Say

Press ⌥⌘T to start and press it again to stop. Talkie transcribes and pastes when you stop. There’s no time limit; a long dictation is transcribed in pieces and joined.

Push-to-talk is off until you turn it on, under Settings › Dictation › Triggers. Then hold Right ⌃ while you talk and let go to paste. Double-tap it to keep recording without holding, and tap once more to stop. Pressing another key while you hold cancels that recording.

To throw a recording away, click the ✕ on the recording pill.

Talkie copies the transcript to the clipboard and pastes it into the window you started in. If you’ve moved to another window by the time it’s done, Talkie doesn’t paste into the wrong place. It tells you the text is on the clipboard, and ⌘V puts it wherever you like.

The transcript stays on the clipboard afterwards. If you’d rather paste by hand every time, set Paste action to Clipboard Only in Settings › Dictation › Delivery.

If text is selected when you start, the dictation replaces it. To send the message as well, turn on Press Enter after paste in TalkieAgent’s Output settings; it’s off by default.

A pill at the bottom of the screen shows that Talkie is listening and warns you if the microphone is silent. On a Mac with a notch, the notch shows it too. You can move the pill, hide it, or add a bar along the top of the screen in Settings › Dictation › Indicators.

To watch the words appear as you speak, turn on the Live panel in TalkieAgent’s Overlay settings. The preview comes from Apple’s speech recognition, which needs its own permission; the text that gets pasted still comes from your model.

Talkie drops filler words like “um” and “uh” and writes numbers and punctuation as you’d type them. Both happen on every dictation.

For names and terms the model keeps getting wrong, add them to your dictionary under Settings › Context › Vocabulary, then turn on Enable Dictionary Processing. Each entry replaces what the model wrote with what you meant: a word, a phrase, a regular expression, or a fuzzy match for near misses. The Test playground there shows what an entry does before you rely on it.

The dictionary fixes the text after transcription. It doesn’t teach the model new words.

Every model runs on your Mac, so your voice doesn’t leave it to be transcribed. Choose one in Settings › Models.

Parakeet V3

The default; 25 languages, about 250 MB

Parakeet V2

English, about 250 MB

Parakeet 110M

English, about 100 MB; the smallest download

Parakeet Japanese

Japanese, about 250 MB

Whisper

English, in four sizes from 39 MB to 756 MB

Talkie pastes what you said, as you said it. To have it rewritten first, add a rule under Settings › Context › Apps: pick the apps, write a prompt such as “turn this into a commit message,” and choose whether to paste the result or open it in the scratchpad to review. Rules use the AI provider from your settings unless you pick another for that rule.

Talkie follows the Mac’s default input unless you pick a microphone in Settings › Dictation › Microphone. If the one you picked is unplugged, it falls back to the default. You can also switch from the terminal:

Each dictation is saved in the Library under Dictations, with the app and window title you were in, so you can find it later by where you said it. At the Full context level, the default, Talkie also keeps the page’s URL and the text of the field you dictated into, when the app exposes them.

To keep less, lower the level in TalkieAgent’s Output settings to Apps & titles only, or turn capture off.