Run Whisper on Windows without touching the command line
Whisper made high-quality speech recognition something you can run yourself: open weights, dozens of languages, no cloud required. The catch was never the model. The setup is what stops people: the reference implementation wants Python, PyTorch and ideally an NVIDIA GPU with CUDA, and the instructions start with pip install.
If that's not your idea of an afternoon, here are the realistic options on Windows.
Option 1: the reference implementation (Python)
The original openai-whisper package works well if you already live in Python. You'll manage a Python environment, FFmpeg, and GPU drivers; on CPU alone the large model is painfully slow. Best for developers who want to script batch jobs.
Option 2: whisper.cpp
A C++ port that runs without Python and uses quantised models, so it's lighter and surprisingly fast on plain CPUs. You still download model files yourself and drive it from a terminal, and it transcribes files rather than live meetings. Excellent for tinkerers, less so as a daily driver for meeting notes.
Option 3: a desktop app that ships the model
This is the "just works" tier: the model, the acceleration layer and the recording pipeline come pre-wired in an installer. LocalTranscript takes this approach on Windows using DirectML, which means GPU acceleration on AMD and Intel graphics as well as NVIDIA. You never touch CUDA or a terminal, and there are no driver versions to match.
What the app adds on top of raw Whisper-class transcription is the part you actually need for meetings:
- Live capture of system audio + microphone, so it transcribes Teams/Zoom/Meet calls as they happen; see the local meeting-transcription guide
- Speaker separation and labels, with voiceprints that never leave the machine
- Custom vocabulary for names, clients and jargon the base model won't know
- Summaries and action items generated locally after the call
- Exports to SRT, VTT, Markdown, TXT and JSON
The complete local app is free and will stay free, with no account, weekly cap or model paywall. Unlimited meetings and the large, highest-accuracy model are included. Everything runs offline, which is the point: your audio stays on your machine.
Which one should you pick?
- You write code and want batch scripts: use reference Whisper or whisper.cpp.
- For live meetings, speakers and summaries with zero setup, download LocalTranscript; you'll be transcribing in about two minutes.
- If a whole team or firm needs managed deployment, contact us about a branded MSI and rollout support.