Run Whisper on Windows.

The official Python package

OpenAI's own code, MIT licensed. It needs Python (the README expects 3.8 to 3.11) and the ffmpeg command line tool, which the README says to install on Windows with Chocolatey or Scoop. The default model is turbo, listed at 809 million parameters and about 6 GB of VRAM. It uses the GPU only when PyTorch can see CUDA. Otherwise it runs on the CPU and gets slow.

The README warns that turbo is not trained for translation. For foreign speech into English text it says to use another multilingual model, such as medium, with --task translate.

scoop install ffmpeg
pip install -U openai-whisper
whisper audio.mp3 --model turbo

whisper.cpp, no Python at all

A C++ port of the same models, also MIT. Its releases page carries prebuilt Windows zips, so you can skip the build: whisper-bin-x64.zip for plain CPU, whisper-blas-bin-x64.zip, and whisper-cublas-12.4.0-bin-x64.zip for NVIDIA cards. Not every tag has them. At the time of checking the latest release, v1.9.4, has no files attached, and its notes link to a build tag, b5130, that does. Unzip, add a model file, and run whisper-cli.exe.

To build it instead you need Git, CMake and a C++ compiler. The README lists MSVC and MinGW as supported on Windows. A .cmd script fetches the model. With Visual Studio's compiler, which the project's own release workflow uses, whisper-cli.exe lands in build\bin\Release. The README's commands, written for Linux and macOS, use build/bin. whisper-cli reads only 16 bit WAV, so convert your audio first with the README's ffmpeg line, the last one below.

git clone https://github.com/ggml-org/whisper.cpp.git
cd whisper.cpp
.\models\download-ggml-model.cmd base.en
cmake -B build
cmake --build build -j --config Release
.\build\bin\Release\whisper-cli.exe -m models\ggml-base.en.bin -f samples\jfk.wav

ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav

faster-whisper, and the two apps

• Const-me/Whisper, the project people call WhisperDesktop, is a Windows only GUI that runs Whisper on any Direct3D 11 GPU, so no NVIDIA card is needed. Unpack WhisperDesktop.zip from the releases page and run the .exe. It is MPL 2.0 and not archived, but the last release is 1.12 from July 2023, and the one commit since is a May 2026 README warning about a lookalike website. Download it only from the GitHub releases page, and treat it as finished, not maintained.
• Buzz is an MIT licensed app for Windows, macOS and Linux. It handles files and YouTube links, live transcription from the microphone, speaker identification, and export to TXT, SRT and VTT. Its README sends Windows users to SourceForge for the installer and says it is unsigned, so Windows warns you. It also installs from PyPI, for which the README asks for ffmpeg and a Python 3.12 environment.

faster-whisper runs the same models through CTranslate2. Its README says it is up to four times faster than openai/whisper at the same accuracy with less memory, and that 8 bit quantization helps further. It needs Python 3.9 or newer, plus cuBLAS and cuDNN 9 for CUDA 12 on the GPU. For those libraries on Windows the README points at Purfview's whisper-standalone-win archive. It is a Python library, not a command: the README's example loads WhisperModel and calls transcribe, with device cpu and compute type int8 when there is no GPU.

pip install faster-whisper

pip install buzz-captions
python -m buzz

Which route suits whom

• An app on any GPU, if a 2023 build is acceptable: WhisperDesktop. An app still being worked on: Buzz.
• A script or a batch job: faster-whisper, or whisper.cpp to avoid Python.
• OpenAI's reference behaviour: the openai-whisper package.
• No GPU at all: whisper.cpp or faster-whisper with a small quantized model.

Whisper is for files. Live captions are a different job

Everything above is built around audio that already exists. Captioning a call or a stream as it happens needs a model that runs much faster than the audio on a laptop, and it needs the sound the PC plays, not a file.

Windows 11 does the same-language half for free. Microsoft's documentation says live captions is in Windows 11 version 22H2 and later, that Windows key plus Ctrl plus L turns it on, and that all audio processing happens on the device. Translation needs a Copilot+ PC on version 24H2 or later, and covers over 40 languages into English and 27 into Chinese (Simplified).

Fraze covers the translation gap on a Windows 11 PC that is not a Copilot+ PC, and it does not run Whisper. Speech is NVIDIA's Parakeet TDT 0.6B v3, int8 ONNX through sherpa-onnx, fed by Silero voice activity detection. Translation is Mozilla's Bergamot models in WebAssembly. Both run on your machine, so nothing is uploaded, and it works with Wi-Fi off after the first download. Lines land about half a second to a second after a speaker pauses. The first 30 minutes are free, then $29.99 once or $2.99 a month, both unlimited.

Parakeet against Whisper, measured

Word error rate in percent, lower is better, from our own runs, which are not an independent test. LibriSpeech clean: Parakeet TDT 0.6B v3 2.15, Whisper large-v3-turbo 1.93. Earnings-22 conference calls: Parakeet 11.20, Whisper turbo 12.35. Do not read that last pair as a win for Parakeet. Our Whisper runs on the English sets went through WhisperKit, a Core ML port, and the benchmark repository notes that the reference PyTorch build scores about 11.1 on the same set, level with Parakeet. FLEURS read speech, 200 clips each, with a hosted Whisper large-v3-turbo: German 6.21 for Parakeet against 5.33 for Whisper, Spanish 6.05 against 5.68, French 7.83 against 6.82, Russian 9.18 against 7.48.

Whisper wins every one of those non-English sets, and its model card lists 99 languages against Parakeet's 25. What Parakeet buys for live captions is speed: in Fraze's own engine it ran at about 0.03 times real time on an Apple silicon Mac mini, fast enough to transcribe each sentence the moment the speaker pauses. Whisper large-v3-turbo is MIT, Parakeet v3 is CC-BY-4.0. If accuracy on a finished recording matters more than latency, use Whisper.

Questions