For years dictation was a tool for people with a particular kind of English. Everyone else got a transcript full of guesses. That’s mostly over, and it’s worth knowing why, because it changes which tools to pick.
What changed
Old dictation models were small and trained on narrow, clean speech. Large speech models are trained on enormous amounts of real-world audio, in every accent, at every speed, with background noise. Scottish, Nigerian, Indian, Australian, Glaswegian at pace: these models have heard it all, and the gap between accents has mostly closed.
The catch: the big models run in the cloud. On-device dictation, including Apple’s, still uses smaller models and still shows the gap. If accuracy on your accent matters, a cloud-backed tool like Speaky Box or Wispr Flow will be noticeably better than built-in dictation.
What still goes wrong
Names and jargon. No model knows how your company spells its product or that your colleague is Siobhán. The fix is vocabulary: add the words once and the model is told to prefer those spellings. Speaky Box’s Words list does this for both keys, in every app. It takes five minutes and it’s the single biggest accuracy improvement for anyone, accent or not.
Small things that help
- Speak at your normal pace. Slowing down and over-enunciating makes it worse, not better.
- A decent microphone matters more than the accent. Laptop mics are fine at a desk; a headset is better in a café.
- Pause where the punctuation goes. The model uses your rhythm.
Try it on your own voice
Every tool has a free tier. Dictate the same three sentences into each, including two names and one technical term, and compare. It takes five minutes and it’s the only test that matters for your voice.