Talk DAT! (often written TalkDAT) is AI dictation for Windows and Mac by Knight AI+AV, in public beta. Install Windows from the Microsoft Store.
This page lists exactly what works offline and what does not. On the Local route, speech recognition stays on this PC after a compatible model is downloaded, and no provider is involved at all. Everything that still touches a network is named further down.
Speech can go two ways, and you choose which in the Talk DAT! bar. Local uses a downloaded on-device model and keeps recognition on this machine. Bring your own key sends it to a provider you supply credentials for, and Knight is not in that path.
Choose Local and the audio stops at your own hardware. One catch: a compatible model has to be downloaded first, so offline dictation starts after that download rather than out of the installer.
If your connection drops. If Windows reports no connection and a local fallback is ready, the session routes local before capture even starts. Repeated provider failures can switch to local temporarily, and when connectivity returns a probe moves the session back toward your preferred route without rewriting your saved choice.
The default local family choice is Parakeet TDT 0.6B v3. The wider catalog covers Parakeet, Canary, GigaAM, Whisper, Distil-Whisper, and related faster-whisper variants, and a compatible Hugging Face repository or local folder can be configured directly for anything outside it.
You can compare, download, prepare, select, and delete models from inside the app. Downloads are verified and resumable, and preparation records readiness rather than treating any partial folder as complete, which is the failure that usually makes local speech look broken. A background warmup can also prepare the selected model before your next dictation.
There is no fixed minimum. The requirement follows the model: small through larger models trade download size, memory requirements, speed, language coverage, and accuracy against each other.
Two things narrow the guesswork. A hardware scan can recommend a model tier instead of letting an old laptop choose something too heavy, and compatible local engines can use available GPU acceleration. Local inference can also cap CPU use at half the machine's cores, and no more than four, at below-normal priority, so a long dictation does not stall the work you are dictating about.
Local speech recognition does not make the entire application offline. These features have their own network behavior:
Formatting and translation. The finishing pass can run through a provider key you supply, or a locally prepared formatter. Only the second is offline, and it has to be prepared first. Local translation needs a translation model you deliberately install, and nothing large downloads silently.
Updates and sign-in. Update checks make their own network calls. Sign-in, by emailed six-digit code or Sign in with Apple, is optional and needs a connection when you use it. There is no Talk DAT! server that receives your audio or text.
How we check these claims. We only say something stays on your computer after testing an installed build with the network switched off, and we say which feature was tested.
Talk DAT! entries describe the shipping public beta. Other entries state only what each vendor publishes, read on 2 September 2026.
| Capability | Talk DAT! | Windows built in | Dragon | Superwhisper |
|---|---|---|---|---|
| Speech stays on the PC | Yes, on Local, after a model download. | Voice typing needs an internet connection, per Microsoft support. | Desktop editions process speech on the PC. | Not confirmed |
| Choose your own model | Yes, including a compatible Hugging Face repository or local folder. | Not confirmed | Not confirmed | Not confirmed |
| Price | Free. | Free and built into Windows. | Paid professional editions; check their site for current pricing. | Paid plans; check their site for current pricing. |
| Compliance claims | None: no SOC 2, no ISO 27001, no BAA. | Not confirmed | Not confirmed | Claims SOC 2 Type II and HIPAA, per their site. |
"Not confirmed" means we could not verify the capability from public information and would rather say so than guess. Competitor names belong to their respective owners; this page is not affiliated with or endorsed by Microsoft, Nuance or Superwhisper. Their published facts were read on 2 September 2026.
What is protected. On Local, recognition runs on your own machine and no provider is involved at all. On bring your own key, your chosen provider receives the request and Knight is not in that path. Neither mode sends desktop content to Knight AI+AV. Optional sign-in traffic uses HTTPS with server-side token verification. Knight does not receive request audio, transcripts, or finishing text.
What is not protected. Talk DAT! writes transcripts, a searchable history, and short recovery recordings to your user profile without encrypting them, relying on your operating system account for protection. There is no permanent audio archive, but a limited recovery buffer is kept on this PC, so "no audio is stored" would be false. Knight has not completed a SOC 2 audit, holds no ISO 27001 certification, and offers no Business Associate Agreement. The optional PII redaction setting is a convenience filter, not de-identification. No system is perfectly secure.
Straight answers.
Speech recognition does, on the Local route, after a compatible on-device model is downloaded. If Windows reports no network and a local model is ready, the session routes local before capture starts. Offline dictation begins after that download, not out of the installer.
Update checks, optional sign-in, and any provider you add your own key for, including a provider-based finishing pass, use the network. Local speech recognition does not make the entire application offline.
No fixed figures are published, because the answer depends on the model you pick. Small through larger models trade download size, memory requirements, speed, language coverage, and accuracy, and a built-in hardware scan can recommend a tier for your machine.
No, and that matters. Talk DAT! writes transcripts, a searchable history, and short recovery recordings to your user profile without encrypting them, relying on your operating system account for protection. Provider keys and sign-in tokens do use protected credential storage where available.
Not as sold. Knight has not completed a SOC 2 audit, holds no ISO 27001 certification, and offers no Business Associate Agreement, and the optional PII redaction setting is a convenience filter rather than de-identification. Regulated work should use a product that carries the paperwork.
Voice Access is Microsoft's built-in accessibility control, and Windows voice typing needs an internet connection per Microsoft support. Dragon is professional desktop software that processes speech on the PC. Talk DAT! sits between them: on-device recognition plus AI cleanup.
Install, download a model, disconnect, and dictate. It is free.
