Whisper on Windows with nothing to install: how a desktop app ships whisper.cpp
If you've ever tried to run OpenAI's Whisper on a Windows PC, you know the drill: install Python, pip install openai-whisper, discover it's running on the CPU, uninstall torch, reinstall it from the CUDA index, fight wit
If you've ever tried to run OpenAI's Whisper on a Windows PC, you know the drill: install Python, pip install openai-whisper, discover it's running on the CPU, uninstall torch, reinstall it from the CUDA index, fight with versions. Fine for a developer. A wall for the people who actually need transcripts: a student with a recorded class, someone with a two-hour meeting, a journalist with an interview.
I make AppGrabbit, a Windows app that downloads and transcribes videos. When I added local transcription, the rule was: the user clicks one button and it works. No Python, no CUDA toolkit, no terminal, and the audio never leaves the PC.
Here's how it's built, with the real numbers.
1. whisper.cpp instead of Python
whisper.cpp is Georgi Gerganov's C/C++ port of Whisper, and it publishes official Windows builds:
- a CPU build:
whisper-bin-x64.zip, 8.6 MB - a CUDA 12.4 build:
whisper-cublas-12.4.0-bin-x64.zip, 674 MB (the cuBLAS DLLs are most of it; about 1.2 GB unpacked)
No Python runtime, no PyTorch: one executable (whisper-cli.exe) plus its DLLs.
The app pins a single version (b5130). Everything is tested against it, and a VERSION file next to the binaries tells the app when an engine on disk doesn't match, so it gets replaced instead of half-working.
2. Download on demand, verify, swap in atomically
Nothing Whisper-related ships in the installer. It stays small, and people who only download videos never pay for 1+ GB they won't use.
When someone turns on transcription:
- Only what that PC can use is downloaded: the CUDA build if there's a usable NVIDIA card, the CPU build otherwise, plus one model and the voice detector.
- Every file has its SHA-256 pinned in the app. They come from a mirror of the official files in our own GitHub release, and anything that doesn't match is thrown away.
- Disk space is checked first: download size + unpacked size + a 200 MB margin.
- It unpacks into a temp folder and swaps it in, so a crash halfway never leaves a broken engine behind.
-
Most of the zip is dropped. It contains 25 executables (server, stream, talk-llama, benchmarksβ¦). We keep
whisper-cli.exeand only the DLLs it loads. Fewer unknown.exefiles means fewer antivirus false positives.
The models are quantized GGML files:
| In the app | File | Size |
|---|---|---|
| Fast | ggml-small-q5_1.bin |
190 MB |
| Accurate | ggml-large-v3-turbo-q5_0.bin |
574 MB |
| Voice detector | ggml-silero-v5.1.2.bin |
0.9 MB |
3. GPU or CPU, without asking the user
def whisper_backend():
gpu = gpu_info() # nvidia-smi first, WMI as a fallback
if gpu.vendor != "nvidia":
return "cpu" # AMD / Intel run the CPU build
if gpu.compute_cap is not None and gpu.compute_cap < 5.0:
return "cpu" # Kepler and older can't run CUDA 12
if gpu.driver is not None and gpu.driver < 551.61:
return "cpu" # CUDA 12.4 needs this driver on Windows
if settings.get("whisper_cuda_failed"):
return "cpu" # the GPU failed its smoke test once
return "cuda"
Things that bit us:
-
nvidia-smi --query-gpu=name,driver_version,memory.totalgives the real driver version. Compute capability goes in a separate query, because oldernvidia-smiversions don't knowcompute_capand would fail the whole call. - With no
nvidia-smi, the app falls back to WMI (Win32_VideoController) and skips remote-desktop and virtual adapters, which are not real cards. - On some laptops the query takes up to 25 seconds, so the result is cached for an hour and never re-queried while a job is running.
4. A smoke test before the first real job
After installing, the app runs whisper-cli on 3 seconds of silence: a 16 kHz mono WAV written with Python's wave module, so not even FFmpeg is involved.
- If the CUDA build exits with an error, the app installs the CPU build and runs the same test. If that one passes, the GPU is to blame: the app remembers it, stays on the CPU, and deletes the 1.2 GB
cuda/folder. - Only a real non-zero exit from
whisper-clicounts against the GPU. A timeout or a missing file is an environment problem, and the card is not blamed for it. - The test waits if another transcription is using the GPU, because sharing it would fail on memory and look like a broken card.
5. The actual call
FFmpeg, already bundled for downloads, converts the audio to 16 kHz mono WAV. Then:
whisper-cli -m ggml-large-v3-turbo-q5_0.bin -f audio.wav -l auto -t 4 ^
--vad --vad-model ggml-silero-v5.1.2.bin -osrt -ojf -of output -pp
-
-l autodetects the language. The user can also pin Spanish, English or Portuguese. - Threads:
min(8, cores)on the CPU, 4 on the GPU. -
--vadruns the Silero voice detector, which skips music and silence. Without it, large-v3-turbo invented a credits line ("Ana Luiz Revisor: Claudia Sander") over the opening music of a TEDx talk, a known Whisper habit picked up from subtitle training data. -
-osrtwrites subtitles, and-ojfwrites full JSON with per-token confidence. -
-ppprints progress, which the app parses into a real percentage and ETA.
6. Honest numbers
Measured on a laptop: Ryzen 7 6800HS (8 cores), RTX 3050 Laptop with 4 GB, Windows 11. The test was a Spanish TEDx talk (7:19), compared against the human transcript. The word error rate (WER) was measured after normalizing case, accents and punctuation.
| Model | CPU (8 threads) | RTX 3050 | Word error rate |
|---|---|---|---|
| small q5_1 | 81 s | 21 s | 8.5 % |
| medium q5_0 | 208 s | 34 s | 5.9 % |
| large-v3-turbo q5_0 | 276 s | 19 s | 5.1 % |
On an English NASA video (5:31), turbo took 13 seconds on the GPU, at 2.6β3.2 % WER.
What we took from it:
- With a GPU, turbo is both the fastest and the most accurate, so it's the default whenever a usable NVIDIA card is found.
- Without a GPU, small is the only sane default: about 2 minutes per 10 minutes of audio. Turbo on the CPU takes about 6 minutes per 10.
- Medium is never the best choice. It's slower than turbo on the GPU and less accurate everywhere, so the app ships only small and turbo.
- Most of small's Spanish errors are accents and spelling ("ΒΏquΓ© hacΓ©s?" becomes "ΒΏquΓ© haces?"), not invented words. The meaning survives.
7. What it doesn't do
- No speaker labels.
- No translation.
- No live microphone.
- AMD and Intel GPUs run the CPU build for now.
From the user's side
All of the above collapses into one download button in Settings and this screen: paste a link or drop your own file, pick a prompt for your AI if you want one, and you get a clean .txt plus a timed .srt.
If you want to see it in action, here's Whisper on Windows with nothing to install. AppGrabbit is free to start (5 credits a month) on Windows 10 and 11.
Questions about the GPU detection, the VAD setup or the numbers? Happy to answer in the comments.
β CrowSo Studio
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.