Dictation

Talk into a terminal, a text field or a web page: on-device transcription with a local Whisper model, started from a button, a chord, or a held key, with a cursor that shows it listening.

TermHQ types what you say into a terminal, into one of its own text fields, or into a text field on a web page in a browser pane. Transcription happens on your own machine, on a Whisper model you download once — nothing is sent anywhere, which matters for a terminal sitting in front of proprietary source. It exists because prompts to coding agents are long and typing them is the slow part, but it works at any shell prompt.

Before the first word

The installers ship without a speech model, so a fresh install shows no mic button at all. Open Settings → Voice, download a model (or import one you already have), close Settings, and the button appears in the title bar. The panel says at the top whether dictation can run on this build and whether a model is ready, before it offers anything to download. Installation covers the models and their sizes; Test microphone in the same panel plays a few seconds back through the exact capture path dictation uses, so a silent microphone is found before any words are spent on it.

Keep Settings open while a model downloads: closing it cancels the download. Press Esc, click its close button or click outside it mid-download and TermHQ asks first — Keep downloading, or Close and cancel it.

Linux support is pending.

Three ways to start

• The mic button in the title bar. Click to start, click again to stop. The words go to the focused terminal, or to the text field you were typing in when you clicked.
• The chord, Ctrl+Alt+Space on every platform — the one shortcut that does not move to ⌘ on macOS, because ⌘+Alt+Space is Finder’s search. It works in a terminal, in TermHQ’s own text fields, and over a web page in a browser pane.
• Hold to talk, off by default. Turn it on under Settings → Voice and holding the space bar in a terminal for a moment starts dictating; let go and it transcribes — the gesture Claude Code’s own dictation uses, here for every program. A tap still types a space, and a key pressed mid-hold flushes the space ahead of it, so typing is unchanged.

Stopping — a second click, the chord again, or letting go of the held key — sends the recording to the model. The button and the chord stop a recording even if you have clicked into another terminal, an editor or a text field since. The exception is ultra focus, which hands the chord to the terminal — use the button there. A recording under 0.4 seconds is dropped rather than transcribed.

The key you hold

Space is the default because it is the key Claude Code uses, and that is also its cost: while Hold to talk is on, holding the space bar inside Claude Code starts TermHQ’s dictation instead of Claude Code’s. If you want both, give TermHQ’s a key of its own. Settings → Voice → Hold key shows the current key; click it and press the one you want — any single key, a modifier such as Right Ctrl included, which is the classic push-to-talk key. Esc keeps the current one. A quick tap of the key still does what it normally does: a character key types its character, a modifier or function key tapped alone does nothing.

The hold is not a shortcut in the usual sense — it is a hold, not a chord — so it lives in Settings → Voice rather than in Settings → Shortcuts.

Where the words land

In a terminal, the transcript is typed into the terminal you started dictating in, not into whichever pane happens to be focused when the model finishes. Start in one pane, click into the next while it records or transcribes, and the words still arrive in the first. Dictate twice into the same prompt without typing in between, and the second lands after a space so the two do not run together.

In a text field — the commit box, a search box, a browser pane’s address bar — the words go in at the cursor, replacing any selected text, with a space in front where they would otherwise run into the word before. The field takes them as if you had typed them.

On a web page in a browser pane, the words go into the text field the page had focused when you started — a search box, a comment form, a chat input — typed through the page’s own editing, so the site sees ordinary typing and its own undo takes them back. They go where you started even if you click elsewhere while they are being transcribed, with a space in front where they would otherwise run into the word before. A page with no text field focused has nothing to type into, and dictation says so before the microphone opens.

Anywhere else there is nothing to type into: an editor, a terminal whose shell has ended. There dictation does not start. The mic button says Nothing to dictate into in the status bar, and the chord goes on to the pane like any shortcut with nothing to do.

Nothing is submitted for you. The words appear at the prompt as if typed, and Enter is yours to press — read them first.

The cursor shows it

While a recording is aimed at a pane, that pane’s cursor becomes a block that shows what is happening. Recording: a bar inside the cell rises with your voice — gray in a quiet room, cycling through color the moment it hears you, tuned so a quiet microphone still swings most of the cell. Transcribing: a chrome block with the light traveling upward, in time with the mic button’s three dots and the Transcribing… label in the title bar. The block follows the cursor as the shell moves it and disappears the moment the words land.

Reduced-motion settings still the animations and keep the block.

When it fails

A dictation that cannot finish says why in the status bar, and the message’s tooltip says what to do:

Beside the mic button, Dictation failed shows for a moment at the same time.

Language and models

Dictation listens in English by default. Settings → Voice → Dictation language offers fifteen more and an auto-detect option that lets a multilingual model identify the language per recording; models labeled English only stay English whatever this says. Models can be switched, deleted or re-downloaded at any time, and they survive updates.

When something is wrong — no button, a flat waveform, the wrong words — see Troubleshooting.