General

A four-way switcher: Push-to-Talk Toggle Listen Wake Word. Push-to-talk holds to record; toggle starts/stops on a tap; listen transcribes system audio from a loopback device; wake word arms a keyword detector. Switching mode shows or hides the mode-specific controls below. See Recording modes for when to use each.

Shown only in Toggle mode. When on, a toggle session runs continuously from the first press to the second — silence-based auto-stop and silence-timing punctuation tuning are both disabled. Use it for long-form dictation where soft pauses were cutting you off.

Shown only in Listen mode. Picks which WASAPI loopback (system-audio) device is transcribed. null uses the default render device. This lets WinSTT caption a call, video, or anything playing through your speakers.

A keyword that arms recording when spoken. The list unifies 18 free keywords across two engines — a 2x badge means both Porcupine (PVP) and openWakeWord (OWW) agree (highest accuracy); single-engine keywords carry just the PVP or OWW badge. The detector backend is chosen automatically from the keyword you select.

A 0.00–1.00 slider in 0.05 steps (21 positions). Lower is stricter — fewer false triggers, may miss soft pronunciations. Higher is more permissive. 0.60 is a sensible middle for most voices.

Seconds the gate stays armed after a detection (1–30). If you say the wake word but never follow up, the engine returns to listening after this window — protection against stray noise triggering a long-tail recording.

Shown only in Listen mode. Colors each speaker in the transcript and tracks identities across the session. First use downloads ~32 MB of ONNX models. Toggles live — no restart.

Now lives on the Output tab. A 6-step slider — Off, 20%, 40%, 60%, 80%, Mute — that ducks your speaker volume while you dictate so playback doesn't bleed into the mic. Intermediate values drop to (100 − value)% of the previous level; the original volume is restored when recording stops.

Hidden in Listen mode. Plays a short chime when recording starts and stops. Turning it on reveals the Sound Library below.

The interface language, independent of the language you dictate in. Six locales are fully translated — Arabic (العربية), English, Spanish (Español), French (Français), Hindi (हिन्दी), and Chinese (中文). Arabic renders right-to-left. (Additional seeded locales appear in the list but still display English until their translation passes land.)

A five-way switcher: Bar, Grid, Radial, Wave, and Aura. The gallery below shows each reacting to speech.

Shown only when the type is Bar. The number of bars, 3–21 in odd steps (3, 5, 7 … 21). More bars read denser; fewer read calmer.

A 6-step slider — Off · XS · S · M · L · XL — that both turns the floating overlay on and sets its size. Position 0 (Off) hides the overlay and reverts overlay-only live-display choices to in-app; XS–XL pick the visualizer height. Greyed out in Listen mode (the overlay never shows there). Size is stored as general.visualizerSize.

Greyed out when the overlay is off. Two layouts (shown below): Floating bottom — a two-piece pill near the bottom of the primary display; and Dynamic island — a morphing capsule docked to the top-center that grows as the live preview fills in.

Where the live preview renders, as two checkboxes — In app (the main-window feed) and In pill (inside the overlay). The combination maps to none, in-app, in-pill, or both. The "In pill" option is disabled while the overlay is off.

Launch WinSTT automatically when you sign in.

Start hidden in the system tray instead of opening the main window.

Closing the window keeps WinSTT running in the tray rather than quitting. With this off, the close button exits the app entirely.

Opt-out anonymized crash/error reporting via Sentry. Audio and transcripts are never sent — only stack traces and error context.

When on, WinSTT presses a submit key right after the paste lands. Auto-Submit Key (general.autoSubmitKey) chooses the combo — Enter for chat boxes or Ctrl+Enter for IDE prompts. Off pastes and leaves the cursor where the target app puts it.

Reads text from the focused window just before each dictation and feeds it to the LLM cleanup step so names and jargon are spelled correctly. This is platform-gated and opt-in: it shows a confirmation dialog before activating, and is only offered when it actually helps (a Whisper model or LLM cleanup is active).

An app/host allow-out list for context capture. Each entry is an executable basename (1password.exe) or a URL host suffix (bankofamerica.com, matches subdomains). Matching windows have their captured text, HTML, and URL stripped before reaching the LLM. Seeded with six common password managers.

Cap on persisted transcription-history rows (10–10,000). The main process trims oldest on each insert; larger histories slow the history dashboard.

Auto-deletes saved recordings: Never keeps everything; Cap prunes recordings beyond History Max Entries; 3 days / 2 weeks / 3 months are absolute age cutoffs. Cleanup runs at startup and whenever the policy changes — "Never" still saves, it just never prunes.

Opt in to alpha/beta auto-updates. Alpha installs always keep updating to newer alphas regardless of this toggle; the knob only changes behavior on stable builds.