Voice input: add a keyboard shortcut to start/stop dictation

[Ady6688]

Feature request: keyboard shortcut for voice input

Companion bug report (silent capture device misreported as "No speech recognized") is filed separately in Q&A.

Problem

Voice input can only be started and stopped by clicking the microphone button. There is no keyboard shortcut:

• The only key handler in VoiceInput is Escape, which cancels an active recording.
• The desktop main process registers no speech / voice related globalShortcut or menu accelerator.
• zh.json / en.json contain only meta.title / meta.description — no shortcut hints.

For users dictating while typing, moving to the mouse for every recording interrupts the input flow.

Proposal

Add a configurable shortcut that toggles recording — idle → start, recording → stop and insert.

Suggested defaults:

Platform	Binding
macOS	⌘⇧V
Windows / Linux	Ctrl+Shift+V

Fall back to ⌘⇧Space or ⌘\ if ⌘⇧V collides with "paste as plain text" in the host app.

Acceptance criteria

• Pressing the shortcut while idle starts recording, subject to the same guards as the button (catalog, usable, locked, no active capture).
• Pressing it again while recording stops and transcribes, then inserts the text at the captured insertion point.
• The key is swallowed (preventDefault) when focus is inside the composer, so no character is typed into the draft.
• The shortcut is inert when locked is true, matching the button's disabled state.
• The shortcut is scoped to the application window (not a global hotkey), so it does not fire when the app is unfocused.
• The binding is discoverable — shown in the microphone button's title / aria-label and in the plugin details page.
• Binding is configurable, defaulting per platform as above.

Implementation notes

start() and finish() already contain complete guards, so the shortcut can call them directly without duplicating validation:

const start = async () => {
    if (!catalog || !usable || locked || current.current) return;
    ...
};

Bind on document to match Escape's existing scope, and clean up when sessionId changes to avoid duplicate handlers across sessions.

Out of scope

• Push-to-talk — requires keydown / keyup pairing, key-repeat handling, and recovery when a keyup is lost on focus change (leaving recording stuck). Consider a follow-up.
• Custom key-recording UI — ship a small preset list first.

---

[Ady6688]

Formatting note: the opening post predates a check of the repository templates; the text below follows .github/ISSUE_TEMPLATE/feature.md (Motivation / Behavior).

---

Motivation

Voice input can only be started and stopped by clicking the microphone button. There is no keyboard shortcut:

• The only key handler in VoiceInput is Escape, which cancels an active recording. Nothing starts one.
• The desktop main process registers no speech / voice related globalShortcut or menu accelerator.
• zh.json and en.json expose only meta.title / meta.description, with no shortcut to document.

For a user dictating while typing, reaching for the mouse before and after every recording interrupts the input flow. The microphone sits between the model selector and the send button, so the pointer round trip is short but constant.

Behavior

Add a configurable shortcut that toggles recording: idle → start, recording → stop and insert.

Suggested defaults, chosen to avoid collisions with common editing bindings:

Platform	Binding
macOS	⌘⇧V
Windows / Linux	Ctrl+Shift+V

Where ⌘⇧V is already taken by "paste as plain text", ⌘⇧Space or ⌘\ are acceptable fallbacks. The binding must be user-configurable rather than fixed.

Expected behavior:

• Pressing the shortcut while idle starts recording under the same guards the button applies: catalog resolved, provider usable, not locked, and no capture already active (start() already checks all four).
• Pressing it again while recording stops, transcribes, and inserts the text at the insertion point captured when recording began (inputActions.captureInsertion() / inputActions.insertText()).
• The key is consumed (preventDefault) when focus is inside the composer, so no character is typed into the draft.
• The shortcut is inert while locked is set, matching the button's disabled state.
• The shortcut is scoped to the application window rather than registered as a global hotkey, so it does not fire when the app is unfocused and cannot collide with system bindings.
• The binding is discoverable: appended to the microphone button's title / aria-label, and settable from the plugin details page alongside the provider and language selects.

start() and finish() already contain complete guards, so the shortcut can call them directly without duplicating validation. Binding on document matches Escape's existing scope; the handler must be torn down when sessionId changes so sessions do not accumulate listeners.

Out of scope for an initial version:

• Push-to-talk. Requires keydown / keyup pairing, key-repeat handling, and recovery when a keyup is lost as focus changes, which leaves recording stuck. Better as a follow-up.
• Custom key-recording UI. A small preset list covers the need without the edge cases of a key-capture component.

Related

A companion bug report (a silent capture device misreported as "No speech recognized") is filed separately under Q&A.