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

You have an MP4 file on your Windows computer and you need the spoken words out of it as plain text. It might be a recorded team meeting, a lecture, or a client recording you are not allowed to upload to anybody else's server. When you search for an offline converter, most of the results are websites asking you to upload the video first, and that defeats the whole reason you searched.

Windows can do this entirely on your own machine, and the file never leaves your computer. But nearly every offline transcription tool still needs the internet on its first run, because it has to fetch its speech model from the web. Run one short video with the internet on. After that you can pull the cable out and it still works.

That first download catches a lot of people out, and it caught me out while testing my own software. I cleared the model folder to see what a fresh install feels like, turned off my Wi-Fi and started a job, and the app just sat there waiting, because there was no model on disk to listen with. The installer does not carry the model files, so the app has to fetch them once before it can work on its own.

Why does an offline transcriber still need the internet once?

The program that does the transcription is small, but the speech model it depends on is big. A speech model is a large file of learned weights, which is the data that lets the software recognise English speech, and nobody wants to pack half a gigabyte of that into a setup file for users who may never transcribe anything.

The standard solution is to download the model the first time you transcribe something and keep it in a cache folder on your drive. VideoDoc uses faster-whisper, which downloads the model from the Hugging Face Hub on your first run and loads it from your own disk on every job after that. Once that file is in your cache, transcribing a video stored on your drive needs no internet at all.

That is what offline transcription really looks like on a Windows PC. One download, then nothing. If your workplace rule is that the computer must never make an outside connection, even during setup, then somebody has to copy the model files onto that machine by hand, because no tool can get around that first download on its own.

What does the first download actually cost in megabytes?

Most documentation glosses over these file size details, so I checked the numbers inside the model cache folder on my Windows PC. VideoDoc offers five sizes and defaults to small, which is the only one most people will ever download.

Take a close look at the small model numbers, because the disk usage difference surprised me when I measured it. 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?

No single number can predict your transcription time, and any website that gives you one is guessing about your hardware. What I can give you is the faster-whisper project's own published benchmark, and the two things that change 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 faster than the audio plays, 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 things change that number. The first is the model you pick, because medium is much slower than small, so only use it when your audio has heavy background noise or difficult speech. 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 speed claim you may have seen does not apply to a file on your disk. When you paste a YouTube link, VideoDoc takes the caption track the video already has and is done in seconds, without transcribing the audio at all. An MP4 sitting on your drive has no caption track, so your machine has to transcribe the whole recording from start to finish. So if you have been told this kind of tool is instant, that is the reason, and it is not the same job.

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

There are two practical alternatives you can use on your machine. Both are genuine tools rather than limited trial programs that expire after a few days.

1. You can run Whisper or faster-whisper directly from the command line. Both are free and open source, they run on your own hardware, and the transcript is the same quality. The hard part is the setup, because you need a working Python install, ffmpeg on your system PATH, and working CUDA drivers if you want the GPU speed. Getting all of that working the first time usually takes an afternoon.
2. You can also use a browser tab that runs the model on your own machine. Your audio stays private because nothing is sent to a server, and there is nothing to install. The downside shows up on long recordings, because the tab has to stay open the whole time and your machine has to hold the whole model in memory until the job finishes.

Both options get the spoken words out of your file, but neither one captures what is on the screen. That matters on a screen recording, where the screen usually carries most of the meaning.

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

Transcription only listens to the audio, and it never looks at the picture. 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 on the screen. The transcript comes out correct and useless at the same time, which is a frustrating thing to find out after waiting eight minutes for it.

That gap is why I built VideoDoc to save timestamped screenshots next to the transcript and put both into one PDF or Markdown file. The text carries what was said, the screenshots carry what was shown, 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 a short video on your computer and run it once tonight with the internet on, so the speech model downloads. Once the model is on your disk, that PC can do this with the cable out.