Is voice typing private? Eight things a dictation app can send, and how to check each one

Where your voice becomes words

One question decides more than all the others: does the app recognise speech on your device, or on a server?

If recognition happens on your device, the recording never has to go anywhere. If it happens on a server, your voice is uploaded every time you dictate, and what happens to it next depends on that company's policy and its settings. Microsoft's support page for Windows voice typing, the Win+H feature, said when we checked on 5 September 2026 that it uses online speech recognition and needs an internet connection. Wispr Flow's privacy page said, on the same date, that its transcription always happens in the cloud.

Neither is hidden; it is simply the first thing to look for. You can test it in a minute: switch off the wifi and try to dictate. A tool that stops working most likely recognises speech on a server — or insists on a sign-in check, which is worth knowing too — and offline dictation on Windows lists the tools that do neither.

Eight things a dictation app can send

Recognition is only the first row. Here is the whole list, the same one the byte table on this site fills in for each tool.

1. Audio. The recording of your voice. It carries more than the words: who you are, and who else is in the room.
2. The dictated text. Some apps send the transcript to a language model to tidy it up, even when recognition is local. Ask how often, and whether the text is kept.
3. Context around the text. To punctuate a sentence well, an app may want to know where it is going. That can mean the app's name, or the text already in the box, the window title, what is on your screen, or a screenshot. Wispr Flow's Context Awareness help page, checked on 5 September 2026, listed a screenshot among the context sent with each dictation request, and said the feature is on by default on Mac and Windows.
4. Telemetry and analytics. Usage reports, device details, logs. Typeless's privacy policy, checked on 5 September 2026, listed usage data such as IP address, device type and system logs, and named Google Analytics.
5. Update checks. A request to see whether a new version exists. Usually harmless, but still a call home, so know what it carries.
6. Model downloads. An app that recognises on your device needs a speech model on it, fetched from somewhere at least once.
7. Account and sign-in. An email address, a name, a payment method, and a record that ties everything else to you. Ask whether you need one to dictate at all.
8. Integrations. Anything the app sends to other services for you: a note to a chat channel, a ticket, a webhook. Check where they go and what they include.

How to find out what a given tool sends

Open the tool's privacy policy and help pages and go down the eight rows. For each, you want three facts: what is sent, where it goes, and when. Vague wording about collecting information to improve a service answers none of them. A few habits help:

• Read the defaults, not the options. A setting that exists but ships switched on is what most people run. Wispr Flow's pages said on 5 September 2026 that "Improve the model for everyone" is pre-selected during onboarding, and that with it on, dictation audio, transcripts and edits may be used to train its models.
• Treat silence as unknown. A row the policy does not mention is unknown, not "nothing".
• Look for the words "discarded" and "retained". Typeless's privacy policy, for instance, said on 5 September 2026 that audio and contextual information are processed on its cloud servers in real time and discarded once the result is returned to your device. That is a clear answer, and it is still an upload.
• Test with the network off. It tells you whether the recognition is local.
• Watch the traffic. A firewall that logs outbound connections shows what the app actually does.

What Deixis sends

Deixis recognises speech inside the app on your own device: on Windows, Mac and Android, and on iPhone, where its keyboard is in beta on TestFlight. Here are the eight rows filled in.

• Audio: never for a dictation or a meeting. That audio stays on your device under every setting and on every plan, and a meeting recording cannot even be shared. The only audio that can leave is a screen recording you narrate and then choose to share with Share a link.
• Text to a language model: only if you set up a cleanup endpoint. Out of the box, a rules engine on your machine punctuates and tidies what you said, and no text leaves. With an endpoint set, the default, When it helps, sends a sentence only when you ask for a transform ("make this a bullet list") or dictate a long stretch with no punctuation. Translation, meeting notes and recording summaries use the same endpoint. Deixis's own endpoint routes only to model endpoints with zero data retention.
• Context: names, not contents. On Windows and the Mac, with that same request go your vocabulary terms, the name of the app and the field the text is going into, and the sentence you last dictated into that field, so the next one follows on. Never the window title, which often carries a document or customer name, and never a screenshot. The Android app has a switch, off by default, that also sends the 400 characters before the cursor.
• Telemetry: none. Deixis has no telemetry.
• Update check: nothing but the request. On Windows and the Mac it runs at every launch, and the version comparison happens on your machine; on Android, Google Play handles updates.
• Model download: once, when you press Download. On first run the speech model is fetched from huggingface.co, and the request carries nothing but itself.
• Account: not needed to dictate. Deixis never asks you to sign in to dictate. You give an email address if you choose to start the free trial of the cloud features, or to make a share link.
• Integrations: only to addresses you add. A webhook receives the finished note or translation, and for a screen grab its path and file name, never the image itself.

So on Windows and the Mac a fresh install makes two calls on its own, and neither carries anything but the request: the update check at launch, and, when you open the Settings window, a check for an announcement from us, which has an off switch of its own. Everything else waits until you configure it or press something. One key, for Claude Code users, asks Anthropic how much of their Claude plan is left, and only when pressed. The one feature that uploads a file of yours is Share a link, which puts the grab, note or screen recording you picked where somebody can open it by URL.

What Deixis does not offer is cloud cleanup without the context: a sentence either goes with its app and field names, or it does not go at all. For a language model's help with nothing leaving, on Windows point cleanup at a model on your own machine, such as Ollama; Settings, and where cleanup runs shows how. Clearing the endpoint stops cleanup, translation and meeting notes calling out at all, and dictation carries on with the local rules.