At a glance
| WhisperDesktop | |
|---|---|
| Made by | Const-me (open source, MPL-2.0) |
| Latest release | 1.12.0, July 2023; no updates since |
| Platform | 64-bit Windows only |
| GPU | Any Direct3D 11 GPU: NVIDIA, AMD or Intel |
| Models | ggml files from whisper.cpp; medium is the tested one |
| Speaker labels | ✗ |
| Price | Free |
Download it from GitHub, not whisperdesktop.com
The only official download is the Releases page of the Const-me/Whisper repository on GitHub. The author warns that whisperdesktop.com impersonates the project: it links to GitHub today, but could swap those links for something else at any time.
- Open github.com/Const-me/Whisper, go to Releases and download WhisperDesktop.zip from release 1.12.0.
- Unpack the ZIP and run WhisperDesktop.exe. There is no installer.
- On the first screen, point it at a model file. Download ggml-medium.bin from the ggerganov/whisper.cpp repository on Hugging Face.
- Pick an audio or video file and transcribe, or use the capture screen to transcribe from a microphone.
Which model file to use
WhisperDesktop reads the ggml model files published for whisper.cpp. The author tested mostly with ggml-medium.bin (about 1.4 GB), and that is the safe choice. ggml-small.bin (about 500 MB) and ggml-base.bin (about 150 MB) are faster on weak GPUs but make more mistakes.
Files with q5 or q8 in the name are compressed versions made after WhisperDesktop's last release. If one fails to load, use the plain file instead. The large-v3 and large-v3-turbo files are covered in the next section.
Errors people hit
- "failed to generate timestamp token" with large-v3 or large-v3-turbo. WhisperDesktop predates these models. The support requests are still open on GitHub. Use ggml-medium.bin, or a tool that supports the newer models.
- It won't start on an older PC. The CPU needs AVX and F16C instructions, and the GPU needs Direct3D 11. Nearly every PC from the last ten years has both.
- An .ogg file won't open. Audio is read through Windows Media Foundation, which handles most formats but not Ogg Vorbis. Convert the file to MP3 or WAV first.
- The command-line version ignores my model. Its default path is models/ggml-base.en.bin (an open issue on GitHub). Put the file there or pass the model path explicitly.
What it does well, and where it stops
WhisperDesktop's strength is still its GPU support. Through DirectCompute it accelerates on AMD and Intel graphics, not only NVIDIA, and it ships as a small DLL with no Python. On a laptop without an NVIDIA card, it was for years the fastest free way to run Whisper.
It stops at plain text and subtitles. It doesn't say who is speaking, doesn't write summaries or Word documents, and won't get the newer models or fixes. For a single-voice recording that's fine. For interviews and meetings you end up listening again to work out who said what.
Alternatives that are still maintained
| Price | GPU on Windows | Newer models | Speaker labels | |
|---|---|---|---|---|
| Scrieb | Free 1 h/day, Pro paid | NVIDIA, AMD, Intel; native ARM64 | Large v3 Turbo | ✓ |
| Buzz | Free | NVIDIA | ✓ | ✗ |
| whisper.cpp | Free | Depends on the build | ✓ | ✗ |
| Const-me's successor | Paid after 30-day trial | Direct3D 12, Vulkan | Voxtral | ✗ |
- Buzz is the closest free swap: an installer, a window, Windows, Mac and Linux. Text only.
- whisper.cpp is the project WhisperDesktop was ported from, and it is actively developed. It runs from a terminal; see Whisper for Windows for the steps.
- Const-me's commercial successor is a new closed-source app by the same author, with a different speech engine. It is free for 30 days, then a one-time purchase.
- Scrieb is for recordings with more than one voice. It is covered below.
Scrieb: WhisperDesktop's GPU support, plus speakers
Scrieb runs Whisper locally on Windows and Mac like WhisperDesktop does, and also uses AMD and Intel GPUs (through Vulkan and OpenCL). It adds what WhisperDesktop never had: speaker labels, summaries and Word export. There is a native build for Snapdragon and other Windows on ARM laptops.
No model files to fetch
Choose Whisper Small, Medium or Large v3 Turbo in Settings; Scrieb downloads it once and keeps it updated.
Who said what
Speaker identification runs on your PC, in every plan including the free one.
Fewer invented lines
Silence is skipped so Whisper does not make up text in pauses, and suspicious transcripts are flagged.
Documents, not just text
Interview, meeting and lecture summaries; export to DOCX, PDF, TXT or Markdown.
Download Scrieb. The free plan gives 1 hour a day, with no card needed.
Pricing
Free plan: 1 hour a day with speaker labels. Pro: no minute limits, one licence for Mac and Windows.
Frequently asked questions
- Where do I download WhisperDesktop?
- From the Releases page of the Const-me/Whisper repository on GitHub: download WhisperDesktop.zip, unpack it and run WhisperDesktop.exe. The latest release is 1.12.0 from July 2023. The website whisperdesktop.com is not the official project; the author warns that it impersonates it.
- Is WhisperDesktop still maintained?
- No. The author's note on GitHub says the open-source project hasn't been updated since 2023, and points to a separate, closed-source commercial successor. The 1.12.0 release still runs, but reported bugs are not being fixed.
- Which model file should I use with WhisperDesktop?
- ggml-medium.bin (about 1.4 GB), from the ggerganov/whisper.cpp repository on Hugging Face. The author says he mostly tested with that model. ggml-small.bin and ggml-base.bin are smaller and faster but less accurate.
- Does WhisperDesktop support large-v3 or large-v3-turbo?
- Not reliably. Users report "failed to generate timestamp token" errors with large-v3 and large-v3-turbo, and the requests for large-v3 support are still open on GitHub. Use ggml-medium.bin, or an app that supports the newer models.
- Does WhisperDesktop work on Windows 11?
- It needs 64-bit Windows, a Direct3D 11 GPU and a CPU with AVX and F16C. The author says it should work on Windows 8.1 or newer but tested only on Windows 10.
- Can WhisperDesktop identify speakers?
- No. It produces text and subtitles without saying who spoke. For speaker labels you need a diarization step: Scrieb does it in the app, WhisperX does it from the command line.
- What is the best WhisperDesktop alternative?
- For free and text only: Buzz, or whisper.cpp if you are comfortable with a command line. For speaker labels, summaries, Word export and the newer Whisper models without setup: Scrieb, which runs on Windows (x64 and ARM64) and Mac.