger than the gap between accents.

This fits the wider research on EdAcc, where the authors found that every English recognition model they tested dropped in performance on speakers of Indian, Jamaican and Nigerian English ¹². Our test shows the same direction for Nigerian English with Apple and Whisper, and a much smaller effect with Parakeet.

What the numbers do and do not measure

• Two speakers for Kenyan English. Kenyan English rests on 2 speakers and 100 utterances. Nigerian English has 5 speakers and 99 utterances, and Mainstream US English has 9 speakers. A single voice or microphone can move a small group's figure, so a small difference, such as Parakeet's Kenyan result against US English, is not a finding. Compare engines within a group more than groups with each other.
• Not all of Africa. These are two varieties of English. They say nothing about other African accents or about languages spoken in Nigeria and Kenya.
• Conversation, not dictation. The audio is informal talk over video calls, selected for at least five words, 1.5 to 30 seconds and no overlapping speech. Dictating alone to a microphone is a different task, so treat these rates as a comparison, not a forecast.
• Digits count as errors. Numbers spoken as words in the reference count as errors when an engine writes digits.
• Three engines on one Mac. We tested an M2 Pro. Cloud services are not in the test, and newer versions can change the results.

What this means for you

If you speak Nigerian or Kenyan English, dictation is usable, but do not accept the default blindly. Try the same paragraph in two engines. On this evidence Parakeet had the lowest error rate in both groups, and Apple and Whisper made more errors, especially on Nigerian English. Plan to proofread, because even the best result here is not error free. Our own app, Yap, runs on-device dictation on Mac and Windows; see /market for how tools compare.

Related reading: the full dataset, the noise robustness data, the on-device benchmark, dictation across languages and the guide to speech recognition and accents. For the full ranking see which speech recognition handles accents best.