On-device vs cloud speech recognition on Windows

Two places speech can become text

Every dictation tool has to turn sound into words somewhere. There are two places it can happen:

• In the cloud (online). Your audio is sent over the internet to a server, and the text comes back.
• On the device. A speech model is installed on your PC, and the audio is turned into text there.

Microsoft uses the same split in Windows. It describes device-based speech recognition as one that “processes your voice locally on your device”, with no voice data sent to Microsoft. Online speech recognition “uses Microsoft cloud-based services for more accurate recognition”, and voice data is sent to Microsoft to provide the service and create the transcription.

Source: Microsoft Support, “Speech, voice activation, inking, typing, and privacy”, checked 23 Oct 2026.

Which Windows tools use which

• Voice typing (Win+H) uses online speech recognition, powered by Azure Speech services, and needs an internet connection.
• Fluid dictation, a voice typing feature on Copilot+ PCs, corrects grammar, punctuation and filler words using small language models on the device. Microsoft still describes voice typing itself as using online speech recognition.

Source: Microsoft Support, “Use voice typing to talk instead of type on your PC”, checked 23 Oct 2026.

• Voice access, in Windows 11 version 22H2 and later, lets you control your PC and write text by voice “without an internet connection”, according to Microsoft. It is built for controlling the whole PC, so it works differently from Win+H.

Source: Microsoft Support, “Get started with voice access”, checked 23 Oct 2026.

In the browser it depends on the tool. For Google Docs voice typing, Google says your web browser controls the speech-to-text service. For Chrome's talk to type, Google says Chrome collects your voice recording, typed words and page details. Details for both are in voice typing in Gmail and Chrome.

Source: Google Docs Editors Help, “Type & edit with your voice”, checked 23 Oct 2026.

Source: Google Chrome Help, “Use talk to type in Chrome”, checked 23 Oct 2026.

For more on what Win+H does without a connection, read does Win+H voice typing work offline?

What changes between the two

• Connection. Cloud recognition needs a working connection while you speak. On-device recognition can work without one once the model is installed.
• Where your audio goes. With cloud recognition, the audio travels over the internet to the provider. Its privacy page says what it keeps and for how long. With on-device recognition, the audio is turned into text on your PC, though the app may still connect for other things, such as updates or a licence check.
• Download and disk space. On-device recognition needs the model on your disk, so there is a download first.
• Your PC. On-device speed depends on your hardware and the model. Cloud speed depends more on your connection.
• Languages. The two can cover different languages, so check the list for the mode you plan to use.

How to check what a tool does

• Read the maker's own description of where speech is processed, and its privacy page.
• In Windows, Start > Settings > Privacy & security > Speech has the Online speech recognition switch. Microsoft says that when it is off, only device-based features (such as Narrator) are available.
• Test it: try the tool with the network off. If it still types, the recognition is on the device.

Source: Microsoft Support, “Speech, voice activation, inking, typing, and privacy”, checked 23 Oct 2026.

Where VoltType fits

VoltType has both. In the cloud, your voice is sent over the internet to be transcribed. The provider is listed on how the AI works. On Pro, in Offline mode, your voice is transcribed on your PC.

• Offline mode is part of Pro and covers 25 languages. Pro covers all supported languages in total; the other 74 go through the cloud. Which 25 is on offline dictation in other languages.
• The voice pack is a one-time download of about 670 MB. The model runs on the processor, so no special graphics card is needed.
• A few things can still go out in Offline mode, such as the subscription check (at most every 6 hours) and the daily update check if automatic updates are on. The app shows a “Sent since Offline” counter, and the verify page shows how to check with Windows' own tools.
• Spoken commands such as “comma” and “new line” work in both modes. See punctuation commands.

Hold Ctrl + Space to talk, or press Ctrl + Shift + D to start and stop. How Offline mode works step by step is in offline dictation on Windows, and a side-by-side is in VoltType vs cloud dictation. Plans are on the pricing page.

Sources

• Microsoft Support, “Speech, voice activation, inking, typing, and privacy”: support.microsoft.com, checked 23 Oct 2026.
• Microsoft Support, “Use voice typing to talk instead of type on your PC”: support.microsoft.com, checked 23 Oct 2026.
• Microsoft Support, “Get started with voice access”: support.microsoft.com, checked 23 Oct 2026.
• Google Docs Editors Help, “Type & edit with your voice”: support.google.com, checked 23 Oct 2026.
• Google Chrome Help, “Use talk to type in Chrome”: support.google.com, checked 23 Oct 2026.

Frequently asked questions

What is the difference between on-device and cloud speech recognition?

Cloud recognition sends your audio over the internet to a server that returns the text. On-device recognition runs a speech model on your own PC, so the audio is turned into text there.

Does Windows voice typing (Win+H) run on the device?

No. Microsoft says voice typing uses online speech recognition, powered by Azure Speech services, and needs an internet connection. On Copilot+ PCs, Fluid dictation cleans up the text with small language models on the device.

Which Windows tool works without the internet?

Microsoft says voice access, in Windows 11 version 22H2 and later, lets you control your PC and write text by voice without an internet connection.

How can I tell whether a dictation app works on the device?

Read the maker’s description and privacy page, then try it with the network off. If it still types, the recognition is happening on your PC.