Whisper on Windows without installing anything
No Python, no pip, no CUDA toolkit. whisper.cpp downloads itself on first use and uses your NVIDIA card if you have one.
Updated: September 2026 · 7 min read
The easiest way to use Whisper on Windows is not a tutorial: install AppGrabbit, open Transcript, paste a video link or drop an audio/video file and click Transcribe. The first time, the app downloads whisper.cpp (the official Windows build, CUDA 12.4 if you have an NVIDIA card, CPU otherwise) and the model you pick — small (181 MB) or large-v3-turbo (547 MB) — verifies both, tests them on your machine and carries on. No Python, no environment, no command line. You get a clean .txt with a prompt on top for ChatGPT, Claude or Gemini, plus an .srt. 1 credit · 5 free a month
Why the usual way is painful
Search "whisper windows" and you get tutorials: install Python, create a virtual environment, pip install openai-whisper, install FFmpeg and put it on the PATH, then, for the graphics card, the CUDA toolkit and the right PyTorch build. It works, and if you like that world it is fine. But most people just want the text of a recording, and a wrong PyTorch build can cost an evening. The desktop apps that wrap whisper.cpp fix this, but they leave you to hunt for model files, and several cap the free tier by words per day.
What AppGrabbit does under the hood
Exactly what is downloaded and where it lives
- Engine — whisper.cpp
b5130, the official Windows builds: CPU (8 MB) for every PC, or CUDA 12.4 (643 MB download, 1.1 GB on disk) when an NVIDIA card with driver 551.61 or newer is detected. Onlywhisper-cli.exeand its DLLs are installed. - Models —
ggml-small-q5_1(181 MB) as Fast andggml-large-v3-turbo-q5_0(547 MB) as Accurate, plus the Silero voice-activity model (under 1 MB) that skips music and silence. - Integrity — every file is mirrored in our GitHub releases and checked by size and SHA-256 before it is used; then a test on a 3-second audio clip confirms it actually runs on your machine. If the CUDA build fails the test, the app falls back to CPU and says so.
- Where — the app's
tools\whisperfolder. Remove or switch models any time in Settings › Transcription; uninstalling removes everything.
How fast, how accurate
| Model | With an NVIDIA card | Processor only | Words wrong |
|---|---|---|---|
| Fast · small (181 MB) | ≈ 30 s per 10 min | ≈ 2 min per 10 min | ≈ 8 in 100 |
| Accurate · large-v3-turbo (547 MB) | ≈ 30 s per 10 min | ≈ 6 min per 10 min | ≈ 5 in 100 |
These are the reference figures the app itself shows before it starts: the times come from a laptop-class RTX 3050 and its processor, and the error rates are the published figures for each model. Your hardware will move the times; the app shows the expected time for your PC and for that exact video before you confirm. With an NVIDIA card, Accurate is as fast as Fast and better, so it is the recommended one; without a GPU, Fast is the sensible default.
What you don't get (on purpose)
AppGrabbit is not a Whisper front end. There is no live dictation from the microphone, no speaker diarization, no translation, no choice of beam size or temperature. It transcribes videos from a link (YouTube, TikTok, Instagram and more) and your own files, and hands you a text file with a prompt on top so your AI does the summarizing. If you need knobs, whisper.cpp itself is the tool; if you need the text, this is the shortest path.
AppGrabbit vs the alternatives
| AppGrabbit | Python + openai-whisper | Desktop whisper.cpp apps | Whisper web services | |
|---|---|---|---|---|
| Setup | One installer; engine + model download themselves | Python, pip, FFmpeg, CUDA toolkit, PyTorch | Download app, then find model files | None |
| GPU | ✓ NVIDIA, automatic | ~ If PyTorch matches your CUDA | ~ Varies | ✓ Theirs |
| Offline / private | ✓ Yes | ✓ Yes | ✓ Yes | ✗ You upload |
| From a video link | ✓ Paste and go | ✗ Download first | ✗ Download first | ~ Some |
| Prompt for your AI | ✓ On top of the .txt | ✗ | ✗ | ✗ |
| Free tier | 5 credits a month, no minute cap | Unlimited | Varies; some cap words per day | Minutes |
Step by step
- 1Install AppGrabbit — 109 MB, Windows 10 and 11. Windows will say it protected your PC because the installer is not code-signed: More info › Run anyway.
- 2Open Transcript and give it something to transcribe — Paste a video link, or drop an mp4, mkv, mp3, m4a, wav, ogg or opus file.
- 3Let it get ready, once — Pick Fast or Accurate (Accurate is recommended with NVIDIA) and click Set up and transcribe. It downloads the engine and the model, checks them and tests them.
- 4Click Transcribe — Progress shows Audio · Transcribing · Saving. You can cancel at any step.
- 5Manage it in Settings › Transcription — See which backend runs (your card or the processor, and why), switch models, remove them, or re-check the card after a driver update.
Frequently asked questions
Any card with compute capability 5.0 or higher and driver 551.61 or newer (CUDA 12.4). Older cards or drivers fall back to the processor, and the app tells you which case you are in.
It runs on the processor instead. whisper.cpp has no official Windows build for those cards today; the app says so plainly in Settings.
Not from the interface: the app manages the two models it ships (small and large-v3-turbo) so that every download is verified. If you need other models, use whisper.cpp directly.
No. The audio, the model and the text stay on your PC. The app only contacts its own server for your license, the free credits and anonymous usage counters.
Into the app's tools\whisper folder, and the transcripts into your Transcripts folder. Settings › Transcription removes models; uninstalling removes the whole folder.
1 credit per transcript. Free comes with 5 credits a month and 1 more to claim every 48 hours when you run out; Pro (USD 8.99 a month) or the 7-day pass (USD 3.99) remove credits.
Related guides
Whisper on your PC, with none of the setup
Windows 10/11 · whisper.cpp with GPU · 1 credit per transcript, 5 free a month