If your mental model of speech recognition software is Dragon, with a headset, a training session and a licence key, it’s out of date by about three years. Here’s what changed, what it means for a Mac, and which software is worth using now.
What changed
Older speech recognition was built per language, per accent, from carefully labelled audio, and it learned your voice over weeks. The new generation, starting with OpenAI’s Whisper and continuing with the models from Deepgram, AssemblyAI, Google and others, is trained on hundreds of thousands of hours of real audio. It understands most voices, most accents and most background noise out of the box. Nobody trains it to their voice any more.
The second change is what happens after recognition. The words now go through a language model that adds punctuation, removes ums and restarts, fixes grammar, and can restructure the text for where it’s going. That step is where the useful software differentiates.
Cloud or local
Cloud models (what Speaky Box and Wispr Flow use) are the most accurate, especially on names, accents and technical terms, and they return text in about a second. Local models (Superwhisper, MacWhisper) run on your Mac, work offline, and keep everything on the machine, at some cost in accuracy and setup. Both are good. Pick by whether offline and local matter to you.
Accuracy in practice
On clear speech, all the modern options are in the high nineties. The differences show on the hard cases: a name the model has never seen, a technical term, someone talking fast with an accent. Two things fix that: a vocabulary list you control, and the cleanup model, which uses context to pick the right word. Software that has both will feel far more accurate than software with a better raw model and neither.
The software
- Speaky Box: cloud model, cleanup model, vocabulary, hold-to-talk, app-aware writing in AI apps. Free for 2,000 words a week.
- Wispr Flow: cloud model, cleanup, cross-platform. Comparison.
- Superwhisper: local Whisper-family models, configurable. Comparison.
- macOS Dictation: Apple’s on-device model, no cleanup. Comparison.
- Dragon: no longer available for Mac. Alternatives.
What to expect
Install, allow the microphone, talk. No training, no headset, no profile. If it gets a name wrong, add the name. That’s speech recognition on a Mac in 2026.