Skip to main content
You’ve chosen a local model and transcription fails, falls back to the cloud, or produces nothing.
Check this first. The transcription engine is set per mode. Dictation, note recording and audio upload each have their own — so switching one to local leaves the others exactly where they were.Settings → Speech-to-Text under AI Models, then the tab for the mode you’re actually using: Dictation, Note Recording or Audio Upload.If you changed the setting and nothing changed, this is almost always why.

​
Quick checks

1

Confirm a model is downloaded

Local transcription needs the model on disk. The model picker shows which are installed. If a download failed, see A model won’t download.
2

Check the mode you're using is set to local

As above — the right tab, not just the right section.
3

Try a smaller model

Large models need substantial memory. If a small one works and a large one doesn’t, that’s the constraint, and the smaller model is a real answer rather than a workaround.
4

Check bundled components weren't quarantined

Security software removing whisper.cpp, sherpa-onnx or FFmpeg breaks local transcription specifically while leaving everything else working. See Antivirus blocks OpenWhispr.

​
GPU acceleration

If GPU acceleration fails — an unsupported card, or not enough video memory — OpenWhispr restarts on the CPU, retries the same request, and tells you it’s using CPU instead. Your dictation still completes; it’s slower. If that happens every time, turn GPU acceleration off from the GPU card in the transcription model picker. It removes the failed attempt and the delay it adds.

​
What “local” does and doesn’t need

  • No network once the model is downloaded — as long as cloud fallback (below) is off, which is the default.
  • No plan allowance — local transcription doesn’t count against your weekly words on any plan.
  • Real CPU or GPU work — so it’s slower than cloud on modest hardware, and a laptop on battery may throttle it.
Cloud vs local sets out the trade-offs; Local models covers the models themselves.

​
FAQ

That only happens if cloud fallback has been enabled — it defaults to off, so out of the box a failed local attempt just fails rather than silently going to the cloud. If you’re seeing this, a debug log will show what’s configured and why the local attempt failed — see how to send one.
Expected on most machines. A smaller model is the usual answer if speed matters more than the last few points of accuracy.
It depends on the model — Whisper and Parakeet support different sets. Local models has the detail.
Local transcription runs on your machine and the audio doesn’t leave it — as long as cloud fallback is off, which is the default for every install. If it’s been enabled, a failed local attempt sends that audio to the cloud instead of failing. Cloud vs local sets out what each mode sends.