Troubleshooting

1. Check Dashboard ▸ Settings ▸ Permissions. Grant anything not showing as granted.
2. If macOS shows the prompt but nothing changes, open System Settings ▸ Privacy & Security ▸ Input Monitoring, click +, and add OpenVoiceFlow from Applications.
3. Check the app is not paused — the menu-bar menu shows this, and pausing lasts an hour.
4. Check another app has not claimed the key globally (Raycast, Alfred, Karabiner). Try a different hotkey to confirm.
5. If you moved the app after granting permissions, macOS may have invalidated the grant. Put it back in /Applications and re-grant.

1. Your text is not lost — it is on the clipboard. Press ⌘V to place it now.
2. Grant Accessibility in System Settings ▸ Privacy & Security ▸ Accessibility, adding the app with + if it is absent.
3. Check Paste automatically is on in Settings ▸ Dictation.
4. Confirm your cursor is in a text field that accepts a paste — some apps intercept ⌘V themselves.

1. Check Microphone permission in Settings ▸ Permissions.
2. Check System Settings ▸ Sound ▸ Input — the right device selected, and its level moving when you speak.
3. Disconnect and reconnect a Bluetooth headset; they occasionally attach with a broken input profile.
4. Quit other apps that may be holding the microphone exclusively.

1. Microphone position. Sit at normal working distance or use a headset. This fixes most cases.
2. Speak in phrases, not word by word — Whisper needs context to disambiguate.
3. Add problem words to your dictionary (requires cleanup on).
4. Move up a model size in Settings ▸ Transcription.
5. For non-English, make sure you are on a large enough model — see Languages.

Full detail: Improving accuracy.

1. Turn on cleanup in Settings ▸ AI cleanup and pick a backend.
2. If cleanup is already on, verify your API key and provider balance — cleanup fails open, so an expired key looks exactly like raw transcripts.
3. Want the dictionary without sending text anywhere? Use local Ollama.

1. Use Try again — the app clears the partial download first, which resolves most cases.
2. Open Details on the error for the actual reason rather than guessing.
3. Check free disk space: Large turbo needs about 1.6 GB plus room to compile.
4. A corporate VPN or proxy can block the model host. Try another network.
5. Manual reset: quit the app, delete ~/Documents/huggingface/models/argmaxinc/whisperkit-coreml/, relaunch, re-select the model.

• Unchanged output: cleanup fails open. A missing/expired key, no balance, or a provider outage all yield the raw transcript. Check your provider dashboard.
• Slow with a cloud provider: requests time out at 10 seconds and fall back to raw text, so this shows up as inconsistency rather than a hang. Try a smaller model or a faster provider.
• Slow with Ollama: local generation is bounded by your Mac. Use a smaller model — a 3B-class instruct model is plenty for cleanup.
• Ollama not responding: make sure Ollama is actually running and listening on localhost:11434.

• Hold the key only while actually speaking, and start speaking promptly.
• Check your microphone is picking you up — a flat waveform makes this far more likely.
• A larger model reduces but does not eliminate it.

• Drop to Small in Settings ▸ Transcription, especially on 8 GB Macs.
• Turn off Show words as you speak, which re-transcribes while you hold the key.
• On Intel Macs, prefer Small or Tiny — there is no Neural Engine to offload to.

• Quit another menu-bar app to make room, or use a menu-bar manager.
• Open the dashboard from the Dock icon instead, or turn Show in Dock back on in Settings.