Troubleshooting
Nothing appeared
Open Home first. If the dictation is there, transcription finished but delivery to the target application did not.
Keep the original window focused until the text arrives. OpenVoice stops delivery if that window loses focus or you cancel, rather than typing into a different window. An application running with administrator privileges can also block input from OpenVoice.
When delivery is held, the transcript stays in OpenVoice. Click Copy in Home, then press Ctrl+V where you want it. Recovery does not copy automatically.
Normal paste briefly borrows the clipboard, then restores its previous contents unless you copied something new. Version 1.3.3 fixes a regression where a copied screenshot could prevent dictation from pasting. Clipboard formats that cannot be safely preserved use paced typing instead.
The next hotkey press does nothing
Update to 1.3.3 if this happens after cancelling hands-free dictation. It fixes the missed next press and recording-state problems after rapid cancel/restart, microphone failures, or reaching the recording limit. Starting hands-free mode also preserves earlier dictations that are still processing.
The wrong window got the text
Start speaking with the target window and text field focused. Current versions check the original window before sending text, but you should also keep the intended text field selected within that window.
The Flow Bar is designed not to take focus from the application you are dictating into.
It says transcription failed
There are two different failures behind that phrasing, and they are now reported separately. A capture failure means the audio never reached the transcriber, which usually means a muted or disconnected microphone. An ASR failure means the speech engine itself failed. If it is capture, check your input device and that it is not muted.
If Home offers Retry transcription, retained audio is available in memory for up to ten minutes. Retrying shows the recovered text in Home; it does not copy or paste automatically. Use Copy when you need it. Retained audio is lost when OpenVoice closes.
A model will not download
Model downloading was completely broken before v0.2.0. No model could be fetched by an installed copy, including the very first download on a fresh machine, which made a new install unusable unless it happened to find weights already cached. Offline mode was read once by the download library at startup and could not be lifted afterwards, so every fetch failed claiming outgoing traffic was disabled.
If you are on an older build, upgrade. If you are on v0.2.0 or later and a download still fails, check that Settings → Updates and any corporate proxy are not blocking the request.
The Speech model screen says I have nothing
Fixed in v0.2.0. It always looked under %APPDATA%\OpenVoice\models, which is only where weights live for a standard install. If OPENVOICE_PYTHON or HF_HOME is set, the app uses a different cache, and the screen now asks the running engine which one it is actually using.
A model reads "Selected — not downloaded"
That is correct and not an error. You have chosen a model whose weights are not on disk yet. The Download button next to it will fetch them.
The update check failed
Nothing is wrong if no release has been published yet, or if you are offline. The message says so while keeping the underlying detail. You can turn update checks off entirely in Settings → Updates, and then no request is made at all.
Dictation cuts off early
Check Maximum recording in Settings. Note that this setting genuinely did nothing before v0.2.0: the whole [limits] section was written to disk and then ignored, because the engine started with built-in defaults instead of your settings. Choosing 5 minutes still cut off at 2.
Very short recordings are discarded on purpose. Anything under 300 ms is treated as a fat-finger press and dropped silently, because a toast for something you did not mean to do is noise.
The output has repeated characters or looping phrases
That is a known decoder artefact and there is a rule that collapses it. Any character repeated more than three times is squashed, and a phrase repeated more than three times in a row is trimmed. Legitimate repetition survives: "no no no" and "that is very very good" are left alone.
If you are seeing it constantly, a larger model usually helps.
Something got redacted that should not have been
Secrets matching privacy.redact_patterns are replaced with [redacted] before a transcript is saved or logged. The defaults cover OpenAI, GitHub and AWS credential shapes. The text delivered to your application is never altered; only the stored copy is.
Reporting something
Include the stage-by-stage trace if it is a formatting problem. It is the difference between a report that can be fixed in ten seconds and one nobody can act on.
ov format "whatever you said" --profile terminal --trace
Then open an issue on GitHub. For anything security-shaped, SECURITY.md has the right route.