Talk DAT!

Privacy Notice

Effective October 27, 2026

Talk DAT! runs on your device, and it is free. With local speech and writing models selected, which is the default, your audio and the text you dictate stay on your computer. In version 0.4.175 for Windows and for the Mac download, command search is the one exception. If you select a speech or writing provider, audio or text for that route goes directly to that provider under your account and its privacy terms. A feedback report contains a support log only if you choose to attach it. On its own, the app uses the internet for the usage counts, which you can turn off, for update checks and for speech model downloads; none of those carries what you said or wrote. The optional browser demo on this website sends what you say to be transcribed and finished, and only when you press it.

Who we are

Talk DAT! is made by Knight AI+AV LLC of Sacramento, California ("Knight", "we"). Questions about this notice, and every privacy request, go to Build@KnightAIAV.com.

What this notice covers

This notice covers the Talk DAT! desktop apps for Windows and Mac, this website and its browser demo, and our server at api.talkdat.app, which runs the optional account, feedback and the usage counts. A copy of Talk DAT! that someone else builds from its open-source code contacts none of our services unless its builder sets it up to.

Local data on desktop

On Windows and macOS, transcripts, protected voice sessions, history, dictionary entries, snippets, settings, and scratchpads remain under your Windows or macOS user profile. Your settings include two random IDs: an install ID for the usage counts, and a device ID that is used only if you sign in. Provider keys and your sign-in token use the operating system's protected credential storage when available.

Protected voice sessions are recordings of your recent dictations, kept so that a dictation that fails can be recovered. Talk DAT! keeps at least the last five. A recording of a dictation that failed is kept until you recover or clear it, for up to 14 days and no more than 50 of them. These recordings are kept even when History is off. On the local route they stay on your computer; for a provider route, see Bring your own key.

These files are written unencrypted under your user account, so anyone with access to that account, or to an unencrypted disk, can read them. Turning on full-disk encryption is the single most useful thing you can do about that, and Settings can delete history, recordings and the dictionary at any time. The Windows uninstaller keeps these files unless you choose to remove them.

There is no Talk DAT! cloud

Earlier versions offered to transcribe and rewrite on Knight's servers. That service is gone. There is no setting in the app that turns any of it back on: the code that did it was removed rather than switched off. If a local part cannot run, Talk DAT! tells you instead of sending the work elsewhere.

When the app uses the internet

With the default settings, Talk DAT! connects to the internet for three things, and none of them carries your audio or what you dictate:

It also goes online when you choose to: when you sign in, send feedback, use your own key for a speech or writing provider, or use command search; when you set up smart formatting, which installs Ollama and its writing model, through winget on Windows and through Homebrew or ollama.com on a Mac; and when you turn on GPU speed-up on an NVIDIA graphics card, which downloads NVIDIA's runtime from PyPI.

For advanced users, the environment variable TALKDAT_NO_PHONE_HOME=1 turns off every connection to our services and the update check. It does not stop model downloads, command search or a provider route you set up.

Usage counts

To learn whether Talk DAT! works for people, the app sends four kinds of report to our server: first open, once, when the app starts for the first time; first dictation, once, after your first successful dictation, with how many seconds it took; still in use, at most once a day, when the app starts; and day 7, once, when you dictate again after seven days.

Each report carries a random install ID, not linked to you or your account, which the app creates the first time it runs. It also carries the app version and whether the app runs on Windows or Mac. The first open and first dictation reports carry the time the app was first opened, and the first open, still in use and day 7 reports carry a yes or no flag that marks our own test computers. No audio, no text, no email address and nothing that names you, your account or your computer is part of a report.

Our server keeps one record per install ID. It holds the time of each kind of report, how long the first dictation took, the app versions and platform the reports named, and the test-computer flag. A new still in use report replaces the one before it. We do not link these records to your account. The install ID does let us tell that several reports came from the same install. Our server sees your IP address when a report arrives, as with any internet connection, and does not store it with the report. Railway, which hosts our server, keeps its own request logs under its own rules.

Your choice. The usage counts switch is in Settings, Privacy. It is on by default, setup mentions it once, and you can turn it off at any time; while it is off, the app sends none of these reports. In version 0.4.175, the first open report and the first still in use report go out when the app first starts, before setup shows the switch, and resetting the app settings with Reset turns the switch back on. It is separate from Local-only privacy: that setting keeps your audio and text on your device, and these reports never contain either. A copy of Talk DAT! built by someone else from its open-source code does not send them to us.

How long we keep them. We delete an install's record 12 months after its last report.

Deleting your record. We cannot find your record from your name or email address, because we do not link it to you. Send your install ID to Build@KnightAIAV.com and we will delete the record for that ID. The install ID is in Talk DAT!'s settings file, config.json, under "metrics" as "install_id". On Windows the file is in the TalkDat folder in %APPDATA%, or in the TalkDatData folder beside the portable app; the Microsoft Store version may keep it in a folder that Windows manages for Store apps. On a Mac it is in ~/Library/Application Support/TalkDat. If you cannot find it, write to us and we will help. The ID stays the same when you update Talk DAT! and when you use Reset, and it stays after you uninstall until Talk DAT!'s files are removed.

Bring your own key

If you add an API key for a speech or writing provider, requests on that route go from your device straight to that provider, under your account with them and their privacy terms. Knight is not in that path, never sees the request, and adds no charge; the provider bills you directly. Remove the key and the route stops.

Local-only privacy ("Keep speech and finishing on this computer"), which is on by default, keeps these routes off until you turn it off. A speech request carries your audio, and it can include your custom words so names come out right. A writing, translation or rewrite request carries the text, your custom words and the style you chose. Requests to Deepgram ask it not to use your audio to improve its models, unless you turn that option off. If a dictation on a provider route is interrupted, Talk DAT! can send its saved recording to that provider again to retry it. When you ask Talk DAT! to transcribe a saved provider recording again, it sends the recording through the speech route you have selected. If you point Talk DAT! at a server of your own, such as an Ollama address or a custom endpoint, your text goes to that server.

In command mode, if no text is selected, Talk DAT! opens a web search for the words you said at perplexity.ai, in your default browser. Perplexity receives those words, and whatever your browser sends, under Perplexity's own terms. Knight does not receive them. In version 0.4.175 for Windows and for the Mac download, this happens even when Local-only privacy is on. To avoid a search, select text before you use command mode.

The optional account

You do not need an account to dictate. If you sign in, with a six-digit code sent to your email address or with Sign in with Apple, our account service stores:

If your account was created before September 22, 2026, it may also hold records from earlier versions, such as word counts and plan dates from before Talk DAT! was free. They are deleted with the account.

It never stores audio, transcripts, dictated text, dictionaries, provider keys, or history. The account service is our own API at api.talkdat.app, hosted on Railway with its database in Postgres on Railway. Sign-in codes are emailed through Resend. Sign in with Apple is handled by Apple under its own terms. We do not use a hardware fingerprint.

Feedback you send

If you use the in-app form to report a problem or share an idea, we store the words you type, the reply address if you give one, the app version, the platform and which part of the app you sent it from, so that we can read and answer it. The form can also attach a formatting log, but only when you tick the box that says so. That log can contain text you dictated, which is why it is never attached without your say. We do not link feedback to your account. We keep a report for 12 months after we receive it and then delete it, and we delete one sooner if you ask.

This website

The website has no advertising, no analytics, and no tracking pixels. The browser demo, when you press it, works as described under Browser demo. Your browser keeps your chosen site theme and music setting, and, if you sign in here, your sign-in session, in its own local storage on your device. Each page asks GitHub for the latest release details when it loads, so GitHub sees that request, and the account panel loads Apple's sign-in script from Apple when you reach it. "Email me the link" opens your own mail app; nothing is sent through us. Our hosting provider, Railway, keeps ordinary access logs, such as IP addresses and request times, to run and protect the site. If you type talkdat.app without "www", our domain provider, Squarespace, forwards you to www.talkdat.app. Datadog checks from outside that this website, our server and the demo service respond; it receives none of your data.

Browser demo

The home page has an optional browser demo so you can hear what Talk DAT! does before installing it. This preview uses remote speech and writing services. The desktop apps do not use Knight's demo service. Their local and selected provider routes work as described above.

The microphone starts only when you press it, and only after you tick a box agreeing to it the first time. While it listens, your browser streams your microphone audio straight to Deepgram, which transcribes it, using a token from Knight's demo service that lasts about 30 seconds, with Deepgram's model-improvement opt-out switched on. The rough text is then sent through Knight's demo service to OpenRouter, which passes it to a language model to finish it in the Chill or Executive style you picked; those requests deny data collection and require providers that keep no data. OpenRouter receives them from our server, so it does not see your IP address. If you use the written box instead of the microphone, only the text you typed is sent, to OpenRouter the same way, when you press Clean it up; the box does not ask you to tick the agreement first. If Deepgram cannot be reached, the page may fall back to your browser's built-in speech service, which follows your browser maker's privacy terms.

Nothing is stored by Knight. The demo service holds the text only for as long as the request takes, and it never receives your audio at all. To limit use, it keeps your network address and a random demo identifier in memory for up to five minutes. Like the website, it is hosted on Railway, which keeps ordinary access logs, such as IP addresses and request times. Nothing from the demo is added to an account. Your browser keeps that demo identifier, a count of how many voice tries you have used in the last five minutes, and whether you have agreed to the demo, in its own local storage, so the demo can limit how often it runs; these are not used for advertising or analytics. The demo microphone turns off when you release, stop, leave or hide the page, and each take is capped at 30 seconds.

Features we retired

On September 22, 2026 we retired the demo-to-desktop email handoff, the price-reminder list and every paid plan. The website no longer offers them. The browser demo was retired the same day and brought back as described above, without its email handoff and without the first-party demo funnel that used to measure it. If you used one of them before then and want any record of it deleted, write to Build@KnightAIAV.com and we will delete it.

Services that receive data

Services you choose yourself, each under its own terms: the speech or writing provider you add a key for; Perplexity, for command search; PyPI, for NVIDIA's runtime; Ollama, and on a Mac Homebrew, for smart formatting setup; and your browser maker, if the demo uses your browser's speech service. If you install from the Microsoft Store, the Store's own terms and privacy policy cover the download and its updates.

Children

Talk DAT! is not directed to children under 13, and we do not knowingly collect personal information from them. If you believe a child under 13 has created an account or sent us information, write to Build@KnightAIAV.com and we will delete it.

Keeping and deleting your data

Regulated data

The desktop app does not send your dictation to Knight unless you attach a formatting log to a feedback report, so do not attach one that holds regulated data. The app transcribes on your own device, or at a provider you supplied a key for, so the content does not reach Knight's servers to begin with. The browser demo on this website is not for regulated data: it sends what you say to Deepgram and OpenRouter. Talk DAT! still writes transcripts, history, and short recovery recordings to your user profile unencrypted, so the device remains your responsibility. Knight claims no HIPAA or SOC 2 certification, and the optional PII redaction setting is a convenience filter rather than de-identification.

Security and changes

Account traffic uses HTTPS. Sign-in codes and session tokens are stored hashed, sessions expire and rotate, and the desktop app's sign-in token is signed so it cannot be forged. No system is perfectly secure; see the security policy for how to report a problem. Material changes to this notice will be dated here.

Colophon

DAT is Digital Audio Tape, the format professionals recorded and mastered onto from 1987 onward. It also reads as Digital Audio Transcription, which is the job. Built by Knight AI+AV.