Troubleshooting
“It didn’t type anything”
Section titled ““It didn’t type anything””Work through these in order:
- Is Voice Mode running? Check the menu bar for the Voice Mode icon. If it’s not there, launch the app.
- Did the HUD appear when the hotkey was held? If not, the hotkey didn’t register. This is almost always a permissions issue — open System Settings → Privacy & Security and confirm Voice Mode is enabled under both Accessibility and Input Monitoring.
- Was the user speaking into the right microphone? Check the input device in Voice Mode’s Settings → General → Microphone. If the selected device is muted or the wrong one, VAD will trim everything as silence.
- Did the app receive the paste? The paste targets whichever app has focus at release time. If focus shifted mid-sentence, the paste followed the new focus.
”The correction is too aggressive / changing things I didn’t say”
Section titled “”The correction is too aggressive / changing things I didn’t say””AI Correction uses a small on-device LLM, which will sometimes over-edit.
The simplest fix is to turn it off in the Dashboard’s AI Writing pane —
dictation still works and the text is the raw STT output plus the
Dictionary.
For specific words that get consistently mis-rewritten, adding a
Dictionary entry (see replacements.md) is more reliable since the
Dictionary runs before correction.
”It keeps mishearing a specific word”
Section titled “”It keeps mishearing a specific word””Add a Dictionary entry for that exact spoken form. The Dictionary is deterministic and runs before correction, so it always wins.
”Assistant says I need Pro”
Section titled “”Assistant says I need Pro””The assistant (Claude/Codex/Cursor/OpenCode/Ollama/MLX), TTS, character voices, personas, and augments are Pro-only. Dictation, the Dictionary, and AI Correction are always free. Open Settings → License to activate or manage a subscription.
”First launch is really slow”
Section titled “”First launch is really slow””The on-device STT model has to compile the first time it runs, which can take anywhere from a few seconds to a minute depending on the model selected. Subsequent launches reuse the compiled model and are fast.