The offline-inference challenge¶
YazSes transcribes on your own machine. This page is how you check that claim yourself, in about ten minutes, and report what you saw.
What this demonstrates, and what it does not
This is an offline-inference demonstration: it shows that transcription keeps working with networking disabled. That is a strong, checkable claim and it is the one we make.
It is not proof of perfect privacy, and it is not proof that no software on your computer ever talks to the network. A single run on one machine cannot establish either. If you want the stronger claim, read the code and the privacy statement — and audit the traffic yourself with tcpdump or Little Snitch rather than taking this page's word for it.
The split that makes the demonstration meaningful: installing and downloading a model needs the network; transcribing does not. Step 1 is deliberately online, and everything after step 3 is deliberately offline. Conflating the two is the usual way this kind of demo becomes misleading.
1. While online: install and pre-download the model¶
pipx install yazses # or: uv tool install yazses
yazses setup # system packages, input group, injector
Download the speech model now, while you still have a network. Nothing later will fetch it, and a first run with no model and no network fails for a boring reason that has nothing to do with the claim being tested:
yazses doctor # confirms which model is configured
yazses transcribe --help # loading the CLI is enough to trigger the fetch path
If doctor reports the model is not cached, run one dictation or one yazses transcribe <file> while online to pull it, then re-check.
2. Record what you are testing¶
Please capture these before going offline — a report without them cannot be compared with anyone else's:
yazses --version
yazses doctor # model, backend, session type, injector
python3 -VV
uname -a # or: systeminfo (Windows), sw_vers (macOS)
yazses report collects most of this into one local bundle. It redacts paths and identifiers and never uploads anything — read it before you paste it.
3. Disable networking (reversibly)¶
Pick one, and note which you used. Each is reversible with the command beside it. Prefer the airplane-mode or interface method over firewall rules if you are not comfortable restoring them.
| Method | Disable | Restore |
|---|---|---|
| Desktop toggle | Airplane mode / turn Wi-Fi off in your OS settings | toggle back |
| Linux (NetworkManager) | nmcli networking off | nmcli networking on |
| macOS | networksetup -setairportpower en0 off | ... on |
| Windows (admin) | Disable-NetAdapter -Name "Wi-Fi" | Enable-NetAdapter -Name "Wi-Fi" |
| Unplug | remove the Ethernet cable | plug it back in |
Confirm you are actually offline — an unplugged cable with Wi-Fi still up is the classic false result:
4. Run the two things that matter¶
A dictation. Start the daemon, hold your hotkey, say a sentence you are happy to publish, release:
A file transcription, which exercises the same engine without a microphone:
No sample handy? scripts/download-sample.py fetches one — but run it in step 1, while you still have a network.
5. Record the result¶
Note, for each of the two runs: what you expected, what you got, how long it took, and anything in yazses logs that looked wrong.
Do not paste your dictated text if it is private
We never need your words to diagnose this. Timings, the model name, the daemon state and the log's metadata lines are enough. Use a sentence you chose for publication, or describe the result without quoting it.
6. Restore networking¶
Undo step 3 using the matching command in the table, then confirm:
7. Report it¶
Open an offline-inference report. The template asks for exactly the fields above and nothing else.
Negative results are the valuable ones. If transcription failed with the network off, that is a bug worth more than a hundred confirmations, and it is precisely what this exercise is for. Report it the same way.