Dictation that never leaves your PC.
Most dictation apps stream your voice to a server and bill you by the month for the privilege. Voxmelt does the recognition on your own Windows machine, works with the network cable unplugged, and needs no graphics card to do it. Here is how offline dictation works, how to check that an app is genuinely offline, and what to look for before you commit.
Windows 10 and 11 · free tier, no card · 15-day full trial · one-time purchase, no subscription
“Private” and “offline” are not the same claim
A cloud dictation app can be perfectly well behaved and still be the wrong tool if your work is confidential. The audio goes to a server, gets transcribed there, and you are trusting a policy about what happens next. Offline dictation removes the question instead of answering it: the recognition model runs as a process on your own computer, so there is no upload to have a policy about.
That distinction matters most for the people who ask for it by name: lawyers dictating privileged material, clinicians handling patient notes, journalists protecting sources, researchers under NDA, and anyone working on a machine that is not allowed to talk to the internet in the first place. For that last group an offline tool is not a preference, it is the only category that qualifies.
How to prove an app is really offline
You do not have to take anyone’s word for this, including ours. Turn off Wi-Fi or unplug the cable, open the app, and dictate a paragraph.
If the text appears normally, the model is running on your machine. It is genuinely offline.
If it stalls, spins, or errors, the recognition was happening on a server the whole time.
For a stricter check, leave Windows Resource Monitor or Wireshark running while you speak and watch the app’s outbound traffic. Voxmelt is designed to be boring under that microscope: after setup, dictation and AI cleanup send nothing.
No graphics card required
Offline used to mean “bring your own GPU”. That is no longer true, and it is the single biggest thing that has changed in this category.
Dictation runs on your CPU
The default engine is NVIDIA’s Parakeet TDT v3, an int8 model built to run fast on ordinary processors. English plus 24 European languages, ready in seconds, on a laptop with no discrete graphics.
A GPU is the optional upgrade
If you do have an NVIDIA card, it is not spent on transcription. It runs a local language model that rewrites your raw dictation into an email, a summary, or a commit message before you paste it.
Whisper is still there
Five Whisper sizes from tiny to large-v3 ship alongside Parakeet, on CPU or CUDA, for roughly 100 languages. Switch engines from the Models page whenever the job calls for it.
On our published real-clip benchmark, Parakeet on a CPU wins on fast natural speech at 2.0 percent word error against 3.1 percent for Whisper large-v3 on an RTX 3080 Ti, with no GPU and a roughly four-second load. Whisper large-v3 is still ahead overall, 7.8 against 11.9, which is exactly why both engines ship. See the full method and every number, or read the case for dictating without a graphics card.
Six questions worth asking any offline dictation app
Does the audio leave the machine?
The only answer that matters. Ask whether recognition happens locally or on a server. "We delete your audio" means it was uploaded first.
Does it still work with no internet?
Pull the network and dictate. If the app stalls, spins, or errors, the recognition was never running on your PC.
What hardware does it demand?
Plenty of local tools quietly require a strong NVIDIA GPU. Check whether plain CPU dictation is supported before you buy a graphics card.
Does it do anything after the transcript?
Raw dictation is rarely publishable. The useful question is whether the tool can clean up, reformat or rewrite the text, and whether that step is local too.
Subscription or purchase?
A local app has no per-word server cost to pass on. Recurring pricing for software running entirely on your own hardware deserves a second look.
Can you prove any of it?
Privacy policies are prose. A network monitor is evidence. Prefer tools whose claims you can check yourself in two minutes.
There are other genuinely local options for Windows, and some of them are good. We keep an honest side-by-side comparison here, including the cloud tools most people are switching away from.
Offline dictation on Windows, answered
What does offline dictation actually mean?
It means the speech recognition runs as a program on your own computer instead of on a company server. Your microphone audio is turned into text by your own CPU, so nothing is uploaded, nothing is stored on someone else’s disk, and the app keeps working with no internet connection. Many apps marketed as private are still cloud apps that merely promise to delete your audio afterwards. Those are different things.
How can I verify an app is really offline?
Unplug your network cable or turn off Wi-Fi, then dictate. A genuinely offline app keeps transcribing exactly as before. You can go further and watch it with a network monitor such as Windows Resource Monitor, Wireshark, or Fiddler while you speak: a local app sends no audio anywhere. This is the useful test because it does not rely on trusting anyone’s privacy policy. Voxmelt is built to pass it.
Do I need a graphics card for offline dictation?
Not for dictation. Voxmelt’s default engine is NVIDIA’s Parakeet TDT v3, which runs on any modern CPU, so a laptop with no discrete graphics card works fine. A GPU is optional and only unlocks the local AI text studio, which rewrites, summarises and translates what you dictated. Older offline dictation tools assumed you had a powerful GPU; that assumption is out of date.
Is offline dictation as accurate as cloud dictation?
It is now close enough that privacy is no longer a tradeoff you have to pay for in accuracy. On our published real-clip benchmark, Parakeet running on a CPU wins on fast natural speech, 2.0 percent word error against 3.1 percent for Whisper large-v3 on a GPU. Whisper large-v3 is still ahead overall, 7.8 against 11.9, and it ships inside Voxmelt for exactly that reason. Every number and the method are on the benchmark page.
Does it work on Windows 10, or only Windows 11?
Both. Voxmelt runs on Windows 10 and Windows 11. It installs from the Microsoft Store or as a direct download, and after the first-run setup it needs no internet connection at all.
What is the difference between this and Windows built-in voice typing?
Windows voice typing (Win+H) sends audio to Microsoft servers in its standard mode, so it is not an offline tool in the sense most people mean. Windows 11 also has Voice Access, which can run on-device after you download a language pack, but it is designed as an accessibility and device-control feature rather than a writing tool, and it has no AI cleanup of the resulting text.
Why do offline dictation apps ask to download a model on first run?
The speech model is the part that does the recognition, and it is several hundred megabytes to a few gigabytes. Shipping it inside the installer would make the installer enormous, so most local apps download the model once during setup and then never need the network again. Voxmelt does the same: a small installer, then you pick which engine to install.
Is Voxmelt free?
There is a free tier with 60 minutes of dictation a day and no card required, plus a 15-day trial of everything. Beyond that Voxmelt is a one-time purchase rather than a subscription, so there is no renewal date and nothing to cancel.
Try it with the Wi-Fi off
Free tier with no card, a 15-day trial of everything, and a one-time purchase after that. No subscription, no renewal date, nothing to cancel.