Talk DAT! (often written TalkDAT) is in public beta. Install the Windows build from the Microsoft Store, or take the notarized Mac disk image for Apple silicon on macOS 14 or later.
Runs on your computer. On the Local route your prompts and your code never leave the machine: hold a key, speak, release into the terminal. Your own provider key is there when you want it. See the prompt forming while you are still saying it. On a local model, with the switch on, the Talk DAT! bar shows the text so far about 1.2 s behind your voice, on a GPU or on six CPU cores, so you can see a sentence go wrong before it is typed into Cursor or Claude Code. Nothing leaves the machine. Prompts are long, and typing them is the slow part.
A prompt to a coding agent is more like a note than code: it is long, it is plain English, and you only use it once. That makes it easier to say than to type. With Talk DAT!: hold Ctrl+Win, say the whole thought, release.
It does not read your editor. It is not an extension, it has no view of the file you have open, and it does not know what sits to the left of your caret. It is a system-wide overlay that delivers text into the focused window. Because of that, one setup works in Claude Code, Cursor, a browser, and a plain terminal.
Install from the Microsoft Store on Windows, or use the notarized disk image on an Apple silicon Mac running macOS 14 or later. Both are public betas. Talk DAT! is free and needs no account to dictate; run the guided setup, which walks you through the trigger.
For an editor, nothing further is needed: Auto delivery picks the insertion method from the focused process, and ordinary applications get a verified clipboard write followed by Ctrl+V. For a terminal, set the delivery method to Shift+Insert, which terminal-style targets accept where a plain paste chord may not. Recognised remote clients, including RDP, VNC, TeamViewer, AnyDesk, RustDesk, and Parsec, already route to direct typing.
Copy-only mode is a safe way to start: Talk DAT! puts the finished text on your clipboard and inserts nothing. The push-to-talk page covers the trigger side of the same loop.
Speech recognition often gets names wrong, and code is full of names. Words and Phrases takes a preferred spelling plus optional sounds-like aliases, so a repository, a service, or a library is written the way you write it. Terms are stored as canonical spelling, aliases, and phonetic keys rather than blind replacement, and your own terms outrank the built-in defaults.
A raw dictionary editor takes a batch paste for a long list, and a vocabulary pack can be carried to another machine. The vocabulary pass runs again after formatting, so a model pass is less likely to undo a name. Proper nouns in the foreground window title can become transient local hints ranked below your explicit terms; that reads the window title only, not the screen.
The limit. This post-corrects recognised text. It does not retrain speech model weights, and one example does not guarantee a name is always spelled correctly.
Two routes, picked from the Talk DAT! bar. Local runs a downloaded on-device model on your machine, with NVIDIA Parakeet TDT 0.6B v3 as the default, and local inference is deliberately restrained: it can cap CPU use at half your cores and no more than four, at below-normal priority, so a dictation does not fight your build. Bring-your-own-key sends audio from your machine straight to the provider whose key you configured, billed by that provider under your account, and Knight is not in that request path.
Note that running speech recognition locally does not make the whole application offline, because updates, and sign-in if you choose it, reach the network. The bring-your-own-key page covers providers and key storage; the offline dictation page covers the local models.
Chill, the default, keeps your own phrasing. Executive is a fuller professional rewrite that can restructure prose, add headings, and remove profanity; it is one click away in the Talk DAT! menu. For a prompt, Chill is usually the right choice, because Executive may reword your instructions.
Both finishes run output validation. Empty responses, refusals, code fences, severe shrinkage, malformed lists, lost quotes, and changed numeric, URL, email, or hashtag anchors can be rejected, and a rejected result falls back to the deterministic pass. These checks catch many problems but not all, so check names and numbers yourself.
Short, safe dictations can paste a fast deterministic pass first and refine it moments later if the same window is still focused. That path has a real gap: the safety helper accepts a signal for user typing since the paste, orchestration does not currently supply it, and so typing into the same window during that moment can be overwritten. Wait for the Talk DAT! bar to show it is done before typing in that window.
The rest, in short. Delivery goes to the final focused window, not the one you started in. A successful synthetic paste does not prove the application accepted it. The app does not read the text around your cursor. Sessions are bounded rather than unlimited: hands-free defaults to about five minutes and a hold can extend toward thirty. And an opt-in control API binds to 127.0.0.1 on port 4670, can require a bearer token, and exposes status, toggle, cancel, paste-last, and copy-last, with matching command-line flags.
Straight answers, matching what the product actually does today.
Yes. Talk DAT! is a system-wide overlay rather than an editor extension, so it delivers finished text into whichever window has focus at delivery time. It has no view of your files and does not read the text around your cursor, so one setup works in an editor, a browser, and a terminal.
Set the delivery method to Shift+Insert, which terminal-style targets accept where a plain paste chord may not. Simulated typing is available as an explicit method or a fallback, at a configurable interval that defaults to about 2 milliseconds a character. Recognised remote clients route to direct typing.
Add them in Words and Phrases with a preferred spelling and optional sounds-like aliases. Your own terms outrank the built-in defaults, a raw dictionary editor accepts a batch paste, and the vocabulary pass runs again after formatting. This post-corrects recognised text rather than retraining a speech model.
Yes, on the Local route. A downloaded on-device model transcribes on your own PC, with NVIDIA Parakeet TDT 0.6B v3 as the default selection, and local inference can cap CPU use at half your cores and no more than four. Local speech recognition does not make the whole application offline: updates, and sign-in if you choose it, still reach the network.
Chill, most of the time, and it is the default. Executive is a fuller professional rewrite that can restructure prose and add headings, which can change what your prompt asks for. Chill keeps your phrasing and tidies it. The switch is one click in the Talk DAT! menu.
Yes, through an opt-in localhost control API. It is disabled by default, binds to 127.0.0.1 on port 4670, rejects non-loopback Host values, and can require a bearer token. It exposes status, toggle, cancel, paste-last, and copy-last, with matching command-line flags. There is no remote or cloud control.
Install Talk DAT!, hold Ctrl+Win, and say the whole thought at once. Free to use, with no account needed to dictate.
