TL;DR: WhisperDesktop by const-me is a free, native Windows GUI for Whisper that runs on any Direct3D 11 GPU, including AMD and Intel. Download
WhisperDesktop.zipfrom the GitHub releases page, grab a GGML model file, point the app at it. It still works. It also hasn’t been updated since 2023, can’t load large-v3, and has no real dictation mode. For files, Buzz or Easy Whisper UI are the maintained replacements. For dictation, that’s Invoke.
What WhisperDesktop actually is
The Const-me/Whisper project is a from-scratch C++ port of OpenAI’s Whisper that runs the model on the GPU through DirectCompute instead of CUDA or PyTorch. WhisperDesktop is the small Windows app that ships with it: a file picker, a model picker, a language dropdown, and a transcribe button. Output goes to text, SRT, or WebVTT.
Two things made it popular in 2023. It was fast, often faster than the Python reference implementation by a wide margin, and it ran on AMD and Intel graphics cards when nearly everything else needed an NVIDIA card. If you searched “whisper desktop” back then, this was the answer.
Download and setup
There is no installer and no store listing. The steps:
- Open the Releases page and download
WhisperDesktop.zipfrom the latest release. - Unzip it anywhere and run
WhisperDesktop.exe. Windows SmartScreen will complain because the binary isn’t signed; that’s expected for this project. - Download a model. The app wants GGML-format files, the same ones whisper.cpp used at the time. The usual source is the ggerganov/whisper.cpp model repo on Hugging Face.
ggml-medium.binis the sensible default for English;ggml-large.bin(v1 or v2) if you have 4 GB or more of VRAM to spare. - In the app, point the model path at the
.binyou downloaded, pick your GPU in the advanced settings if it doesn’t auto-select, and transcribe.
That’s the whole setup. On first run it compiles shaders for your GPU, so the first transcription is slower than the rest.
What still works well
AMD and Intel GPUs. This is the reason to still consider it. CUDA-based tools (faster-whisper, Whisper-Standalone-Win, Invoke’s Windows build) only accelerate on NVIDIA. WhisperDesktop accelerates on anything with Direct3D 11, which includes integrated Intel graphics and every Radeon from the last decade.
Zero dependencies. No Python, no CUDA toolkit, no environment. A zip file and a model.
Batch mode. It can transcribe a folder of files in one go, which is more than some newer GUIs offer.
What doesn’t
No updates since 2023. The repo isn’t archived but it’s effectively frozen. Issues pile up, nobody merges.
No large-v3. Whisper large-v3 changed the mel spectrogram from 80 to 128 bins, and WhisperDesktop’s engine predates that. It loads large-v1 and large-v2 fine. If v3’s accuracy on accented or noisy audio matters to you, this is a hard stop.
No proper dictation. The Capture mode listens to a microphone and prints text into its own window. There’s no global hotkey, no push-to-talk, and nothing lands at your cursor in another app. If your goal is “hold a key, talk, text appears in my editor or terminal,” this isn’t that tool, and neither is any other file-transcription GUI.
Stale ecosystem. Newer quantized models, speaker diarization, word-level timestamps from later whisper.cpp releases: none of it reaches this app.
What to use instead, by job
| You want to… | Use | Why |
|---|---|---|
| Transcribe files with a GUI, any GPU | Easy Whisper UI | Maintained, Vulkan acceleration, current models |
| Transcribe files with a GUI, simplest possible | Buzz | Drag, drop, done |
| Batch-transcribe a folder as fast as possible on NVIDIA | Whisper-Standalone-Win | faster-whisper with CUDA, no Python |
| Script your own pipeline | whisper.cpp or faster-whisper | See whisper.cpp vs faster-whisper |
| Talk to your computer all day | Invoke | Push-to-talk, GPU, pastes at your cursor |
The full landscape, including the tools I didn’t list here, is in the best Whisper desktop app for Windows.
Invoke is $49 once. Free 7-day trial, all features, no credit card.
If you came here for dictation
A lot of people search “whisper desktop” because they want to stop typing, not because they have a folder of podcast recordings. WhisperDesktop was never built for that, and the gap is what I built Invoke to fill: hold a hotkey, speak, release, and the transcription lands wherever your cursor is. It runs faster-whisper on your NVIDIA GPU, or on the M-series GPU on a Mac, and it’s sub-second on a mid-range card. It can also reformat what you said into a clean prompt using your project’s context, which is the part that matters if you spend your day talking to Claude Code or Cursor.
If you’re on an AMD or Intel GPU on Windows, Invoke falls back to CPU. It works, it’s just not instant. That’s the honest trade-off, and WhisperDesktop remains the better pick for file transcription on those cards.
Quick answers
Where do I download WhisperDesktop? From the Releases page of the Const-me/Whisper repository on GitHub. The file is WhisperDesktop.zip. It is not on the Microsoft Store and there is no installer, you unzip it and run WhisperDesktop.exe.
Which model file does WhisperDesktop need? A GGML model in the older whisper.cpp format, such as ggml-medium.bin or ggml-large.bin, downloaded from the ggerganov/whisper.cpp repo on Hugging Face. It does not load the newer large-v3 models.
Is WhisperDesktop still maintained? No. The last release was in 2023 and the repo has been quiet since. It still runs fine on Windows 10 and 11, but it will never get newer models or bug fixes.
Does WhisperDesktop work on AMD and Intel GPUs? Yes, and that is its best remaining feature. It uses DirectCompute, so any Direct3D 11 GPU accelerates it, not just NVIDIA cards with CUDA.
Can WhisperDesktop do live dictation? Only in a limited way. Its Capture mode transcribes from a microphone into its own window, but there is no push-to-talk hotkey and nothing pastes at your cursor. For that job use a dictation app like Invoke.
What should I use instead of WhisperDesktop? For file transcription with a GUI, Buzz or Easy Whisper UI. For the fastest batch transcription, Whisper-Standalone-Win. For talking to your computer, Invoke.
Try it
Keep WhisperDesktop around for the odd AMD-accelerated file job. For everything you’d say out loud: Local Whisper on your GPU, Windows and Mac, $49 once. Free 7-day trial, no credit card.