Skip to main content
Transcription is working, but what lands isn’t what you said. Names come out mangled, technical terms are guessed at, or the whole thing is in the wrong language. These have different fixes, so it’s worth knowing which one you have.

​
Names, jargon and acronyms

If it’s specific words that are wrong — people’s names, product names, medical or legal terms — the custom dictionary is the fix, and it’s the highest-value change you can make. Add the words you use in Dictionary in the sidebar, on the Dictionary tab. Type them into the box at the top — you can paste several at once, separated by commas — and press Add. They’re then supplied to transcription so it knows to expect them. Full guide: Custom dictionary.
Add the spelling you want, not the one you’re getting. The dictionary tells transcription what the right answer looks like.

​
The wrong language

Check the language set for the mode you’re using. The engine and its language are configured per mode — dictation, note recording and audio upload each have their own — so changing it in one place doesn’t change the others. Settings → Speech-to-Text under AI Models, then the tab for the mode you’re using: Dictation, Note Recording or Audio Upload.
Auto-detect works from the audio, so a short or quiet take gives it very little to go on. If you consistently work in one language, setting it explicitly is more reliable than auto-detect.

​
Accuracy that used to be better

Work through these in order — the first two account for most of it.
1

Check which engine is running for this mode

Local and cloud models differ in accuracy, and so do the local model sizes. If you switched to a smaller local model, or fell back to one, that alone explains a drop. Settings → Speech-to-Text under AI Models, then your mode’s tab.Local models are covered in Local models; the trade-offs between local and cloud are in Cloud vs local.
2

Check your microphone and your distance from it

Accuracy is dominated by input quality. A headset microphone near your mouth outperforms a laptop or monitor microphone across the room by a wide margin, and a device change you didn’t intend is common — see My microphone isn’t working.
3

Rule out AI processing

If the transcription is right but the final text isn’t, something rewrote it afterwards. Compare with View raw transcript in your history, and see It answers me instead of typing.
4

Reduce background noise

Other voices, music and typing all degrade transcription. If you can’t change the environment, get closer to the microphone.

​
If a whole passage is missing

A very long dictation can display truncated in the app even though the full text was saved. Open the history item and use View raw transcript or Copy raw transcript — that gives you everything, and it’s the reliable route for long recordings.

​
FAQ

Larger models are more accurate and slower; smaller ones are faster and less accurate. Local models sets out the options and what each needs from your machine.
That’s exactly what the custom dictionary is for. Add it there rather than changing model.
You can re-run it — each history item offers Re-transcribe, which is useful after you’ve added dictionary words or changed model. See I lost a dictation.
Adding words to your custom dictionary is the reliable way to teach it. Everything about how that works is in Custom dictionary.