Proof that transcription stays on your PC

Every dictation app says it respects your privacy. That is a claim, not evidence. So we measured ours. Below is what the operating system saw while PrivateTranscribe transcribed a recording with a local Whisper model, including the parts that make the test weaker.

0
connections leaving the PC
60
checks during the run
67ms
longest unchecked gap
106 chars
transcript produced

What we did

We transcribed a 313KB audio file with the local base Whisper model on Windows. While it ran, a separate program asked the operating system, roughly every 37 milliseconds, which network connections PrivateTranscribe and its child processes were holding open.

Every connection found was sorted into one of two piles. Connections to somewhere else on the internet, and connections that never leave your machine. The first pile was empty for the whole run, and a real transcript came out the other end.

The connections we did find

There was traffic, and we would rather show it than quietly leave it out.

TCP 127.0.0.1:8178TCP 127.0.0.1:63900

127.0.0.1 is your own computer talking to itself, and both ends of that conversation can show up separately. Local transcription runs the Whisper engine as a second program on your PC, and the app hands it the audio over that address. Nothing on that path touches a network card. If any of these had been an address anywhere else, this page would say so, because the test fails on any connection that leaves the machine.

Where this test is weak

We check repeatedly rather than watching without blinking. In theory something could open a connection and close it again inside one of the gaps between checks. The longest gap in this run was 67 milliseconds, and we publish that number so you can judge the hole rather than take our word that it is small.

An earlier version of this test had a ≈1.3 second blind spot at the exact moment the Whisper engine starts, which is the worst place to be looking away. We found it, moved the checker into its own program so it could not be interrupted, and the worst gap dropped to 67 milliseconds. The mistake is worth mentioning because a test that is not looking during the interesting moment proves nothing.

Check it yourself

You do not have to trust our measurement. Turn off Wi-Fi, or pull the network cable, then dictate normally with a local Whisper model selected. If the text still lands at your cursor, the network was never part of it. It takes about thirty seconds and needs no administrator rights.

Blocking the app in Windows Firewall is the obvious alternative, and it is easy to get wrong. Local transcription runs in a second executable, whisper-server, so a rule aimed only at PrivateTranscribe.exe leaves the engine's network access untouched and the test proves less than it appears to. Disconnecting covers every process at once, which is why we recommend it instead.

What this does not say

This covers local transcription, which is the default. PrivateTranscribe also offers cloud transcription as an option, and if you choose it, your audio goes to that provider, because that is what cloud transcription is. We would rather say that plainly than claim everything is offline.

It is also evidence from one machine, one version and one run, not a universal guarantee. The test ships with the app so it can be run again on other hardware and other versions.

Run details

Date
24 August 2026
App version
0.15.0
Operating system
Windows, build 26200
Whisper model
base
Transcription time
2307ms
Median gap between checks
37ms

audio sha256 bc35c86329b1b14d1472b4b630d8e26474e40bf3c36910dbdb67b63db879e41f

PrivateTranscribe is free, has no word limit, and does not ask for an account, so you can run the airplane-mode test yourself in about five minutes. Download it here, or read how local and cloud transcription actually differ in the private dictation guide.