How to turn an MP4 into text on Windows, with the internet off

You have an MP4 sitting on your Windows PC and you want the words out of it. Maybe it is a recorded meeting, maybe it is a lecture, maybe it is a client file you are not allowed to put on anybody else's server. So you search for an offline converter, and most of what comes back is a website that wants you to upload the file first, which is the exact thing you were trying to avoid.

Here is the answer before the explanation. Windows can do this entirely on your own machine and the file never leaves it. But almost no offline transcriber is actually offline on the very first run, because it has to fetch its speech model once. Run one short video with the internet on. After that you can pull the cable out and it still works.

That first download catches people out, and it caught me out on my own tool. I emptied the model folder to see what a brand new install feels like, turned the wifi off and started a job, and the app just sat there waiting, because there was nothing on disk to listen with. It is not a bug and it is not a lie in somebody's marketing. It is simply how these models get onto your computer in the first place.

Why does an offline transcriber still need the internet once?

The program that reads speech is small. The model it reads with is not. A speech model is a big file of learned weights, which you can think of as the part that actually knows what English sounds like, and no installer wants to carry half a gigabyte of that inside the setup file for every user who might never transcribe anything.

So the normal design is to download the model the first time you need it and then keep it in a cache folder forever. VideoDoc uses faster-whisper, which pulls the model from the Hugging Face Hub on first use and loads it from your own disk on every run after that. Once that file is cached, transcribing a video already on your drive needs nothing from the network at all.

That is the honest shape of the word offline on Windows. One download, then nothing. If your rule is that the machine must never make an outbound connection, including the first run, then somebody has to carry the model onto that machine another way, and no tool clears that bar on its own.

What does the first download actually cost in megabytes?

This is the part almost nobody publishes, so here are the real figures from the model cache on my own Windows machine. VideoDoc offers five sizes and defaults to small, which is the only one most people will ever download.

Read the small row twice, because that gap is real and it surprised me. The model file is 484 MB and the folder it lives in came to 973 MB, because the cache kept a second copy of the same file. Two of the four sizes I have on disk did that and two did not, so on a small SSD the safe plan is to budget double the download and be pleasantly surprised.

How long does an hour of video take on a Windows PC?

Nobody can give you one number for this, and any page that gives you one is guessing about your computer. What I can give you is a published measurement and the two things that move it.

The faster-whisper project benchmarks 13 minutes of audio with the small model on an Intel Core i7-12700K, using 8 threads and int8, and it finishes in 1 minute 42 seconds. That is roughly seven and a half times quicker than the audio is long, so an hour of video lands at about eight minutes on a desktop processor like that. A thin laptop with four cores will take two or three times longer, and that is normal rather than broken.

Two levers change it. The first is the model size, because medium is far slower than small and you only need it when the audio is genuinely hard. The second is your graphics card: VideoDoc runs the transcription on an NVIDIA GPU when the CUDA runtime is installed, and that path is roughly 5 to 10 times faster than the processor, which is the difference between an evening job and a coffee break.

One thing that gets quoted at you and does not apply here. On a YouTube link VideoDoc takes the caption track that already exists and is finished in seconds, with no transcription at all. An MP4 sitting on your disk has no caption track, so a local file always takes the slow honest path. If you have been told this kind of tool is instant, that is why, and it is not the same job.

Is there a free way to do it, and where does free stop?

There are two, and both are real answers rather than trial buttons that expire.

1. Whisper on the command line, or faster-whisper directly. Free, open source, runs entirely on your own hardware, and the transcript is the same quality. It stops at the setup: you need Python, a working ffmpeg on your PATH, and a working CUDA install if you want the GPU speed. The first time, that is an afternoon.
2. A browser tab that runs the model on your device. Nothing is posted anywhere and there is nothing to install. It stops at long files, because the tab has to stay open and your machine has to hold the whole model in memory for the length of the job.

Both of those give you the words. Neither gives you the screen, and on a screen recording the screen is usually the half that carries the meaning.

What does a transcript miss when the MP4 is a screen recording?

Transcription listens, and it never looks. If your MP4 is somebody demonstrating software, the sentence in the audio is "and then you click here", and here is a button that only exists in the picture. The transcript comes out correct and useless at the same time, which is a frustrating thing to discover after waiting eight minutes for it.

That gap is the reason I built VideoDoc to save timestamped pictures of the screen next to the transcript and put both into one PDF or Markdown file. The words carry the explanation, the frames carry the screen, and an AI reading that document can use both. It runs on your own machine on Windows and on Apple Silicon Macs, and it is $19 once for two computers.

Quick questions

Pick the shortest recording you have and run it once tonight with the internet on, so the model lands. From tomorrow that PC can do this with the cable out.