What local really means in a video to text converter

The question that blocks this work is never whether a tool can transcribe. It is where the file goes. Somebody in IT asks it, and the answer decides whether you are allowed to use the tool at all.

I got asked it on a call about a set of internal training recordings. I said nothing is uploaded, which was true, and it changed nothing, because every transcription site on the internet says that on its homepage. What ended the conversation was opening Resource Monitor and letting him watch the app for the length of a job.

What does local actually mean when an app says nothing is uploaded?

Local is a claim about your media, not about the program and its internet connection. It means the audio and video are read from your disk, decoded by your own processor, and turned into text by a model that is also sitting on your disk. Nothing about the content leaves the machine.

It does not mean the program never opens a connection. Almost every local transcriber still needs the network for three things: fetching the speech model the first time, checking whether a newer version exists, and confirming a licence. All three are about the software, not about your video.

That gap is worth taking seriously. A reviewer who was promised fully offline and then sees any outbound traffic at all will assume the worst, and they are right to. It is far better to list the traffic than to claim there is none.

What does VideoDoc contact, and what does it send?

Here is the whole list for my own tool. It is a table because this is the part people actually want to paste into an email.

The model arrives from the Hugging Face Hub the first time you transcribe, which is simply how faster-whisper works: you name a model size, the matching model is downloaded once, and every run after that loads it from your own cache. Once it is cached you can transcribe local files with the network switched off.

One detail is worth stating plainly, because it usually gets assumed the other way round. The fingerprint that ties a Pro activation to a particular computer is calculated and stored on that computer. It is checked locally against the cached activation. It is not sent to anybody.

And the two rows that would break the promise are missing on purpose. Your video is never sent. Your transcript, PDF, Markdown and saved frames are never sent.

How can you prove nothing was uploaded, without buying anything?

This is the free method, and it is the one I would run against any tool including mine. It needs no software you do not already have and it takes about five minutes.

1. Copy the video onto the machine itself, so no network drive is involved in the test.
2. Run one short transcription normally, with the internet on. This is what downloads and caches the speech model.
3. Turn on airplane mode, or unplug the cable, and wait until Windows admits there is no connection.
4. Transcribe a different local file. If it finishes, transcription itself needed nothing from the network.
5. Turn the network back on. Press the Windows key, type Resource Monitor, open it, go to the Network tab and sort by Send.
6. Start a third job and watch the row for the app. Compare the bytes sent against the size of your video.

Step six is the one that ends arguments, because it produces a number instead of a promise. An upload has to move the whole file. A 400 MB recording cannot hide inside a few kilobytes of traffic.

Where it stops, honestly: this proves what happened on your machine, on that day, with that version. It is not a code audit and it says nothing about what a future release will do. For work that is genuinely regulated you also want a written statement from the vendor, and you want to pin the version you tested.

Is there a genuinely free local converter, and where does it stop?

There are two, and neither costs anything. Both are real answers rather than trial buttons.

• A browser tab that transcribes on your device. The model is fetched once and then runs inside the tab, so the audio is never posted anywhere. You can transcribe a video free without uploading it this way. It stops at long files, because the tab has to stay open and your machine has to hold the model in memory for the whole job.
• Whisper on the command line. Free, open source, entirely on your own hardware, and the same transcript quality. It stops at setup: you need Python, a working ffmpeg, and a working CUDA install if you want it finished this afternoon rather than this evening.

Both of those give you the words. Neither gives you the screen. On a screen recording or a slide deck the screen is usually the half that carries the meaning, which is why my own tool captures frames alongside the transcript.

What do you actually tell the person doing the security review?

Give them four sentences and a version number. The media never leaves the machine. The app contacts these hosts, for these reasons, and none of them receive content. Here is how I tested that, and here is the date I tested it.

That reads as an engineer talking rather than as marketing, and in my experience it is the only version that gets approved quickly. Vague reassurance gets escalated to somebody more senior. A list gets ticked off and the ticket closes.

If your policy forbids outbound connections outright, say so at the start rather than at the end. The whole thing runs that way once the model is cached, but somebody still has to allow that one first download.

Quick questions

Pick the recording you have been avoiding because of the upload question, run it once with Resource Monitor open, and you will have an answer ready the next time somebody asks.