Improving accuracy
Ordered by effect, largest first. Most accuracy complaints are solved by the first two items, and neither of them involves changing a setting.
1. Fix the microphone situation
This is the biggest single factor and the most overlooked. A MacBook's built-in microphone is decent at conversational distance and poor across a room. If you are leaning back from an open laptop, the mic is largely hearing your ceiling.
- Sit at a normal working distance, or use a headset — AirPods or any wired headset mic beat a built-in array at distance.
- Reduce steady background noise where you can: fans, air conditioning, a nearby TV.
- Check that macOS is using the input you think it is, in System Settings ▸ Sound ▸ Input. Bluetooth headsets sometimes select a low-quality input profile.
2. Speak in phrases
Whisper resolves ambiguity from surrounding words. Delivering a complete clause at a natural pace gives it that context; word-by-word delivery removes it and makes the model less accurate, not more.
- Speak as if to a colleague, not as if to a machine.
- Do not over-enunciate — unnatural stress patterns hurt.
- Hold the key a beat before you start and a beat after you finish, so the first and last words are not clipped.
3. Teach it your vocabulary
Proper nouns are not a model-size problem — a name absent from training data stays absent at every size. The personal dictionary is the durable fix for colleagues' names, product names, and internal jargon.
The dictionary and the Know-Me profile are applied during AI cleanup. With cleanup set to None, they have no effect on your output. If you rely on them, cleanup needs to be on.
4. Move up a model size
If the first three did not get you there, try the next model up in Settings ▸ Transcription. The jump matters most for accented speech, noisy rooms, and non-English languages. On Apple Silicon, Large turbo is usually worth it.
5. Turn on cleanup for prose
Raw transcripts contain your filler words and false starts because that is what you said. If you are dictating email or documents rather than notes, AI cleanup removes the “um”s, fixes grammar, and handles spoken corrections like “no wait, I mean”. It also activates your dictionary and profile.
What accuracy to expect
We do not publish a benchmark, because word error rate depends on your microphone, accent, vocabulary, and room far more than on anything we could measure in a lab. The honest test is the one you can run in a minute: dictate a real paragraph of your own work and judge the result. If it disappoints, work down this page in order — the fix is nearly always here.
Documentation for OpenVoiceFlow 0.5.5 · last updated 2026-07-31. Something wrong or missing here? That is a bug — tell us at shimoverse@gmail.com.