Troubleshooting

Open Settings → Model and download one. Nothing works without it.

Large v3 Turbo on a discrete GPU, Small on integrated graphics or CPU. See Speech models.

Sotto asks your desktop for the shortcut through the GlobalShortcuts portal, which Plasma 5.25 and later provide. Some desktops do not have it.

Fix: disable the shortcut in Settings → Shortcut, then bind sotto --toggle in your desktop’s own shortcut settings — System Settings → Shortcuts → Custom on KDE, or your compositor’s config file elsewhere. It behaves identically.

Also check the binding has not been claimed by something else, in your desktop’s shortcut list.

In order:

1. Install wl-clipboard. Nothing clipboard-based works without it.
2. Set up ydotool, or accept the one-time RemoteDesktop permission prompt when it appears. One of the two is needed to synthesise the paste keystroke.
3. If the target is a terminal, the default Ctrl V will not paste — terminals use Ctrl Shift V. Switch to Type it or Clipboard only in Settings → Output.

In the worst case, the text is always left on the clipboard, so it is never lost.

With LayerShellQt installed and the screen setting on auto, your compositor places the overlay on the active output — and it is re-evaluated every time it appears.

Without LayerShellQt, the fallback window uses the screen under the mouse pointer instead.

Fix: install layer-shell-qt, or pin a specific monitor in Settings → Appearance.

Look at the backend named at the bottom of the Settings window.

• It says cpu — that is the problem. Rebuild with -DSOTTO_GPU=vulkan (or hip for ROCm, cuda for NVIDIA). See Installing.
• It names a GPU backend — try a smaller model. Large v3 is significantly heavier than Large v3 Turbo for a modest accuracy gain.

If your GPU is not officially supported by ROCm, set HSA_OVERRIDE_GFX_VERSION before running — 10.3.0 covers many RDNA2 cards.

If that does not help, build with -DSOTTO_GPU=vulkan instead. On AMD, Vulkan is often within a few percent of ROCm for this workload and far less sensitive to driver versions.

In rough order of impact:

1. The microphone. A headset beats a laptop microphone across a room by more than any model upgrade will.
2. A bigger model. Small → Large v3 Turbo is a real step up.
3. Pin the language in Settings → Model instead of leaving it on auto. Detection can guess wrong on a short first utterance.
4. Speak in complete sentences. Whisper uses sentence context; fragments give it nothing to work with.

You are probably saying a phrase that contains “new line” or “new paragraph” — “a new line of products” is the classic. Turn voice commands off in Settings → Formatting and use pauses for paragraphing instead.

If paragraphs are breaking where you did not pause deliberately, raise the pause threshold in the same place.

Sotto keeps a short rolling buffer specifically so the first words are not clipped, so this is unusual. Check that the right input device is selected in Settings → Audio, and that it is not being brought up by a noise-suppression layer that gates the first fraction of a second.

That is correct behaviour. A running Sotto owns its session-bus name, so a second invocation forwards its command to the first and exits — it does not start a second process. See Notepad and CLI.