What OpenAI Whisper actually is
OpenAI's repository describes Whisper as a general-purpose speech recognition model that can also do speech translation and language identification. The code and model weights are released under the MIT License.
The official package is a Python library and command-line tool. Its own example transcribes audio files, as in whisper audio.mp3 --model turbo. On its own, it has no microphone shortcut and doesn't type into Word, Slack or your browser. That is the job of the app or script you put around it.
Route 1: the official Python package on Windows
This route suits you if you want to transcribe files or script your own workflow.
- Install Python. OpenAI says the code was trained and tested on Python 3.9.9 and is expected to work with Python 3.8–3.11.
- Run pip install -U openai-whisper.
- Install FFmpeg. The README gives choco install ffmpeg (Chocolatey) or scoop install ffmpeg (Scoop) for Windows.
- If pip fails on the tiktoken dependency, the README says you may also need Rust and pip install setuptools-rust.
- For an NVIDIA GPU, use the CUDA build of PyTorch from PyTorch's install selector. It's a separate command, and PyTorch notes that ROCm is not available on Windows.
Which Whisper model fits your Windows PC
OpenAI lists six model sizes. Four also come in English-only .en versions. The memory column is OpenAI's approximate VRAM figure, which matters most if you run on a graphics card.
| Size | Parameters | English-only | Multilingual | Required VRAM (approx.) |
|---|---|---|---|---|
| tiny | 39 M | tiny.en | tiny | ~1 GB |
| base | 74 M | base.en | base | ~1 GB |
| small | 244 M | small.en | small | ~2 GB |
| medium | 769 M | medium.en | medium | ~5 GB |
| large | 1550 M | — | large | ~10 GB |
| turbo | 809 M | — | turbo | ~6 GB |
OpenAI notes that turbo is not trained for translation. If you need non-English speech translated into English, use a multilingual model such as medium or large.
Route 2: whisper.cpp builds for Windows
whisper.cpp is a separate C/C++ implementation of the Whisper model, also under the MIT License. It lists Windows as a supported platform with MSVC and MinGW builds.
- Clone the repository, download a model converted to its ggml format (for example base.en) and build with CMake.
- Choose acceleration at build time: -DGGML_CUDA=1 for NVIDIA cards, -DGGML_VULKAN=1 for a Vulkan-capable driver, or OpenVINO for Intel CPUs and GPUs. A plain CPU build also works.
- The whisper-cli example accepts 16-bit WAV files, so convert other recordings with FFmpeg first.
- For live microphone input, the whisper-stream example needs SDL2 and prints its transcript in the terminal. Getting that text into another app is still up to you.
Route 3: a packaged Windows app around a speech model
If you want a shortcut and text in any app without building anything, look at finished dictation apps. Check each one's own documentation for which model it runs, whether it works offline and what it costs. Examples people search for:
- WhisperTyping: a Windows 10/11 voice typing app. Its support page says you press a hotkey (CapsLock by default), speak, and it types the text with a virtual keyboard.
- Whisper Key Local: an open-source project with global hotkeys that transcribes locally and pastes at your cursor. It installs from PyPI (Python 3.11–3.13) or as a Windows .exe, and its default Windows start hotkey is Ctrl+Win.
- Superwhisper: has a Windows app and documents both local and cloud models. See our Superwhisper comparison for its Windows status.
Whisper on Windows vs SayDart: what changes
SayDart is an alternative for one specific job: dictating your own words into the app you are using. It does not use, bundle or download Whisper, so there is no model to choose and no GPU to configure.
| Running Whisper yourself | SayDart for Windows | |
|---|---|---|
| Setup | Python + FFmpeg (+ CUDA PyTorch), or a whisper.cpp build, plus a model download | Install the Windows x64 app (version 0.1.17), sign in, choose a microphone |
| Getting text into apps | Not part of the Whisper CLI; needs a wrapper app or your own script | Press your shortcut (Right Alt by default; change it on the Home screen), speak, and the text goes into the field you were typing in. If it can't, copy and paste it |
| Cleanup and vocabulary | Depends on your wrapper | Auto cleanup tidies dictated text; Dictionary holds project names and technical terms; Transforms (for example Polish or Create an AI prompt) rewrite selected text |
| Questions | Not included | SayDart Assistant, a separate window for questions (Right Alt + Space by default) |
| Offline | Local models can work offline once set up | Requires internet; recognition, cleanup and the assistant run in the cloud. History is stored on your PC |
| Cost | Whisper is free under the MIT License; your hardware and any app have their own costs | Free with an allowance every 7 days; Pro $12/month or $99/year ($8.25/month equivalent, billed yearly), before tax; all current Pro features unlimited |
Choose by the thing you can't compromise on
Must work with no internet, or audio must never leave the PC? Choose a local Whisper setup or an app that runs models on-device, and check every optional cloud feature. SayDart is not the right fit.
Want to transcribe recordings, subtitles or long files? The Whisper CLI or whisper.cpp is built for files. SayDart is for live dictation, not file upload.
Want to talk and have clean text land in Outlook, Slack, Notion or an AI chat, with nothing to set up? That's exactly what SayDart does. Hold Right Alt and talk. Free every 7 days.
Using a terminal, such as Claude Code or Codex CLI? With SayDart, dictate, review, then copy and paste into the terminal.
Try it: your first dictation with SayDart on Windows
- Download the Windows installer from saydart.com. If SmartScreen warns on first run, confirm the file came from saydart.com, then choose More info → Run anyway.
- Sign in and pick your microphone. Your shortcuts are shown and set on the Home screen.
- Click into an email, document or prompt box, press Right Alt, say what you want to write and stop.
- Review the text. Add recurring names to Dictionary, or select the text and open Transforms if it needs reshaping.
Whisper, Wispr Flow, WhisperTyping and SayDart are different things
OpenAI Whisper is a speech model. Wispr Flow is a commercial dictation app. WhisperTyping and Whisper Key Local are separate Windows dictation tools. SayDart is an independent dictation app for Windows, Mac and Linux, and it doesn't use Whisper. If you meant "Whisper Flow", see our Wispr Flow comparison.