Dictation apps for Windows: 12 compared on where your voice goes

The short answer

Which Windows dictation apps keep your voice on your PC?

Six of the twelve apps say, in their own documentation, that they can recognise speech on your PC: Dictro, superwhisper, Handy, OpenWhispr, SnailText and Microsoft's own Voice access. Dictro, Handy and SnailText have no cloud transcription mode at all. superwhisper and OpenWhispr offer one you can switch on, so with those the answer depends on the model you have selected.

Windows' built-in voice typing (Win + H) is not on that list. Microsoft says it “uses online speech recognition, which is powered by Azure Speech services.” Wispr Flow, Typeless, Willow Voice and Voibe also transcribe in the cloud on Windows. Dragon is installed on your PC, but Nuance's pages do not say where its recognition runs.

How this was put together

Method, scope, and a disclosure you should read first

Dictro is one of the twelve apps compared here, and this page is published by Dictro. Read it with that in mind. What keeps it useful anyway is the method: this is a documentation audit, so you can check every claim against the same public sources we did, including the ones about us.

One idea shapes the whole comparison: a privacy policy and a privacy architecture are different things. A policy is a promise that can be revised; an architecture decides what is even possible. An app that recognises speech on your PC cannot leak that audio to a server, whatever its policy says next year. That is the axis this audit sorts on.

At a glance

Eight Windows dictation apps, compared on architecture

“Local by default” and “local only” are not the same guarantee. Several good apps here recognise speech on your PC until you choose a cloud model, which is a real feature, but it means the privacy answer depends on your settings rather than on the app. Four more apps are covered briefly under Also checked.

The evidence, app by app

What each vendor actually says

Dictro — on-device only, no cloud mode to leave on

Dictro recognises speech on your PC's processor, with no graphics card needed, and there is no cloud processing mode to opt into or forget to switch off. Hold Right Ctrl, speak, release, and the text is placed in the field you were already typing in — Outlook, Word, Slack, a browser, a terminal or a code editor. It keeps working with the network off.

What happens between hearing you and typing is deliberately mechanical on Windows: filler words are removed, spoken numbers become digits, sentences get capitals, and each app can have its own writing style. It does not rewrite your grammar — the Mac app adds a language-model pass for that, and on Windows the rules run alone, so they can never reword or invent what you said.

What Dictro does send: the same scrutiny applied to everyone else here. Dictro requires a sign-in, and the app sends an anonymous heartbeat with a random identifier, the app and operating system version, and an aggregate count of words dictated. It never sends audio, transcripts, or the text you dictate. The detail is in the privacy policy.

Limits worth knowing: 64-bit (x64) Windows 10 and 11 only, English only today, and the installer is not code-signed yet, so Windows shows an “unknown publisher” prompt the first time. The free plan is 10,000 words a week with no card; Pro is $5 a month, $48 a year or $99 once (see pricing). More on the app itself at Dictro for Windows.

Windows voice typing (Win + H) — free, built in, and online

Microsoft is unambiguous about where this runs: “Voice typing uses online speech recognition, which is powered by Azure Speech services.” It adds that “To use voice typing, you'll need to be connected to the internet, have a working microphone, and have your cursor in a text box.” Source: Microsoft Support

Windows' privacy page describes the online speech recognition setting the same way: it “Uses Microsoft cloud-based services for more accurate recognition. Voice data is sent to Microsoft only to provide the service and create text transcriptions.” Source: Windows privacy

To its credit, it is free, already installed, and on Windows 11 it offers “Automatic punctuation, to let voice typing automatically insert punctuation based on what you're saying.” On Copilot+ PCs, a newer feature called fluid dictation corrects grammar and filler words, and Microsoft says “This experience is powered by on-device small language models (SLMs) to ensure fast and private processing.” Note what that covers: the tidy-up step. The speech recognition beneath it is still the online service quoted above.

Voice access — Microsoft's on-device option

Voice access is the built-in tool most “best dictation app for Windows” lists miss, and it is the one that runs locally. Microsoft says it uses “modern, on-device speech recognition to accurately recognize speech and works even without the internet.” “Voice access is available in Windows 11, version 22H2 and later,” and “To complete the initial set up voice access in Windows, you need an internet connection.” Source: Microsoft Support

It is built first for controlling the whole PC by voice, with dictation as one part of that. “To dictate text into a text box, first move the focus to the text box,” and “By default, automatic punctuation is turned off.” Check your language before relying on it — Microsoft's own pages disagree: the FAQ says “Voice access is currently available in English, Spanish, French, and German,” while the set-up page lists more. Source: Microsoft Support

superwhisper — local models on Windows, with cloud models too

superwhisper now has a full Windows app: “Superwhisper runs natively on Windows 10 and 11,” and it describes itself as “AI dictation that works in any app on Windows 10 and 11.” Its Windows page is direct about the local option: “On-device AI models. Your audio never leaves your machine and you don't need internet.” Source: superwhisper.com/windows

The nuance is that local is a choice, not the only path. Its model docs say “There are two types: cloud models run on hosted servers with no download, and local models run entirely on your device.” Source: superwhisper docs Its privacy policy, last updated in June 2024, still describes local-only transcription, so treat the model docs as the current word. One thing it does that Dictro does not: its FAQ confirms “Both Windows 10 and Windows 11, x64 and ARM64.” We compare the two apps in more depth in Dictro versus superwhisper.

Handy — open source and fully offline

Handy is the purest local option here, and the easiest to audit because you can read the code. Its README calls it “A free, open source, and extensible speech-to-text application that works completely offline,” and explains: “Press a shortcut, speak, and have your words appear in any text field. This happens on your own computer without sending any information to the cloud.” It “Works on Windows, macOS, and Linux” and is released under the MIT License. Source: github.com/cjpais/Handy

Its README also documents an optional post-processing mode you can toggle on and off, alongside the plain transcription.

OpenWhispr — open source, local by default, cloud if you choose

OpenWhispr runs on Windows as well as macOS and Linux, and states plainly: “Cloud is a choice, not a default.” In local mode, “When you dictate locally, your words are transcribed on your own computer by open models. Nothing is sent to us or to anyone else.” If you opt in, “Switch the cloud on and your audio goes out to be transcribed, then it's gone the moment the words come back.” “OpenWhispr is open source on GitHub under the MIT license.” Source: openwhispr.com

Wispr Flow — cloud transcription, with training and storage controls

Wispr Flow is “Available on Mac, Windows, iPhone, and Android,” and it is the clearest cloud architecture in this list, by its own account: “Transcription always happens in the cloud to provide the best speed and accuracy.” Source: wisprflow.ai/privacy

Its controls are real and worth crediting. One setting lets you “choose whether Wispr may use your Dictation data—audio, transcripts, and edits—to evaluate, train, or improve AI models,” and “Dictation Cloud Storage controls whether Wispr stores Dictation transcripts, audio, and dictation history on its servers for persistent, cross-device features.” Separately, “Wispr always maintains zero data retention agreements with all third-party AI providers.” Source: wisprflow.ai/data-controls

The honest summary: you can stop Wispr Flow storing or training on your dictation, but your audio is still transmitted and recognised off your PC. Whether that matters depends on whether your worry is retention or transmission. For confidential or regulated work it is usually transmission. Dictro versus Wispr Flow works through that trade-off, including what Wispr Flow does better.

Dragon Professional v16 — the veteran, sold through a sales contact

Nuance lists Dragon Professional v16 under “Locally installed productivity,” and says “Dragon Professional v16 is optimized for Windows 11 and backwards-compatible to Windows 10.” Its cloud sibling, Dragon Professional Anywhere, is “cloud-hosted speech recognition for the enterprise and public sector.” Source: dragon.nuance.com

Two things we could not find on Nuance's pages: a statement of where the desktop product's speech recognition actually runs (“locally installed” describes the software, not the processing), and a way for an individual to buy it online — the product page offers a sales contact instead. If you are evaluating Dragon for a practice or a team, ask Nuance the first question directly.

Briefly

Also checked

These come up in Windows dictation searches too. We confirmed from each vendor's own site whether it runs on Windows and where the audio goes.

• Voibe — cloud on Windows. “On Windows, dictation runs in zero-retention cloud mode,” and “Fully on-device mode, Live Dictation, and manual language selection are Mac-only today.” Source: getvoibe.com
• Typeless — cloud. “Transcription is performed on the cloud to ensure the highest accuracy and low-latency performance.” Source: typeless.com
• Willow Voice — cloud on Windows. “Willow uses cloud-based AI models so you can get the most accurate and responsive dictation experience.” Its own blog puts its offline mode on “Mac and iOS” only. Source: willowvoice.com
• SnailText — on-device. It supports “Windows 10/11 and macOS 12+ on Apple Silicon (M1 or later),” and says “We don't currently ship a cloud STT option.” Source: snailtext.app
• Sotto — sometimes listed for Windows, but it is Mac-only: “Sotto requires macOS 13 or later.” Source: sotto.to

Matching the tool to the job

Which one to pick

• You want live dictation, clean text, and no cloud mode to worry about: Dictro. There is no setting that sends audio away, because there is no cloud path.
• You want to read the code, not trust a promise: Handy or OpenWhispr, both MIT-licensed.
• You want the most features, or you are on an ARM64 PC: superwhisper, with a local model selected.
• You want to control your whole PC by voice: Voice access, built into Windows 11.
• You want free and built in, and cloud is fine: Windows voice typing on Win + H.
• You want polished rewriting and accept cloud transmission: Wispr Flow, with model training and cloud storage switched off.
• You are an organisation with an IT department: Dragon, after asking Nuance where recognition runs.

Trust, then test

The two-minute test that settles it for any app

Every claim on this page is a quote from a vendor. This check needs no vendor at all, and it works on any Windows dictation app.

1. 1

   Turn off the network

   Press Win + A and switch on Airplane mode, and unplug any network cable.

2. 2

   Dictate into a local app

   Open Notepad, put the cursor in it, and dictate a sentence with the app you are testing.

3. 3

   Read the result

   Text with no connection means recognition ran on your PC. An error or an empty page means it needed a server.

4. 4

   Repeat with your real settings

   For apps with optional cloud models, test the setup you actually use, not the default you installed.

On a Mac instead? The same audit for macOS is private dictation apps for Mac, compared. Every page we publish is listed in guides and comparisons.