Whisper is a model, not an app
Whisper is OpenAI’s open speech-to-text model, and it’s genuinely good — better than most cloud dictation services, and it runs entirely offline. The problem is that “Whisper” ships as a Python library and a set of model weights. There’s no install button. If you search for a “Whisper desktop app for Windows,” you’re really asking: who wrapped this thing in something I can actually use?
Here are the honest options.
Option 1: Run it yourself
If you’re comfortable in a terminal, you can run Whisper directly. faster-whisper and whisper.cpp are both excellent and free. You pip-install, point them at an audio file, and get a transcript.
The catch is that this gives you transcription, not dictation. There’s no hotkey, no live recording, no pasting into the app you’re working in. You’re transcribing files, not talking to your computer. For a one-off, it’s perfect. For an all-day workflow, you’d be building the app around it yourself — which is exactly what I ended up doing.
Option 2: A desktop app that wraps it
This is what most people actually want: hold a key, talk, and have the text appear where your cursor is. A few apps do this, but the Windows landscape is thin.
| App | Local? | Platform | Price | Built for devs? |
|---|---|---|---|---|
| Invoke | Yes (GPU) | Windows | $49 once | Yes |
| SuperWhisper | Yes | Mac-first, limited Windows | Subscription / lifetime | Partly |
| MacWhisper | Yes | Mac only | Free / one-time | No |
| Talon Voice | Yes | All | Free (Patreon) | Power users |
| Cloud dictation apps | No | Mac/Win | Subscription | No |
Most of the polished local-Whisper apps are Mac-first. On Windows, you’re choosing between Mac apps with a half-finished Windows port, or power-user tools like Talon that have a steep learning curve. (I compared Talon vs Invoke and the SuperWhisper Windows situation separately.)
Invoke is $49 once. Free 7-day trial, all features, no credit card.
What I built and why
Invoke is faster-whisper wrapped in a Windows app that’s actually designed for the push-to-talk loop. You bind a hotkey — Mouse5, Ctrl+Shift+D, whatever — hold it, speak, release. It records while held, transcribes locally on your GPU, and pastes at your cursor. That’s the whole loop, and it’s sub-second on an NVIDIA card.
On top of transcription it can reformat what you said into a clean prompt using your project context. That part is built for developers dictating into AI tools, and it’s the reason I made it instead of just using faster-whisper from a script.
GPU matters more than you think
This is the thing people miss. Whisper’s accuracy is the same on CPU and GPU — but the speed isn’t close. On CPU, a few seconds of speech takes a few seconds to transcribe, and that lag breaks the flow. On an NVIDIA GPU with CUDA, it’s effectively instant.
Invoke ships the CUDA runtime bundled, so if you have an NVIDIA card it just uses it — no separate CUDA install. No GPU? It falls back to CPU and still works, just slower. More on the local-GPU setup in local voice to text for coding.
Try it
If you want Whisper on Windows without building the app yourself: 7-day free trial, no credit card. Download here.